The Architecture

Three major institutions shape American foreign policy outcomes in the Middle East. The government. The military-intelligence apparatus. And the national press.

The first two are subject to oversight, elections, and consequences when they get things wrong. The third is not — by design and by constitutional necessity. Freedom of the press is foundational. What that freedom does not confer, however, is immunity from analysis. An institution powerful enough to shape geopolitical outcomes is powerful enough to be examined as a geopolitical actor.

That examination rarely happens. Journalists cover power. They do not typically regard themselves as an instance of it.

The record says otherwise.

Iran Reads the Coverage

For several years running, the dominant narrative in elite American and European press coverage of the Middle East has organized itself around a set of premises: that Israeli military operations lack legitimacy, that American pressure on Iran is counterproductive or futile, that world opinion is turning decisively against the U.S.-Israel posture, and that the trajectory of events favors a negotiated outcome on terms closer to Tehran’s than Washington’s.

Each of these premises has been reported as fact. Some are contested. Some have proven wrong. The important point is not whether any individual claim was accurate — it’s what that sustained narrative communicated to its full audience.

The press is fluent in fact-checking. It is not in the business of framing-checking. A fact can be verified and a frame can still mislead — by what it emphasizes, what it omits, what it treats as signal versus noise, which voices it sources and which it doesn’t. The individual facts in Middle East coverage were often defensible. The composite picture they assembled was not a neutral rendering of events. It was an argument. The press did not label it as one.

That audience included Iran.

The Iranian government is reading the international press now. It tracks domestic American polling. It watches congressional opposition, campus demonstrations, UN votes, and the tone of editorial boards. What it saw, consistently, over a period of years, was a picture of an America losing the argument — the forever wars, domestically divided, internationally isolated, and facing a midterm political environment that would constrain its ability to maintain pressure.

To reinforce the picture, the press parades a rotating cast of ex-generals and former officials who confirm that the U.S. is losing the war, that the policy is failing, that the costs are unsustainable. Watch Fox News and you will find an entirely different set of ex-generals and former officials who hold the opposite view — and who never appear on the other channels. The selection is not neutral. It is the argument. What goes unreported is how much has actually been accomplished, and how quickly. The particular ex-generals are chosen because they reinforce the narrative. No attempt is being made at balanced, unbiased reporting.

If you are in Tehran and that is your information environment, the rational response is to wait. Not to make concessions that are politically costly at home. Not to offer flexibility on the nuclear program. To run out the clock on an adversary the press has told you is already losing the debate.

Iran made strategic decisions consistent with that assessment — the one the press has been shouting from the rooftops.

Those decisions were wrong. The pressure did not break. The political environment in Washington did not constrain the administration’s ability to act. The international isolation the press described was real in the press’s opinion polls and not real in operational terms. Those same polls had missed consistently on Donald Trump — in 2016, in 2020, in 2024 — and Iran was reading them as reliable intelligence. Iran held out for a favorable negotiating position that was not available — because the conditions the press told them would produce it did not materialize.

The consequences of that miscalculation are now arriving. Iran is facing a Hormuz blockade, inflation above 40%, and negotiations resuming under evident duress — with the U.S. demanding a twenty-year suspension of enrichment that bears no resemblance to the favorable terms Tehran was waiting for. Whether the final outcome is economic collapse, some form of structured receivership over Iranian energy infrastructure, or a settlement negotiated from genuine weakness rather than manufactured strength — it will reflect decisions made inside a false picture of the correlation of forces.

The press built that picture.

This is not an argument the press will make about itself. And the outcome record is not one it will dwell on: Israel is still a functioning state with a functioning military, and an adversary that calibrates its strategy to press-generated opinion rather than the actual balance of power is an adversary that will be surprised.

The Connective Tissue

Events that appear unrelated are not always unrelated.

The current administration’s pressure campaign against elite universities reads, in most coverage, as a culture war story — an attack on academic freedom, an authoritarian intrusion into institutional independence. That framing is available. It is not the only framing.

Universities are the credentialing and formation system for the national press. The journalists, editors, researchers, and producers who generate elite American media are, in disproportionate numbers, products of the same institutional culture the administration is now pressuring. That culture has a set of assumptions about the Middle East, about American power, and about the moral weight of various actors in the conflict — assumptions that move from the seminar room into the newsroom without being substantially interrogated at either end.

The administration is not shy about having identified the press as an adversary. The university pressure and the press antagonism are expressions of the same analysis: that a set of interconnected institutions — academic, journalistic, nonprofit, legal — constitute a coherent power center, and that the correct response to a power center is counter-pressure.

You do not have to agree with that analysis to recognize it as an analysis. And you cannot understand why the administration is doing what it is doing without engaging with it on those terms.

The press covers the individual events — the executive orders, the funding freezes, the legal challenges. It does not, in the main, report on itself as the subject of a counter-campaign, because that would require it to acknowledge itself as a power center. That acknowledgment does not fit the professional self-image. So the connective tissue goes unreported. The Supreme Court is routinely labeled “conservative” — a political characterization, not a legal one — which frames every decision before the reader reaches the argument. That is not reporting. That is framing. The press does not label it as such.

Neither does it correct the record on the Nobel Prize in Economics — which does not exist, but the press and universities play this game because the authority of the name does rhetorical work. The press that deploys fact-checking as a weapon against others has no interest in applying it to credentials that serve the narrative. (The Fake Nobel: Time to Tell the Truth.)

The Market Corrects What the Institution Won’t

The media’s control is now being eroded by social media and the alternative information ecosystem it has produced. The press’s response is instructive. It has moved aggressively to delegitimize unregulated platforms — pushing for content moderation, amplifying misinformation narratives, lobbying for regulatory frameworks that would restore gatekeeping authority. For the same reason, it argues against chatbots — another uncontrolled source of framing and information it cannot gatekeep. The structural logic is identical to what authoritarian regimes do when confronted with a free press: uncontrolled information is an existential threat to any institution whose power depends on controlling the narrative. The press does not recognize the parallel. It should.

The press has a standard explanation for why people don’t trust it: education, manipulation, the systematic campaign by political actors to undermine institutional credibility. Absent is “the public smells a rat.”

The Gallup numbers are worth sitting with. As of late 2025, 28% of Americans say they have a great deal or fair amount of trust in the mass media — the first time that figure has fallen below 30% in the poll’s fifty-year history. Republican trust stands at 8%. Independent trust at 27%. And Democrats, the press’s most reliable constituency, have declined to a bare majority: 51%, down from comfortable supermajorities a decade ago. Younger Democrats trust the media at rates below a third.

These are not the numbers of an institution winning its argument. They are the numbers of an institution losing its audience while telling itself the audience is wrong.

The Democratic Party, whose narrative has been largely coextensive with the elite press narrative for the past decade, is experiencing a parallel decline. The two institutions share an intellectual infrastructure, a professional class that circulates between them, and a string of outcomes that did not match the predictions — elections lost, coalitions that did not materialize, adversaries that did not fold under the pressure of coverage and condemnation.

The public is not reading the New York Times the way it once did — they play Wordle and use the food app. Cable news viewership is structurally declining across every demographic under sixty-five, and the panel discussions, op-ed ecosystems, and Sunday shows are all talking to a smaller room every year.

This is the democracy correcting for what internal accountability could not. Not through regulation, not through censorship, but through the oldest mechanism available: people stop paying attention when they stop believing what they’re being told.

The press will cover its own decline with considerable seriousness. It will attribute the decline to external forces. That is, after all, what it does.

How Power Centers Behave When They’re Losing

Every power center, when threatened, behaves the same way. It doesn’t concede. It escalates.

From inside the press, the case always looks airtight. Public opposition to the war. The U.S. losing ground. Midterms approaching with affordability and gas prices as wedge issues. Europe aligned against Washington. Every variable that has historically broken a foreign policy is on the board. The narrative is complete. The tipping point is surely imminent.

And yet nothing bends. The policy continues. The outcomes don’t follow the coverage. This is genuinely confusing to an institution that has mistaken the assembly of a compelling narrative for the exercise of actual power. The map is not the territory. The story is not the outcome. That distinction is the one the press is least equipped to make about itself.

Authoritarian regimes, when their propaganda stops working, don’t stop producing propaganda. They produce more of it, louder, with greater certainty. The tell is the volume and the desperation. An institution confident in its influence doesn’t need to scream. The Democratic leadership is doing the same — raising the rhetoric, amplifying the alarm, growing more desperate to reclaim a power seat that the public is no longer offering.

You don’t need a Gallup poll to see this. Watch the coverage for a week. The escalating alarm, the pivoting from one crisis frame to the next when the previous one fails to move the outcome, the same voices delivering the same verdict with increasing urgency. The levers are being pulled. Nothing is moving. The desperation is visible to anyone paying attention.

The press has an answer for everything. Governments, corporations, intelligence agencies, the military — all subject to its scrutiny, all held to account for their failures, their blind spots, their institutional self-interest. It is remarkably thorough. The one institution it cannot bring itself to examine with the same rigor is the one holding the pen.

The Iranians made decisions. Those decisions will cost them. The coverage will record the costs faithfully.

It will not explain what the coverage had to do with the decisions.

It never does.

The answer is not obvious. In fact, AI may be the rare technology where both outcomes are happening at once.

The Illusion of Commoditization

At first glance, the case for commoditization is compelling.

AI capabilities are spreading rapidly. Models that once seemed extraordinary—writing code, summarizing documents, reasoning through problems—are now widely accessible. APIs price intelligence per token. Open-weight models continue to improve. Costs are falling. From the perspective of many users, AI already behaves like a commodity.

This is not an illusion. But it is incomplete.

The New Search

The clearest example of genuine commoditization is also the most visible: AI is replacing search.

When someone asks a question that used to go to Google — a recipe, a definition, a product comparison, a travel recommendation — a commoditized model handles it fine. Google, Amazon, and others have already built this tier. It does not require the latest frontier model. It requires something fast, cheap, and good enough. That bar is being cleared by a widening range of models, including open-weight ones.

The same shift is coming to reference material. Searching a technical manual is not just keyword matching — it requires understanding context and intent. Books and documentation will likely come with question-and-answer capabilities built in. That tier of AI, too, will commoditize.

This is a real and large market. The economics favor it, the use case demands it, and users benefit from it.

The new search is commodity AI. Still, people over time may demand better answers — just as any camera was fine for a mobile phone fifteen years ago, and now they require professional-level resolution.

These Are Not Your Father’s Models

Training runs now cost on the order of hundreds of millions of dollars. They require massive compute clusters, specialized infrastructure, and tightly coordinated teams. More importantly, success depends on a growing body of tacit knowledge: how to curate data, stabilize training at scale, and design systems that improve reliably with size.

This begins to resemble industries that have historically resisted commoditization.

Consider commercial aircraft manufacturing. The physics is well understood. Many countries have the talent and capital to attempt it. Yet only a handful of firms—Airbus and Boeing—can reliably produce modern aircraft. The barrier is not just cost, but decades of accumulated experience, integrated systems, and unforgiving validation cycles. China, with its massive manufacturing base, deep engineering talent, and state resources, has spent decades trying to build a competitive commercial aircraft industry. It still relies on Airbus and Boeing for the aircraft its airlines fly.

Frontier AI may be closer to these industries than to traditional software.

The Open Source Counterargument

There is, however, a powerful counterargument grounded in history.

Open source has successfully commoditized some of the most complex software systems ever built. The Linux kernel runs much of the world. Modern compilers like LLVM rival or exceed proprietary alternatives. These systems are maintained by thousands of contributors and represent decades of cumulative expertise.

If such systems can be commoditized through distributed collaboration, why not frontier AI?

This is the right question—but it misses a key difference.

Compilers and operating systems are fundamentally design problems. Once the architecture is understood, improvements can be developed incrementally, tested locally, and integrated by a distributed community. The cost of verifying a contribution is relatively low.

Frontier AI is increasingly a training problem at scale.

Improvements often require running large, expensive experiments. Validation is not a unit test; it is a training run that may cost millions and take weeks. The system is tightly integrated, and small changes can have unpredictable effects. This makes it difficult to distribute innovation across a broad base of contributors.

Open source scales code. It struggles to scale expensive experimentation.

The EDA Analogy

Consider electronic design automation — EDA tools like Synopsys and Cadence produce software used to design advanced semiconductors. They are similar to compilers like LLVM in design and complexity. In principle, EDA tools should be excellent candidates for open-source development. In practice, they are dominated by a small number of firms—those two alone hold roughly 74% of the market, with high retention and recurring revenue that has persisted for decades.

The reason is not technical complexity alone. It is the cost of validation and deep integration with cutting-edge fabrication processes.

Google has been attempting to build open source EDA tooling for years. The effort has made progress, but remains validated primarily at 28nm process nodes — while the leading commercial fabs are operating at 3nm and pushing toward 2nm. That is roughly seven generations behind: 28nm → 20nm → 16/14nm → 10nm → 7nm → 5nm → 3nm. Each generation represents years of development and billions in R&D. That gap illustrates the problem precisely: it is not that the open source tools cannot write code. It is that closing the last distance requires deep, expensive integration with proprietary manufacturing processes that open source cannot easily replicate.

A bug in an EDA tool can result in a failed chip, with losses measured in millions or billions of dollars. Testing and verification are expensive and slow to iterate. As a result, these systems require extremely high reliability and deep integration with external constraints.

Training runs are expensive, failures are costly, and validation requires full-scale experiments rather than incremental tests. This makes the system difficult to modularize and even harder to improve through distributed contribution.

The Pipeline, Not the Model

Perhaps the most important shift in thinking is this: frontier AI is not an isolated app—it is a pipeline.

From the outside, it is easy to focus on the model itself: a transformer trained on vast amounts of data. The architecture is widely known. Open-source implementations exist. It can appear that success is primarily a function of scale.

But in practice, performance differences between leading models suggest something more complex.

Users consistently report that some models are more reliable, more coherent, better at reasoning, or simply “feel smarter” in ways that are difficult to capture in benchmarks. These differences persist even when architectures appear similar and training approaches are broadly understood.

This gap points to the importance of the end-to-end process used to build and refine models.

That process includes data sourcing, filtering, and weighting; experiment design and evaluation; training stability and optimization; infrastructure and throughput; post-training alignment and fine-tuning; and feedback loops from real-world usage.

Each component matters. More importantly, they interact.

A change in data affects optimization. A change in optimization affects stability. A change in architecture affects scaling behavior. The system is tightly coupled, and improvements emerge from how these elements work together.

This is not easily reducible to a set of published techniques.

Much of the advantage lies in tacit knowledge: lessons learned through failed experiments, subtle tradeoffs, and accumulated experience. This knowledge is expensive to acquire and difficult to transfer.

The Benchmark Problem

Benchmarks are the primary tool used to measure AI progress. They are also increasingly unreliable as a guide to real-world performance.

The problem is structural. Benchmarks measure what they measure — a defined set of tasks, evaluated in a defined way. Labs know what the benchmarks are. Training and post-training processes can be tuned, intentionally or not, to perform well on them. Over time, benchmark scores converge even when practical capability gaps persist.

Consider what a typical coding benchmark actually tests. Asking a model to write Pong, or solve a self-contained algorithmic problem, tells you something — the model can produce working code for a bounded task with a known solution that exists throughout its training data. But real programming is long-term and cumulative. It requires sustaining coherent architectural decisions across a large codebase, handling compounding complexity as systems grow, and recovering gracefully when something breaks deep in a structure the model itself built. Benchmark tasks are essentially Pong. They measure whether the model knows what code looks like. They say very little about whether it can build, develop, and maintain something real.

This creates a systematic illusion of commoditization.

Anthropic’s annualized revenue run rate reached $30 billion as of early 2026 — a roughly 14x increase from a year earlier. The number of enterprise customers spending over $1 million annually doubled in just two months. That is not the behavior of a market that has concluded the models are interchangeable.

A Personal Benchmark

Abstract arguments about benchmark reliability are one thing. Here is a concrete test.

I have spent roughly fifty years in computing and software, including significant work on compilers. To put the models to a real test, I set out to build a working C++ compiler at the C++20 standard — entirely AI-written, with me providing high-level direction but writing no code and doing no debugging myself. The task is not trivial. A modern C++ compiler is one of the most complex software systems a developer can attempt. The language specification runs to thousands of pages. The edge cases are unforgiving.

My company had already accumulated significant experience with Claude Code by this point and had made it our agentic tool of choice. But I started this project with Gemini, since at the time the benchmarks rated it at or near the top and Google was offering a good price to upgrade to their higher level subscription. After a week or so, it stalled. It got stuck on certain problems and went around in circles, making no progress while burning through all the compute time my pro subscription allocated. I would have had to debug the problems myself. It was getting nowhere. I switched to Claude Code and the project has proceeded for months now with no issues. In fairness, these models are improving rapidly — the same experiment run today might yield a different outcome.

To validate the work, I had it build a language verification suite — now at approximately 30,000 tests, which Claude Code assessed as comparable in coverage to commercial suites like Plum Hall. The compiler currently passes more than 60% of those tests — and that figure includes complex template processing and other notoriously difficult parts of the C++20 specification, not just the simpler conformance checks. It is also able to handle the Standard Template Library, which rules out any characterization of this as a toy implementation. Getting to 100% is not a technical barrier; it is a question of how much time I can allocate to a side project while running other work.

This is one experiment, one person, one task. It does not settle the question of which model is best across all use cases. But it is not a casual test either.

One frontier model could not do the task. The other could. That’s all we need to know here.

That gap does not show up in the benchmarks. It showed up in the work.

You may have heard of a C compiler that Anthropic built using Claude. It was a C compiler, not C++20. The difference matters — as the complexity note below explains.

For readers unfamiliar with compiler engineering: despite the similar names, a C++20 compiler is conservatively at least 100 times more complex than a C compiler — and likely more. C is a small, stable language. C++20 is one of the most specification-heavy, edge-case-dense targets in all of software development.

The Role of Liability

In high-stakes domains—medicine, law, software development—errors are costly, delayed, and hard to detect. A misdiagnosis may not surface for months. A legal argument built on a hallucinated precedent may not fail until it reaches a judge. A coding error in a financial system may sit undetected until it becomes a breach. When mistakes are cheap and visible, good enough wins. When they are expensive and invisible until they aren’t, you buy quality.

The accuracy math is asymmetric. In high-liability applications, the difference between a model that is 98% accurate and one that is 99.9% accurate is not a 1.9% improvement — it is the difference between a functional product and a lawsuit. That gap does not show up in most benchmarks. It shows up when something goes wrong.

Regulation reinforces this. Compliance frameworks like the EU AI Act impose documentation, auditing, and accountability requirements that only well-resourced organizations can absorb — giving regulated industries another reason to stay with providers who can demonstrate compliance, not just capability.

The Self-Improving Wildcard

There is a more radical possibility: the models themselves may eventually handle their own optimization.

Deep reinforcement learning has already demonstrated that AI systems can discover solutions that humans missed entirely, even after centuries of effort. AlphaGo did not just learn to play Go — it discovered strategic principles that human masters, after more than a thousand years of study, had never found. It found them by playing against itself at scale, unconstrained by human assumptions about how the game should be played.

If that same dynamic were applied to AI training pipelines — tuning data selection, stabilizing training dynamics, discovering more efficient architectures — the accumulated human expertise that currently constitutes the moat becomes less decisive. Whoever gets there first could compress decades of process knowledge into months. That changes the valuation question entirely: it is not just whether the current cost structure persists, but whether human expertise remains the binding constraint — or eventually gets automated away by the same technology those labs are building.

The Algorithmic Wildcard

The cost structure that underpins the moat argument is not fixed. Transformers have a fundamental scaling problem: the self-attention mechanism is O(n²) in sequence length. Double the context, quadruple the compute.

Research into linear attention, sparse attention, and state space models is already attempting to break that wall — pushing toward O(n log n) or better. If any of these succeeds at frontier quality levels, the cost per experiment drops structurally. Computer vision is the precedent: algorithmic improvements didn’t just reduce costs incrementally, they changed the economics of the problem entirely. The same kind of leap is plausible in language models.

Investors underwriting IPO valuations that assume the current cost structure persists for a decade are making a bet on architectural stagnation. That is worth being explicit about.

The IPO Question

This debate is about to leave the realm of theory. OpenAI’s most recent funding implied a valuation of roughly $852 billion; Anthropic’s Series G placed it at $380 billion, with reported annualized revenues of $24 billion and $30 billion respectively. Public market investors will be asked to bet on which version of this story is true. For some related articles, see: What Is Anthropic Worth? and What Is OpenAI Worth?.

The moat question is the valuation question. If commoditization wins, these companies are priced like utilities at best. If the oligopoly thesis holds, they may be the Boeings and TSMCs of the next decade — ugly unit economics today, durable position tomorrow. DeepSeek produced a model competitive with frontier systems for an estimated $5-6 million in compute — a striking data point, though one with constraints and caveats that are still being debated. It does not settle the question. It does establish that the question is open.

Conclusion

The answers to these questions need to be tracked for those trying to value investments in frontier companies. There are not any clear answers at the moment.

That said, the evidence available today leans toward the oligopoly thesis holding at the top of the stack — at least for now. Enterprise adoption is accelerating toward the frontier, not away from it. The performance gaps that matter in professional work are larger than benchmarks suggest. But the wildcards that could break the moat — algorithmic breakthroughs, self-improving systems — remain genuinely uncertain rather than imminent.

The floor is commoditizing. The ceiling is moving up faster. Investors who treat those as the same market may misprice both.

• Iran: The Bill Has Come Due

• Will the US Seize Iranian Oil and Natural Gas Assets?

• The Iran War: A Game of Three Card Monte

• The Iran War: Why Now?

• Iran: Ceasefire or Pause of Mutual Convenience?

Those five pieces built the case for action and diagnosed the ceasefire trap. This one answers the question they left open: what does the endgame actually look like?

The Framework: Conservatorship

The United States should place Iran’s oil and natural gas economy under permanent conservatorship. Conservatorship — permanent external control of Iran’s oil export infrastructure, with revenue directed to approved uses and supervised compliance as the only alternative to destruction.

The word matters. Conservatorship implies what the evidence supports: that the Iranian regime has demonstrated, across four decades and ten American presidencies, that it cannot be trusted to manage its own economic resources without weaponizing them. A bankrupt entity placed under court supervision does not get to vote on the terms of its own receivership. It complies or it loses the asset entirely.

The Iranian people may have buyer’s remorse. That is their problem, not ours. They chose this regime.

The mechanism is straightforward. U.S. naval forces encircle and seize Kharg Island, which handles roughly 90% of Iranian oil exports. The offer to Tehran is simple: accept unrestricted conservatorship or lose their oil and gas economy entirely.

The Alternative: Destruction

If Iran resists, the infrastructure is destroyed. Permanently. Not damaged — destroyed.

At that point the responsibility is entirely Tehran’s. The terms were survivable — the clerics could keep their government, keep their ideology, keep their country. They simply could not keep exporting violence funded by their nation’s oil wealth.

If they choose resistance over compliance, they choose rubble. That is not an American decision. It is an Iranian one.

Iran: Decide

Conservatorship or destruction.

Call it what it is: a pause of mutual convenience, not a serious attempt to end the war.

Who Gets What from the Pause

For Washington and Jerusalem, the timeout locks in battlefield gains. U.S. and Israeli assessments indicate Operation Epic Fury left large swaths of Iran’s conventional forces “combat ineffective for years” — air defenses gutted, the navy crippled, missile production disrupted. The pause also buys something Washington needed domestically: oil prices dropped and equity markets rallied the moment the ceasefire was announced. Forces stay poised to resume if Hormuz clogs again — and additional forces will arrive and be put in place. And exactly how much firepower and fight is left in Iran is something both Washington and Jerusalem would welcome two weeks to quietly ascertain. Israel, for its part, can now focus its resources and attention on finishing the job with Hezbollah in Lebanon.

For Tehran, the regime needs to catch its breath. It took devastating hits but survived. With its leverage now constrained and diminished, Iran needs time to regroup. What Iran can do is probe quietly — reconstituting proxy networks, testing gray-zone harassment below the ceasefire threshold, shaping energy market psychology without formally closing the strait. And two weeks of relative quiet gives Tehran something it badly needs: time to ascertain its own situation, make plans, and reconstruct the communication networks that sustained bombing will have disrupted.

The Clock Runs the Same; The Value of That Time Does Not

The pause is also profoundly asymmetric. Washington can continue moving forces, resupplying munitions, refining targeting packages, and planning the next phase — all without firing a shot. Iran cannot rebuild gutted air defenses in a fortnight, reconstitute its navy, or restock missile production under sanctions and surveillance. The clock runs the same for both sides; the value of that time does not.

Iran Is Negotiating as If It Won

Iran’s 10-point proposal — full sanctions relief, enrichment rights, Hormuz “control” with transit fees, U.S. regional withdrawal, reparations — is not a negotiating position. It’s a fantasy document. They are not going to get their ten points or even one.

What’s actually on the table may be nothing more than the U.S. not seizing Iran’s oil and natural gas assets at Kharg Island and elsewhere. But without someone controlling the purse strings, Iran will go right back to doing what it’s been doing — so even that one point seems non-negotiable.

The Casting at Islamabad Tells You the Rest

The U.S. delegation is headed by Vice President JD Vance — a man who opposed the war, was excluded from the February 11 White House meeting where Netanyahu personally lobbied Trump to launch it, and is not a serious player in this administration any longer.

Vance is leading the delegation. When Trump is serious, Witkoff and Kushner lead.

Iran preferred Vance, seeing him as the soft option. Trump obliged — because letting the odd man out lead talks that are just for theater costs him nothing.

Iran has already signaled its level of seriousness — reports of it reclosing the Strait of Hormuz in response to fighting in Lebanon tell you everything you need to know about how committed they are to ending the hostilities in Iran. They want to preserve their proxies, which help them project power beyond their borders — a major point that started the war in the first place.

This will reignite.

Trump just needs some time to decide what the next step is. Nobody believes Iran will agree to anything reasonable — and if they do, nobody believes they’ll adhere to it.

The war is not ready to end.

It should not be a surprise if the U.S. hits the reset button in a big way, unannounced.

Cranky Old Guy publishes at mecrankyoldguy.com

Dario Amodei wants to turn Claude into the default assistant for white-collar work worldwide. Investors just valued Anthropic at $380 billion after a $30 billion round. The press is writing the superlatives. Before the bankers start printing prospectuses, let’s answer the question every retail investor should be asking: should I buy this, and if so, at what price?

The safest answer is also the simplest one. Buy an S&P index fund. You don’t pick the winner in the enterprise AI war. You own collectively the AI revolution. You capture Nvidia’s chip dominance, Microsoft’s and Amazon’s cloud infrastructure, Google’s efforts, and the broad productivity gains that will flow through every industry as companies lower their white-collar costs. These frontier AI companies themselves are liable to make it into the S&P 500 eventually — at which point you own them anyway. The AI transformation happens. You make money. Nobody gets hurt.

If that is too boring, read on. Because the math on Anthropic specifically is more interesting than the headlines suggest.

The Right Question

Most valuation stories you will read about Anthropic use explosive revenue forecasts and growth multiples. That approach misses the point. At maturity, this is not a high-growth story. It is a mature enterprise platform business selling recurring access to AI that augments or replaces white-collar labor. We should value it the way we value stable, recurring-revenue infrastructure and software businesses — with earnings and a multiple appropriate for a zero-growth utility-style company.

That is what we are going to do here. With numbers that exist today. Nothing else.

The Market

Anthropic’s core opportunity is the global cost of white-collar labor. This includes not just salaries but benefits, HR and support staff, office space, insurance, electricity, equipment, compliance, and every other cost of employing knowledge workers.

A reasonable estimate for the fully loaded annual cost of the world’s white-collar workforce is $60 trillion. Assume Claude solves 80% of that work. That creates $48 trillion in potential annual cost reduction. Businesses will not hand over the entire savings. They will only pay for the AI if they still net save half. That means the entire enterprise AI platform market could be worth $24 trillion per year in revenue at full maturity.

The Earnings

Revenue is not the number that matters most. Earnings are.

Right now Anthropic is effectively giving away significant usage on the free tier and through generous enterprise pilots. Current gross margins sit in the low-to-mid 40% range as a result of high inference costs. At maturity, as the mix shifts toward paid enterprise seats, Claude Code, agents, and reserved capacity, margins should improve. We use 55% gross margin in the mature scenario.

$24 trillion × 55% = $13.2 trillion in annual gross profit for the entire market.

The Multiple

This is a mature, zero-growth enterprise platform — recurring revenue, high switching costs in regulated industries, but no assumption of endless expansion. The right comparable is AT&T: essential infrastructure, stable cash flows, grows roughly with the economy.

AT&T trades at roughly 13.6 times earnings.

$13.2 trillion × 13.6x = approximately $179.5 trillion for the entire addressable market.

What You Are Paying

Anthropic’s latest valuation is $380 billion.

Against a potential mature market of $179.5 trillion, $380 billion is less than a quarter of one percent. At this scale, cents on the dollar breaks down as a metaphor. The market is worth roughly 470 times the asking price.

These are aggressive assumptions — 80% of white-collar work solved by AI, and businesses willing to pay if they keep half the savings. Both have to hold. But if they do, $380 billion is not a valuation. It is a rounding error.

The Conclusion

Based on what we know today, $380 billion is not an absurd price for Anthropic.

It assumes white-collar work is 80% solvable by AI, that businesses willingly share half the savings, that gross margins reach 55% at scale, and that Anthropic captures a meaningful share of that $24 trillion revenue pool while maintaining a durable moat. Those are large assumptions. Current margins are still pressured by inference costs. Competition from OpenAI, Google, xAI, and open-source models is intense.

The enterprise market is contested from day one — unlike consumer AI, there is no period of unchallenged dominance. But Anthropic is already winning it. Among U.S. businesses, Anthropic’s share of combined enterprise AI spend has gone from roughly 10% at the start of 2025 to over 65% by early 2026. The enterprise API market tells the same story: Anthropic rose from 12% to 32% market share in eighteen months. Claude Code alone generates $2.5 billion in annualized revenue and accounts for 4% of all public GitHub commits globally.

The IPO is soon — Anthropic is reportedly targeting late 2026. The unknowns — actual automation rates, margin trajectory, competitive moat, regulatory risk — are real but unquantifiable today. If you need those numbers before you can get comfortable, the Anthropic investment is not for you. You cannot value what cannot yet be valued. Buy the S&P and wait. That is not a bad outcome.

If you can live with the math we have — known fully loaded labor costs, a clear value-sharing assumption, observable margins improving toward 55%, and an AT&T-style multiple for a mature zero-growth business — then $380 billion looks like a fraction of a fraction of the potential long-term value.

One number worth keeping in mind: Anthropic’s revenue has grown from roughly $4 billion annualized in mid-2025 to over $30 billion by early 2026 — roughly 10x in less than a year. The free tier exists to get enterprises hooked and to achieve market dominance. Once Claude is embedded in the workflow of a law firm, a bank, or a hospital, it becomes as essential as the phone system. Costs and dependence grow gradually and leaving becomes unthinkable.

Whether it pays off is a question for another day. Right now, today, the disciplined math says the upside is enormous if the vision holds — and $380 billion is a very small bet on a very large market.

If you want to subtract China, where local competitors dominate and Anthropic has no real path to market, the numbers still don’t move enough to matter.

We don’t know precisely how the future will unfold — but with a market potentially worth 470 times the asking price, there is an enormous amount of cushion for being wrong. The formulas here are simple. Tweak the numbers yourself. There is a lot of room to be conservative and still arrive at a number that makes $380 billion look cheap.

If 470x sounds impossible, it isn’t. Walmart returned roughly 15,000x from its 1970 IPO. Microsoft returned roughly 5,000x from 1986. Amazon returned roughly 2,800x from 1997. Apple returned roughly 2,600x from 1980. All of them looked expensive on the day they went public. None of them were selling intelligence to every corporation on earth.

Sam Altman wants a trillion dollars. Wall Street is polishing its pitch books. The press is already writing the superlatives. Before the bankers start printing prospectuses, let’s answer the question every retail investor should be asking: should I buy this, and if so, at what price?

The safest answer is also the simplest one. Buy an S&P index fund. You don’t pick the winner. You own collectively the AI revolution. You capture Nvidia’s chip dominance, Microsoft’s enterprise distribution, Google’s search transition, Amazon’s cloud infrastructure, and whatever emerges from the open source world. You also capture the efficiency gains rippling through every industry that adopts AI — healthcare, finance, manufacturing, logistics — as those savings flow straight to corporate earnings and stock prices. You do not bet on which consumer AI subscription survives the next five years of a brutal, capital-intensive war of attrition. The AI revolution happens. You make money. Nobody gets hurt.

If that is too boring, read on. Because the math on OpenAI specifically is more interesting than the headlines suggest.

The Right Question

Every valuation story you will read about the OpenAI IPO uses revenue multiples. Price-to-sales. Forty times revenue. A hundred times revenue. The argument is always the same: the growth rate justifies the multiple.

It doesn’t. And here’s why.

OpenAI is not a growth company in any meaningful sense. The consumer AI layer it is building behaves like a mature subscription market at scale. The market it is selling into — human beings who want AI access — is fixed. It grows only as fast as the global population grows, plus inflation. There is no new market to create. There is no untapped geography to conquer. Everyone on earth who can afford a phone is the market.

You do not value a fixed market with a growth multiple. You value it the same way you value AT&T or American Express. With earnings. With a price-to-earnings ratio appropriate for a mature, stable, recurring revenue business.

That is what we are going to do here. With numbers that exist today. Nothing else.

The Market

OpenAI is, at its core, a consumer AI company. ChatGPT has 910 million weekly active users. People don’t say they are going to use an AI assistant. They say they are going to ask ChatGPT. That is what brand dominance looks like. Google took twenty years to become a verb. ChatGPT did it in two.

The right comparable for this business is the mobile phone.

Today there are 9.2 billion mobile phone subscriptions worldwide. Not unique users — subscriptions. Lines being paid for every month. That number grows only with population and economic development. It is a mature, universal, utility market.

Consumer AI will get there. The question is not whether people will pay for AI access — they already are. The question is what they will pay and how many of them there are.

Ten dollars a month is a reasonable estimate. It is roughly what a basic phone plan costs in most of the world. It is cheap enough to be a rounding error in a household budget. Expensive enough to generate real revenue at scale. Call it a guess — but it is an informed one, anchored to what people already pay for utility access to a network.

The math is simple.

9.2 billion subscriptions × $10/month × 12 months = $1.1 trillion in annual revenue.

That is the consumer AI market. Every person on earth with a phone, paying ten dollars a month. One number. No speculation about displacement rates or labor market disruption or how many lawyers AI will replace. Just: this is the market if AI access becomes as universal as the phone.

The Earnings

Revenue is not the number that matters. Earnings are.

OpenAI’s gross margin today is 33%. Inference costs — the computing power required to answer every query — eat the rest. That is the number we have. That is the number we use.

$1.1 trillion × 33% = $363 billion in gross profit.

That is the earnings number. The number a rational investor should be applying a multiple to.

The Multiple

This is a consumer subscription utility. It grows with population. It grows with inflation. It does not grow because AI gets smarter — that is already priced into the subscription. It does not grow because white collar workers get displaced — that is the corporate market, a separate business with a separate customer and a separate conversation.

The right comparable is AT&T or American Express.

AT&T trades at 13.6 times earnings. Pure utility infrastructure. Essential, universal, nobody cancels, grows at roughly the rate of the economy.

American Express trades at 21.9 times earnings. A premium consumer franchise with a loyalty moat. Slightly more dynamic, but still mature and bounded.

Consumer AI sits between them. Call it 15 to 18 times earnings.

$363 billion × 13.6x = $4.9 trillion $363 billion × 18x = $6.5 trillion $363 billion × 21.9x = $7.9 trillion

The range for the entire consumer AI market, priced like what it actually is, is $5 to $8 trillion.

That is the whole market. Every human on earth. One company. No competition assumed.

What You Are Paying

One trillion dollars.

Against a market of $5 to $8 trillion, $1 trillion is 12 to 20 cents on the dollar.

For a company that already has 910 million weekly active users — roughly 10% of the total addressable market already in hand — that is not obviously crazy. It is a reasonable entry point if you believe consumer AI becomes as universal as the phone and OpenAI is still standing when it does.

At $1 trillion, OpenAI is not wildly overpriced. It is priced at a fraction of what the market is worth if AI access becomes infrastructure — which, based on everything we can observe today, it will.

The Conclusion

Based on what we know today, $1 trillion is a defensible price for OpenAI.

It is not cheap. It assumes the consumer AI market reaches full penetration, that $10 a month becomes the global price of AI access, and that OpenAI captures and holds that market the way AT&T holds its phone subscribers. Those are large assumptions. But they are not fantastical ones.

One number worth keeping in mind: of OpenAI’s 910 million weekly active users, only 50 million are paying subscribers today — 0.54% of the total addressable market. That looks like a problem. It isn’t. It is the pipeline. The free tier exists to get people hooked and to achieve market dominance. People become addicted to it, or more precisely, it becomes an essential part of their lives — the same way a camera became essential not because anyone decided to buy a camera, but because it came with the phone. Once ChatGPT is that embedded, the free tier shrinks and the price goes up slowly, the boiling frog model, the way a frog does not notice the water heating. The 860 million convert — not because they want to pay, but because they cannot imagine not having it. The market the math describes is real. The bridge from here to there is human nature.

Right now, OpenAI has no meaningful competition for world dominance in the consumer AI market. That will change. Competition will come. But there is no way to estimate when, from whom, or at what cost — and anyone who tells you otherwise is guessing. So we leave it out. The model is built on what exists, not on what might.

The IPO is now. The unknowns — competition, margin trajectory, inference cost curves, regulatory risk — are real but unquantifiable today. If you need those numbers before you can get comfortable, the OpenAI IPO is not for you. You cannot value what cannot yet be valued. Buy the S&P and wait. That is not a bad outcome.

If you can live with the math we have — known market, known margin, known comparables — then $1 trillion is not a bubble. It is a bet on a market that demonstrably exists, priced at a fraction of what that market is worth. As the industry matures, the model can be refined. Right now, today, at IPO time, the honest math produces a range of $5 to $8 trillion for the whole market and a $1 trillion ask for a company already holding 10% of it.

Whether it pays off is a question for another day.

If you want to subtract China, where local competitors dominate and OpenAI has no real path to market, knock off roughly 1.7 billion subscriptions — the market drops to about $4 to $6.5 trillion and $1 trillion is still 15 to 25 cents on the dollar.

Here is the complete official record of why Pete Hegseth fired Army Chief of Staff Gen. Randy George:

“General Randy A. George will be retiring from his position as the 41st Chief of Staff of the Army effective immediately. The Department of War is grateful for General George’s decades of service to our nation. We wish him well in his retirement.”

That’s it. No reason given. He wasn’t required to say anything else. The president runs the military. That’s Article II. It’s not a loophole — it’s the design. If Congress doesn’t like how the Pentagon is being run, they have remedies: hearings, legislation, oversight, the power of the purse. Full stop.

So what did the mainstream press tell you anyway?

The New York Times told you it was about Hegseth blocking promotions for Black and female officers — citing unnamed military officials. CBS told you Hegseth wants someone who will implement his “vision” (whatever that means) — citing an anonymous source. The Daily Beast ran the headline “Pentagon Pete’s Bigoted Reason for Firing Top General Randy George Leaks” — with zero on-record sourcing. Sen. Chris Murphy told you the generals were probably telling Hegseth his Iran war plans were “unworkable, disastrous, and deadly.” MSNBC told you flatly, not as opinion but as a news sentence, that “a scandal-plagued former Fox News host is destabilizing the U.S. military.” Joe Scarborough compared him to Stalin.

Every single reason came from someone who either wasn’t in the room, has an obvious political interest, or is anonymous and by definition unverifiable. The press assembled these fragments into a confident narrative and presented it as news — with “purge,” “destabilizing,” “unprecedented,” and “scandal-plagued” deployed as straight descriptors, not editorializing.

Fox News — Hegseth’s former employer, the network the prestige press treats as a national embarrassment — reported the story this way: Hegseth gave George no reason for asking him to step down. An Army official told Fox this directly. Clean, sourced, accurate. The most honest sentence written about this story came from the outlet that supposedly doesn’t do journalism.

Better. By a lot.

The Brief

The press loves the word unprecedented. Truman fired Douglas MacArthur — five-star general, hero of the Pacific — during the Korean War. The press called it appropriate exercise of executive authority. Obama fired General Stanley McChrystal after reading that he’d been mocking the vice president in a magazine profile. Gone, no process, no Senate hearing demanded. Also appropriate. Hegseth fires a general with no public explanation — exactly what Truman and Obama did — and it’s a five-alarm purge. Same action. Different president. Different coverage.

Here’s another favorite: loyalty. Trump wants loyalty, the press tells you, as if demanding personal fealty is self-evidently corrupt. Did Trump ever say that? Or did he say he wants people who share his vision? Every president in history has wanted people who agree with his agenda. The press calls it loyalty when they don’t like the president. Try succeeding in any business without following what the CEO wants. In the Pentagon press pool it becomes a purge, a destabilization, a threat to the republic.

Trump tried it the other way. His first term was stocked with people the press called the right choices — Mattis, McMaster, Kelly, Tillerson, Bolton. The adults in the room. Every one of them fought him, leaked against him, or quit in protest. Then they wrote books. You were there to serve the president, not to write a book. Trump looked at that record and decided he’d rather have people who actually want to implement his agenda. Most people would call it not making the same mistake twice.

Then there’s the prosecution’s case against Hegseth personally. He’s 45. He has tattoos. He worked at Fox News. That’s the brief.

He’s 45 — Donald Rumsfeld was 43 when he first took the job. The Democratic senators demanding he prove his experience average 64 years old. Half are over 65. One is 92. Silicon Valley is operated by people in their twenties and thirties. Nobody calls that a crisis. Barack Obama became commander in chief with zero executive experience — community organizer, state senator, three years in the U.S. Senate mostly spent campaigning. No management experience. No organization run. The press called it historic. Pete Hegseth has a Princeton degree, a Harvard master’s, combat deployments in Iraq, Afghanistan, and Guantanamo, and years of organizational leadership. The experience argument apparently has an on/off switch.

He has tattoos — the military is probably the most tattooed institution in America. A 2023 Pew Research study found tattoo rates significantly higher among veterans and active-duty service members under 50 than the general population. Hegseth served in Iraq, Afghanistan, and Guantanamo. During his confirmation hearings, senators grilled him on a Jerusalem cross and a “Deus Vult” tattoo as supposed extremism signals. Every generation of old men finds something about younger people’s appearance to declare disqualifying. It’s never actually about the ink. (See What Tattoos Say About the Present.)

He worked at Fox News — which, on this specific story, reported more accurately than every outlet attacking him for it. Fox dominated cable news ratings for twenty straight years. Profitable. Loyal audience. The Washington Post loses money and exists because Jeff Bezos can absorb the loss. Fox is profitable on news. The Times is profitable on Wordle and cooking apps. When the prestige public readership is given a choice, they’re not choosing the prestige press for the wonderful news reporting — they’re playing Wordle.

Their explanation for Fox’s dominance: the audience is uneducated, low-information, easily manipulated. They sort by college degree — college educated equals reliable, non-college equals credulous. The college-educated consensus gave us the Iraq War, the Afghanistan nation-building fantasy, and twenty years of foreign policy failures nobody in the credentialed class was ever held accountable for. The “low information” voters looked at that record and drew a conclusion.

The Record Nobody Mentions

Let’s talk about the actual record. Under Hegseth: Iran strikes executed — successful support for Israel in the first war against Iran, dismantling an evil regime we have been slow-walking and afraid to act against for fifty years. Pilots rescued. Nicaragua operation completed. No catastrophic operational failures.

Under his predecessor Lloyd Austin — the four stars, the confirmed credentials, the biography the press never questioned — the United States lost an entire country in eleven days. Thirteen service members killed at Abbey Gate. Billions in equipment abandoned to the Taliban. Then Ukraine: just hand wringing. Austin ran all of it. The press covered Afghanistan for two weeks and moved on. Ukraine got sympathetic framing. “Destabilizing” didn’t come up much.

Then there’s the part the press buried fastest. Austin secretly hospitalized himself for prostate cancer surgery and didn’t tell the White House, the Deputy Secretary of Defense, or Congress for days. The Pentagon Inspector General later found the hospitalization “unnecessarily raised national security risks by breaking the chain of command notification.” That’s an on-record government finding. Nobody called for his resignation. The press covered it for a news cycle and moved on.

Pete Hegseth fires a general with no public explanation and it’s a five-alarm destabilization of American democracy.

The standard isn’t a standard. It’s a preference.

Fishing for a Narrative

When Trump speaks, the press reaches for “asserts” and “claims” — distancing language signaling to the reader: don’t believe this. When nine anonymous Pentagon officials say Hegseth fired generals because of racial bias, it runs as fact. No “asserts.” No “claims.” “Revealed.” “Leaked.” The Daily Beast: “Bigoted Reason for Firing Top General Randy George Leaks” — as if documented truth escaped containment, not as if an anonymous source with an agenda called a reporter.

Consider what they produced on the George firing. Ten different reasons, all anonymous, several contradicting each other. DEI? Clashing personalities? Long-running Army grievances? The Driscoll relationship? Iran war plans? Austin association? Hegseth’s paranoia about his own job? The press ran all of them in the same news cycle without noting they can’t all be true simultaneously. That’s fishing for a narrative with every line in the water. The press harps on Trump for changing his story on why we’re in Iran. Fair game. But the same press can collectively produce ten contradictory “factual” reasons why a general was fired — all anonymous, all reported as established fact — and nobody calls that a credibility problem. One standard for the president. No standard for the press.

Watching the press cover the Iran war is like watching a sports bar cover the Super Bowl. Everyone has an opinion. Nobody is the coach. Retired generals pronounce the strategy disastrous without a current briefing. Senators walk out of classified sessions straight to the cameras. Think tank analysts who’ve never commanded anything explain what Hegseth is doing wrong. It’s opinion dressed as expertise.

The most serious problem isn’t the sloppiness. It’s what the coverage does during wartime. Democratic senators publicly declare the war plans “unworkable and disastrous” — precisely the message Tehran wants on the record. Joe Scarborough said America’s enemies “must be delighted” by Hegseth’s leadership. Cheerfully. On a news network. During a war.

There is a word for broadcasting to your enemies that your military leadership is in chaos and your war plans are unworkable. It used to have a harder name than journalism.

This piece is not defending Hegseth’s decisions or arguing the firings were wise. Those are legitimate questions.

The treatment of Hegseth’s firings is not factual reporting. It’s innuendo and opinion. And it sounds mostly like sour grapes from sore loser Democrat leadership and a news media that the public rejects — except for their Wordle games.

Remember the movie “What Ever Happened to Baby Jane?” A faded child star, past her prime, living in a decaying mansion off memories of when everybody loved her. Seems we’re in a remake.

The silence is deafening. No price alerts blowing up your phone. No laser-eyed tech bros on CNBC explaining why Bitcoin hits a million by Christmas. No segment at 11. No Wall Street analyst with a fresh price target. Just quiet.

The Number Nobody Is Talking About

Here’s the number nobody is talking about: crypto has lost nearly $2 trillion in value from its peak. The total market is down roughly 40% from its high. Forty percent. Gone.

Now try to imagine that math applied anywhere else. If the stock market dropped 40%, you wouldn’t be able to turn on a television without the latest news on it. There would be congressional hearings, Fed emergency meetings, and a presidential address. Every anchor in America would be doing their grave face.

Crypto just did it. Crickets.

Not one financial advisor telling you to hedge your portfolio with crypto. Not one wealth manager segment on CNBC. Not one “protect yourself from uncertainty” pitch. When inflation spikes, gold ads are everywhere. When markets wobble, bonds. Every real asset has a moment when someone tells you it belongs in your portfolio. Crypto has nothing to say for itself right now.

The banks and hedge funds that quietly bought in aren’t talking either. Nobody wants to explain that one to their investors or retail customers.

Crypto was never an asset class. It was a promotion. And promotions only work when the number is going up. When it stops going up, the promoters move on. There is no CNBC segment titled “Why Your Investment We Have Been Telling You About Has No Underlying Value.” There is no Wall Street analyst whose job is to tell you the emperor has no clothes. The coverage exists to sell the narrative. When the narrative stops paying, the coverage stops too.

What Is It Actually For?

So what is crypto actually for? Strip away the Web3 revolution talk and the DeFi-will-replace-banks pitch and the digital gold mythology, and you are left with two honest use cases. Moving money outside the reach of the law. And providing a banking system for people in countries cut off from the real one — your Venezuelans, your sanctioned states, your North Koreas and Irans. Those are legitimate needs. Did you add it to your portfolio for one of these purposes?

Everything else was marketing copy. Retail investors were the exit liquidity. They just didn’t know it.

The Trump Dimension

Donald Trump ran for president as the crypto candidate. The industry donated $238 million to get him elected — more than oil, gas, and pharma combined. They got a White House Crypto Czar, the dismantling of the DOJ’s crypto enforcement unit, and dropped investigations against Coinbase, Gemini, Ripple, and Kraken.

The Trump family built their own empire on top of it. World Liberty Financial collected 75% of all token sale proceeds — over $800 million in the first half of 2025 alone. A UAE-linked firm bought 49% of the company four days before the inauguration. A Chinese billionaire invested $30 million. His SEC investigation was subsequently dropped. They also launched $TRUMP and $MELANIA memecoins. Supporters bought in. Insiders cashed out.

Trump’s election triggered a frenzy. The market surged from $2.3 trillion to $4.3 trillion by October 2025. Then it collapsed — right back to where it was before the election crypto spent $238 million to buy. Pure hype, no floor.

Trump claims every stock market uptick as personal vindication. By his own logic, he owns this too.

Where’s the press conference?

The Circus Moved On

The promoters, for their part, have already moved on. No forwarding address.

The circus may come back to town. You’ll know.

Maybe crypto fades for good. Maybe it runs the hype roller coaster for another hundred years, taking on new riders every cycle. Or some of the faithful may never leave.

Dutch tulip bulbs only came back as a cautionary tale — another possible continuation.

The midterms are coming up. Democrats are being gifted a lot of fuel as a result of the war. Trump promised no more wars. Gas prices would climb, and affordability is the number one voter issue. Nobody on his national security team would have told him this would be a slam dunk or end quickly.

He did it anyway.

The foreign policy establishment wants to debate that diplomacy was working. The anti-war left wants to discuss the humanitarian cost. The generals want to debate exit strategies. The media wants to know why Trump made different claims on different days.

The right question is simpler: why now with midterms around the corner?

He Wanted to Finish Iran

For forty years, every American president chose to manage Iran rather than confront it. The nuclear program that Obama slow-walked into a decade-long deferral. The proxy network funding terror from Lebanon to Yemen. The regime that kept exporting instability while the international community held summits and issued statements. Confronting it was hard. Deferring it was easy. So they deferred.

History is not on his side to win the midterms anyway. If he had waited until after the midterms, he most likely would not have been able to finish Iran.

He pulled the trigger.

He killed Khamenei. He degraded the IRGC. He did the thing nobody else would do. The midterms have nothing to do with it. This was always on his list. The question of whether it helps or hurts in November is separate — and largely beside the point. Republicans almost certainly lose the House regardless. The president’s party has lost House seats in 26 of the last 28 midterms. There’s a slim chance he holds it. If he loses it, two years of investigations, subpoenas, and the end of his agenda follow. That’s the downside he was already looking at before the first strike landed.

The press, the DC know-it-alls and other pundits consistently underestimate and fail to understand Trump. I have written about this repeatedly. These are people who think they are very smart. They are not. I would have to reference at least a dozen op-eds I’ve written over the past year to cover this.

Being Right Covers a Lot of Ground

And he was right about Iran. The same way he was right about the border. The same way he was right about NATO burden-sharing. He says things that make serious people wince, takes actions that make serious people reach for their hair, and then the results arrive. Being right covers a lot of ground.

If the war wraps before November and he can stand at a podium and say he finished what nobody else would touch, that helps the slim chance of holding the House. If it drags and gas stays elevated, the midterms get harder. But they were already hard.

The Fool’s Mask

The people calling this reckless are making the same mistake they’ve made since 2015.

I’ve written about this directly. Trump is not a fool. He plays one. The caricature of the bumbling impulsive president who blunders into decisions, who listens to whoever whispered in his ear last, who needs to be managed by serious adults — that caricature has been wrong for ten years and the people holding it still won’t put it down. The man survived two impeachments, multiple indictments, bent the Republican Party to his will, banished the Democrats to a dark corner in the basement, and walked back into the Oval Office. That’s not what a fool being played looks like.

The Iran strike looks impulsive to everyone running the same analysis they ran in 2016, 2020, and 2024 — and got wrong every time. They see chaos and conclude he didn’t think it through. He saw a closing window and made his move.

The Putin narrative runs the same way. Before every major decision, the news cycle floods with stories about what Putin whispered in his ear. Trump lets it run — it’s useful cover if something goes wrong, and he gets the credit if it goes right. Whether you agree with his Ukraine policy or not, it’s a choice, not a con.

The Diploma Problem

Here is what unites every instance of Trump being catastrophically underestimated: the people doing it went to the right schools.

Harvard degrees. Nobel Prizes — the real kind, or the ideologically convenient kind they hand out every few years to academics who tell powerful people what they want to hear. The credential class has been trained to believe that certification equals insight. Trump has no such credential. Therefore he cannot be the smart one in the room. Therefore strategy must be chaos. Therefore calculation must be impulse.

The diploma is not a lens. It’s a blindfold.

The Two-Minute Drill

Iran is the mission. But Trump is also running a two-minute drill on the midterms, and he knows it.

Six months is not a lot of time. But it may be enough. He’s reining in Robert Kennedy. He got rid of Pam Bondi. Kristi Noem is gone. The cabinet faces that generated the most friction, that gave late-night television its material, that made Republicans in swing districts uncomfortable — they’re being quietly cycled out. The administration is behaving more like a government and less like a reality show. That’s not an accident.

If the economy cooperates, and Iran wraps with something he can credibly call a win, and gas comes back down — Republicans have a real chance to hold on. Slim, but real.

And the Democrats are helping. They own the shutdowns. Every one. They forced them, accomplished nothing, and made ordinary people’s lives miserable in the process. Now they’re the party that opposed a war while Americans were fighting it. The unpatriotic optics during wartime are not a small thing. Voters don’t forget which side was rooting against the home team.

But history doesn’t account for a Democratic Party this determined to snatch defeat from the jaws of victory and an adversary they consistently underestimate.

Charlie Munger Had It Right

Charlie Munger — Warren Buffett’s partner, one of the greatest investors who ever lived — was asked about Trump early in his first term. He had previously said Trump wasn’t morally qualified for the presidency. Then he reconsidered.

“He’s not wrong on everything,” Munger said. “And just because he isn’t like us, roll with it.”

That’s the whole argument in three sentences. Trump is different from us. He operates by different rules, takes different risks, makes decisions the credentialed class would never make. The instinct is to resist, to dismiss, to explain why he doesn’t understand what he’s doing.

The smarter move is to roll with it — and watch what happens.

The people calling it reckless are the same people who have been wrong about him for a decade. They are still certain they’re the smart ones in the room.

They could be right this time. Even a broken clock is right twice a day.

For more context, see: The Fool’s Mask, Do Democrats Really Have the Midterms in the Bag?, and my earlier Iran and Ukraine series.

In my earlier piece I argued that the dot-com era wasn’t a bubble in any meaningful sense — the transformation was real, the value created was real, and the NASDAQ sits at roughly 30 times where the “bubble” started. The internet was not a bubble. But Pets.com was. Webvan was. The people who bought the Netscape IPO and held it didn’t get rich — even though Netscape was real, mattered enormously, and made the transformation visible to the world.

AI is real. The transformation is real. That is not the question. The question is how to value new and existing companies that participate in this transformation.

The Exit Is Already Being Built

Let’s be blunt. The excitement around AI is lighting a fire under everyone, and the capital markets are going to capitalize on that excitement. They are counting on the fact that most people either cannot or will not ask the right questions to make sound investment decisions. There is no Consumer Reports for IPOs. There is no independent rating agency telling you what the unit economics actually are. There is the prospectus, in four-point font, and the analyst coverage from the banks that underwrote the deal.

Anthropic closed a $30 billion Series G in February 2026 at a $380 billion post-money valuation. That number did not emerge from a discounted cash flow model. It was preceded by a Wall Street Journal report citing anonymous sources familiar with the deal. Pre-IPO narrative management is standard practice. The WSJ story is how it begins.

The institutional investors who led the Series F got in at $183 billion five months earlier with preferred shares and liquidation preferences.

The investment community can sell the growth everyone can see. The $19 billion revenue run rate growing 10x annually is real — the best possible backdrop for moving preferred shares into a public float. The LP does not care whether Anthropic is profitable in 2030. The LP cares whether the fund returns capital on schedule. They will be gone before the math catches up.

The retail investor buying at IPO is not entering the same trade. They get common shares with no downside protection, no fund lifecycle exit, and no access to the unit economics the institutional investor reviewed before writing the check. They have the revenue growth number in the press release and analyst coverage initiated at “buy” by the banks that underwrote the deal. That conflict is disclosed in the prospectus, in four-point font, across two hundred pages nobody will read.

Wall Street will do fine. They just have to sell the general public on it long enough to not make it look like a scam. That is not cynicism — it is a description of a mechanism that has worked the same way for a hundred years.

The Nifty Fifty Pattern

Not every part of the AI bust will look like dot-com. Some of it will look like the Nifty Fifty.

The Nifty Fifty were the “one decision” stocks of the early 1970s — Xerox, Polaroid, Avon. The growth was so certain that any price was justified. The companies were real. The earnings were real. Then the multiple compressed — some businesses became dinosaurs displaced by the next technology wave, others simply never met the expectations baked into the price. Polaroid fell 90% and eventually went bankrupt. Avon fell 86%. The investors who bought at peak multiples spent a decade underwater — if they got their money back at all.

Anthropic at $380 billion — roughly 20 times annualized revenue, pre-profit, pre-moat, and pre-proof that the business model works at scale — is a Nifty Fifty setup with the downside more acute, not less.

Anthropic is doing genuinely impressive work. The revenue growth is real, the enterprise adoption is real, and the models are good. The question is not whether they are executing well. The question is what that execution is actually worth — and how hard it is for others to replicate it. The team that built Anthropic walked out of OpenAI and reconstructed it in months. What stops the next team from doing the same? Is there a real moat, or is the moat just being first and well-funded in a race where the tools to catch up are the very thing being sold? The revenue is real: $1 billion in December 2024 to $19 billion by March 2026, a trajectory with no precedent in enterprise software history. The company is not Pets.com. But roughly 20 times revenue in a compute arms race with no durable moat requires heroic assumptions about what comes next.

Apply the framework from my earlier piece on Buffett-speak: will the cash you take out comfortably exceed the price you paid before the moat runs out? At roughly 20 times revenue, that math has never been demonstrated at this scale in this competitive environment. Every dollar raised goes to compute. Skip a training cycle and you fall behind. It is Sun Microsystems with better press.

Most people buying AI stocks are not doing this math. If you are not asking whether the discounted free cash justifies the price, you are not making an investment. You are making a bet that someone will pay you more than you paid. That is the greater fool theory — and it works until it doesn’t.

The Gold Bar in Every Box

There is a question nobody is asking loudly enough: will there ever be any free cash?

The standard narrative is that AI losses are investment losses — spend now, revenue catches up, margins follow. That is the Amazon story. Amazon’s losses were building the warehouse network, the logistics infrastructure, AWS. Deposits against future dominance.

Anthropic’s losses are structurally different. The cost of delivering the product is the compute required to run inference on every query. That cost scales with revenue. If unit economics are negative at the product level, more revenue makes the losses larger. It is not an investment loss. It is a product loss. If every box comes with a gold bar inside, you cannot fix that by selling more boxes.

Traditional software economics work because marginal delivery cost approaches zero — a SaaS company at scale runs 80% gross margins because the servers are a rounding error. AI inference does not work that way. Every query costs real money in GPU time, power, and cooling. Underneath that is a capital expenditure layer unlike anything in software history: GPUs at tens of thousands of dollars each, purpose-built cooling infrastructure, power consumption measured in gigawatts. A normal software company’s infrastructure is a line item. An AI company’s infrastructure is the business.

The math that would justify these valuations is rarely shown explicitly. Anthropic’s gross margin is currently estimated at 40 to 55 percent. Mature software companies run 75 to 85 percent. The gap is compute cost — and that is not a temporary inefficiency. It is physics. Transformers require matrix multiplications. Matrix multiplications require GPUs. None of that goes to zero. Discount the projected free cash at a reasonable rate, subtract the capital expenditure to stay competitive, adjust for open source pricing pressure, and ask what number you get. Nobody is publishing that number because it is uncomfortable. At honest assumptions the discounted free cash flow does not get you to $380 billion. It probably does not get you within a country mile.

The gold bar is load-bearing. Inference costs are falling — but competitors’ costs fall at the same rate. And the bar may get heavier, not lighter: each new frontier model generation has required more training compute than the last. Agentic workflows compound this further — an agent completing a multi-step task consumes dramatically more tokens than a simple query, so the cost curve may not fall as fast as hardware improvements suggest even as individual token prices drop. Stay at the frontier and your costs go up. Fall behind and the cheaper alternatives — open source, Chinese labs, distilled models — eat your premium pricing from below. The squeeze runs in both directions simultaneously.

There is a deeper problem underneath this. Traditional accounting valuation was built for a world of physical assets — factories, inventory, land. Things you could walk up to and appraise. Nobody has a valuation methodology that works reliably for frontier AI companies.

The Drug Dealer Business Model

Drug dealers give the first hits away free. The economics only work once the customer is addicted and has no clean exit. What the AI companies are doing is not different — just legal, and the product being pushed is productivity.

Right now, corporate America is being hooked. Agentic AI is being woven into workflows, hiring decisions, and organizational structures. Companies are eliminating headcount on the assumption that AI handles it. The institutional knowledge of how to do things the old way is walking out the door with the people who knew how. Once that restructuring is complete, the switching cost is no longer an API key. It is organizational reconstruction. At that point the AI company can charge what it needs to charge.

This may be the only viable business model available. You cannot charge cost-covering prices today because a competitor is running the same subsidy play. Every major AI company races to get deepest into the enterprise stack before the pricing era arrives. It is a rational strategy — it requires surviving years of losses, which requires the capital raises, which requires the narrative, which requires the IPO. The whole apparatus is in service of buying enough time to get deep enough into enough enterprises that pricing power finally materializes.

The darker version: what if the addiction sets but the pricing power never arrives? Open source is always one generation behind but good enough for most enterprise use cases at zero marginal cost. If Meta or other players give away models for free or as open source, the ceiling on closed model pricing may never cover costs. You have restructured the entire global economy around a product that nobody can profitably deliver. That is not a bubble in any historical sense. That is something stranger and harder to name.

The Spinout Economy

The other part of the bust looks like 1999 — real technology, unwinnable unit economics, and a credential bubble layered on top.

Every senior researcher who left OpenAI got a lavishly funded startup. Ilya Sutskever left — funded at a $32 billion valuation with no product. Mira Murati left — funded. The VP of Research left — funded. The funding happens inside the Sand Hill Road network — the Stanford Mafia — the same tight circle recycling credentials into valuations for thirty years. They went to the same schools, sit on the same boards, validate each other’s numbers. The market cannot absorb fifteen frontier model companies each burning nine figures a year on compute. The network doesn’t care. The network gets its fees regardless.

Yann LeCun’s AMI raised $1 billion at a $3.5 billion valuation — twelve people, Paris headquarters, no product, no demo. The Turing Award is a receipt for work done decades ago, not a forecast. The man who funded LeCun for thirteen years — Mark Zuckerberg — is not in this round. In 1999 every ex-Netscape engineer got funded. The pedigree was real. The market wasn’t big enough for all of them.

What the Bust Looks Like

The dangerous game is picking AI pure-plays at IPO valuations on narrative alone. It is like going to the horse races and trying to pick the winner from the race bill. You studied the form. The professionals who set the odds have forgotten more about those horses than you will ever know. The house takes its cut regardless of who wins.

If AI transforms the economy the way the optimists believe, the S&P 500 will reflect it. The index owned Amazon, Google, Microsoft, and Nvidia through all the volatility and captured every dollar of value they created. It will own whatever the AI winners turn out to be. Do what Buffett says: buy an S&P index fund and let the sorting happen there — without making yourself crazy listening to market pundits, without losing sleep over which frontier model survives, without being on the wrong side of a trade that was designed before you sat down.

The sorting mechanism does not announce itself. It just sorts.

Related reading: What Dot-Com Bubble? · The Buffett Indicator: What It Is — and What It Is Not · The Most Misunderstood Phrases in Buffett-Speak · LeCun’s AMI: What Is the Proposition? · The OpenAI-Anthropic Story

Imagine every book and article published from this day forward carried a disclaimer on the first page: “You may read this. But you are strictly prohibited from using anything you learn here to make yourself a better writer, or more knowledgeable on this subject, in any way that competes with the author.”

No court would uphold it. No legislature ever passed it. Copyright law has always drawn a deliberate line: it protects original expression — not the underlying ideas, facts, methods, or knowledge you walk away with. That line exists for a reason. It is how every technological and intellectual revolution in history has worked. You study what came before, you improve on it, and the world gets better.

That disclaimer doesn’t exist in publishing — yet. But it exists right now in AI, buried in the Terms of Service of the two most powerful labs in the world.

OpenAI and Anthropic both have explicit clauses in their API agreements: You may pay us to use our models, but you are forbidden from using the outputs to train or improve any model that competes with ours. They call it “competitive distillation” and treat it like theft. When a dispute arose with Chinese labs in early 2026, Anthropic wrapped it in a national-security flag and ran it to Congress. Theater and misdirection. The ToS clause existed before that dispute and will exist long after it. It was never about national security. It is about protecting the moat. Meanwhile the Chinese government is actively enabling distillation for its own companies. American labs are playing whack-a-mole trying to stop it — and losing. The only people these ToS restrictions actually bind are legitimate American competitors.

The Law Already Drew This Line

Here’s what should make every American taxpayer and innovator angry: we already have a legal framework for protecting intellectual property. It is called copyright and patent law. Copyright protects specific expression — not ideas, methods, or knowledge. Patent law protects specific inventions for a fixed term, then releases them to the public. These are the general legal protections Congress designed after long debate, deliberately balancing investment incentives against the public’s right to learn, compete, and build. When you release a product into the world where others can use it and learn from it, that is the deal Congress made on your behalf. ToS is an attempt to get more than Congress ever intended to give.

The AI labs don’t like where that line was drawn. So they’re using ToS to move it.

Distillation is a standard machine-learning technique. A smaller “student” model learns patterns and reasoning from a larger “teacher” model’s outputs. Multiple legal analyses have concluded this is unlikely to constitute copyright infringement under current U.S. law. AI outputs generally lack human authorship — a point reinforced this month when the Supreme Court declined to hear Thaler v. Perlmutter, leaving in place the ruling that AI-generated outputs generally lack copyright protection. Copyright doesn’t protect ideas, methods, or systems. Patents don’t cover model behavior.

So the labs fall back on the one tool left: the contract you click “I Agree” to. They’ve invented a private rule that no statute or court has ever imposed — you can use our service, but not to get better at competing with us.

Apply that logic consistently and see where it leads. Every textbook publisher could ban students from using what they learned to compete with the author. Every journal could prohibit researchers from building on published findings. Every lecture could come with a clause forbidding the audience from getting smarter in a way that threatens the speaker. The book and article disclaimer from the opening of this piece isn’t a reductio ad absurdum. It is the logical conclusion of the legal theory these companies are asserting.

This is not how a free market works. You can buy a competitor’s car, study it, and build a better one. You can read a novel and write your own.

In AI, the clauses have gone too far.

Fair Use for Me, Not for Thee

The hypocrisy is staggering. These same companies spent years arguing that training their models on millions of copyrighted books, articles, and code snippets was “transformative fair use.” They won (or are still winning) that argument in court. Yet the moment someone wants to do something functionally similar — learn from their model’s outputs — they scream foul and hide behind a contract.

The result? A handful of well-funded labs get to build artificial moats around capabilities that cost them hundreds of millions or billions to develop. Everyone else is told to start from scratch or pay the toll. That is not competition. That is oligopoly dressed up as innovation policy. What they are really asking for is old AT&T status — a government-protected moat, with themselves as the permanent gatekeeper. AT&T needed a regulator to hand it that position. These labs are trying to get there through a clickwrap agreement.

And Congress is letting it happen.

Pulling Up the Ladder

The pattern is not hard to see. Having secured their position, the labs are using private contract law to pull up the ladder.

Just last month, on March 20, 2026, the White House released its National Policy Framework for Artificial Intelligence and handed lawmakers a blueprint. It talks about intellectual property, frontier models, and national security. It even nods to the distillation attacks by foreign labs. But it leaves the core domestic question untouched: Should private companies be allowed to use boilerplate contracts to block the very kind of knowledge transfer that has driven every previous technological revolution?

The answer is no.

Three Things Congress Should Do Now

Congress needs to step in and set clear national rules. Here are three concrete things lawmakers should do:

1. Clarify fair use for model outputs. Declare by statute that the systematic (but non-fraudulent) use of lawfully obtained API outputs for distillation or fine-tuning is a transformative fair use, just as training on publicly available books and web data has been held to be. Ban ToS clauses that attempt to override this.

2. Draw a bright line between legitimate protection and anti-competitive overreach. Allow labs to ban outright fraud (fake accounts, evasion of rate limits, geo-restrictions). But prohibit them from using contracts to forbid good-faith competitive learning from outputs they willingly sold access to.

3. Create a narrow, time-limited safe harbor for “model behavior.” If the labs want real IP-style protection for the emergent reasoning patterns in their models, let them ask Congress for it — with strict limits on duration and scope, just like patents. Don’t let them seize it through clickwrap agreements.

The labs will scream that without these restrictions, no one will invest the insane sums needed for the next frontier model. They made the opposite argument when they trained on copyrighted works — that restricting access would kill innovation. They were right then. They are wrong now. If the investment case genuinely requires monopoly control over knowledge transfer, Congress can debate public funding, tiered access models, or time-limited exclusivity — the same tools it has used in pharma and semiconductors. What it should not do is let private companies seize that protection through a clickwrap agreement nobody voted on. And while Congress is at it — the fair use question around training on copyrighted works deserves resolution too. Authors claiming that training on their work violates copyright are asserting rights the law was never designed to give them. Congress should clarify that as well.

We are at a rare moment. The White House National Policy Framework for AI has handed lawmakers a blueprint. The Supreme Court’s denial of cert in Thaler v. Perlmutter this month confirmed that AI outputs lack human authorship and cannot be copyrighted — leaving ToS as the only tool the labs have. Lawmakers can either rubber-stamp the status quo and let a few companies privately legislate the future of AI — or they can do what Congress is supposed to do: write clear, public rules that balance real investment incentives with genuine competition.

The choice is simple. Do we want an AI future written by corporate lawyers in clickwrap agreements? Or one written by the American people, through their elected representatives?

I’ve written before about how OpenAI was founded explicitly to prevent a handful of megacompanies from controlling AI — and then became exactly what it set out to stop (They Were Going to Save Us From This. Then They Became This.). The distillation ban is the enforcement arm of everything documented in that piece.

First you build the moat. Then you use private contract law to make sure no one can cross it. Blocking the advance of progress through monopolistic methods is not in the public interest.

It’s time for Congress to step in.

• Dario Amodei: The Self-Appointed Ethics Czar for Planet Earth — The Pentagon standoff and the gap between architectural choices and moral authority

• Time for Open Source Large Language Models — Why the AI cartel’s safety arguments are market protection dressed as principle

• They Were Going to Save Us From This. Then They Became This. — How OpenAI’s idealistic governance structure was the vulnerability Dario walked out through

• What Is Dario’s Lawsuit All About? — Why the Pentagon lawsuit is narrative management, not legal strategy

Act I — What Was Happening Inside OpenAI

By 2020, the knowledge of how to build large language models at commercial scale existed in exactly one place: inside OpenAI. Not because the research was secret by design — the original mission was open research — but because OpenAI had quietly stopped being open. GPT-3 was published as a paper but the model wasn’t released. The weights weren’t shared. The training details stayed internal. “OpenAI” had become a misnomer, and the outside world had no way to engage with what was actually happening.

Nobody else had the resources to find out independently. Training GPT-3 cost somewhere between $4 and $12 million in compute alone — a figure that ruled out every university lab, most research institutions, and all but a handful of well-capitalized companies. The knowledge existed in papers but the ability to act on it required a compute budget that essentially nobody outside of Google, Microsoft, and OpenAI could access. The field was not just small. It was captive.

Inside that captive world, two distinct research philosophies had been developing under the same roof. Ilya Sutskever — co-inventor of AlexNet and one of the architects of the GPT series — represented one approach: RLHF, Reinforcement Learning from Human Feedback. Train a model to satisfy human raters. Also opaque: the model learns to please without being able to explain why any given decision was made.

Dario Amodei and his team were developing the alternative. Constitutional AI: give the model an explicit set of principles and train it to reason against them in a self-critique loop. No armies of human raters required — the feedback is synthetic. The result is auditable by design. A compliance officer can examine the constitution. A regulator can trace a decision back to a principle.

This was not simply a safety disagreement — everyone in the field claims to care about safety. It was a technical argument about methodology. Dario’s team’s position was that Constitutional AI produces more predictable, more auditable, more consistent behavior than RLHF. A model trained against explicit principles behaves in ways you can reason about in advance. A model trained to satisfy human raters learns statistical patterns that are harder to predict in edge cases and impossible to fully explain after the fact. The enterprise case for Constitutional AI was not moral — it was engineering: if you can tell a compliance officer exactly what principles the model reasons against, and trace a decision back to a specific rule, you have a product that regulated industries can actually deploy with confidence.

Dario saw the trajectory clearly. He was VP of Research. He had complete visibility into every experiment, every training run, every capability threshold. He understood something that almost nobody in the media or investment community understood: the application code for a large language model is startlingly lean. A working GPT-2 fits in roughly 500 lines of Python. This is possible because AI development builds on vast open source libraries — PyTorch alone, the framework that dominates the field, represents millions of lines of highly optimized infrastructure that every lab uses equally. The proprietary layer on top is thin.

The intelligence is not in the software — it’s in knowing the process needed to make the model. His team carried in their heads the knowledge of what to write and the far more valuable intuitions about what data to use and which experiments to run. Reconstruction would take months, not years. This was not a leap of faith. It was a calculated extraction with a known timeline.

And the outside world still had no idea any of this was happening.

Act II — The Founding

In late 2020, Dario left. He took his sister Daniela, who ran Safety and Policy. He took a carefully selected team — not random disgruntled employees but the specific people required to reconstruct what OpenAI had built, for themselves, somewhere else.

Kaplan & McCandlish — Physicists who wrote the Scaling Laws: the mathematical rules governing how compute buys intelligence.

Tom Brown — Lead author of the GPT-3 paper. The engineer who could build what the theorists conceived.

Chris Olah — The world’s foremost mechanistic interpretability researcher. The person who could look inside a model and explain why it reasons as it does.

Jack Clark — Policy and narrative: positioning the work to attract capital and credibility.

Daniela Amodei — Operations and recruiting: building the company around the research.

Theory, engineering, interpretability, narrative, operations. Every component required to build and commercialize a frontier model. No redundancy. No gaps. And crucially — a team that had already solved the hardest problem in team-building. They knew how to work together. They shared a technical vision and philosophy, a common vocabulary, and years of accumulated experimental intuition. The “forming-storming-norming” phase that kills many startups was already behind them.

OpenAI’s governance structure — a nonprofit controlling a capped for-profit, with limited non-competes and a mission-driven culture that gave researchers enormous latitude — had no effective remedy. Nothing legally prevented a senior VP from walking out with his entire team. And so he did.

The “safety” framing was the story told publicly. The more complete explanation: Constitutional AI was a genuine technical belief, a cleaner corporate structure offered a clearer path to liquidity, and the Microsoft entanglement made OpenAI an increasingly uncomfortable home for researchers who wanted to control their own direction. “Safety” attracted mission-aligned early capital, created regulatory positioning, and repackaged a product argument — our alignment approach is auditable, theirs is not — as a values argument.

The sophisticated early investors — Eric Schmidt, Amazon, Google — were not buying a moral position. Schmidt said publicly he invested in the person, not the concept. Amazon put in $4 billion because Constitutional AI produces the auditable, consistent behavior their enterprise customers need. Google invested because they understood the technical methodology was genuinely differentiated and wanted access to it. They were buying a product architecture and a proven team. The investors who mattered read the footnotes.

The safety narrative made it legible to a broader audience.

“Safety” was a well-chosen word for that audience. The hallucination problem was real and visible: models that confidently made things up (hallucinated) looked unsafe to every enterprise buyer, every legal team, every regulator who touched them. Constitutional AI’s auditability spoke directly to that.

There is a technical dimension to this that conventional analysis misses entirely. In traditional software, the code is the moat — millions of lines of accumulated architecture that takes as long to rebuild as to build. That is why non-competes and trade secret law developed the way they did. Machine learning broke that model. The application code for a large language model is startlingly lean — a few hundred lines. Constitutional AI compounded this further: by removing the human rater bottleneck, it meant reconstruction required primarily compute and time, not infrastructure. The uncertainty was not “can we rebuild this?” — it was only “how much compute do we need?”

It is entirely plausible that the team arrived at Anthropic and had working prototypes running within weeks. They took nothing — no code, no weights, no data. They didn’t need to. Everything that mattered lived in their heads: the experimental intuitions built over years of running the same training loops, the architectural decisions they had already made and unmade, the failure modes they had already ruled out. Writing the code was a short order task for people who had already written it. The hard part — knowing what to write and what to feed it — was already solved.

The outside world still didn’t know any of this was happening. ChatGPT wouldn’t launch for another two years.

Act III — What Remained

The team that remained at OpenAI inherited the factory but lost the architects — and lost the technical rudder entirely.

Ilya Sutskever represented the competing methodology — RLHF — and was never part of Dario’s camp. Dario took everyone he needed. He did not take Ilya. After Dario left, what remained at the top was no coherent technical vision. Mira Murati, a capable product and operations executive, absorbed the research portfolio. The direction got filled by momentum and market pressure — not by anyone with the standing to set it deliberately.

Ilya tried to fight from inside. He voted to fire Altman in November 2023, reversed himself days later when 770 employees threatened to walk, was sidelined, and left in May 2024. By September 2024, Murati and the VP of Research had also gone. By late 2024 the original research leadership was gone entirely.

Ilya’s successor as Chief Scientist is Jakub Pachocki — a serious researcher who led GPT-4 and the o1 reasoning models. He is not a lightweight. Whether he and Chief Research Officer Mark Chen can generate the kind of integrated technical vision that Dario’s group had is an open question. OpenAI may be building that team. But it is a question mark, not an answer.

Anthropic’s seven co-founders are all still there — more than four years later. The real retention question is the second wave: researchers who joined in 2022 and 2023, fully vested or approaching it, who know exactly how the sausage is made. The co-founders have golden handcuffs. The second wave does not.

The Takeaway

The story most people tell is that Anthropic emerged from OpenAI’s shadow, played catch-up for years, and gradually proved itself as a serious competitor. That story is wrong.

Anthropic was not catching up to OpenAI. OpenAI was catching up to Anthropic. The knowledge, the team, the methodology, and the head start all belonged to the people who left — not to the organization they left behind. OpenAI spent the next three years reconstructing what had walked out the door in 2020, while Anthropic spent those same years executing on a plan that was already mapped before the company was formally announced.

That head start was built on a specific and unrepeatable set of conditions: captive knowledge inside one organization, a governance structure with no defection remedy, application code lean enough to reconstruct in months, and a methodology that removed the human rater bottleneck.

Dario did not leave because of safety concerns. He left because he had a complete team, a technical conviction, a cleaner corporate vehicle, and the precise understanding that reconstruction would take months not years. The “safety” narrative made it sound like a sacrifice. The technical reality made it look like a business plan.

By 2025, OpenAI had published a Model Spec and introduced Deliberative Alignment — both close peers to Constitutional AI in structure and intent. The field converged on Dario’s methodology without announcing it. In hindsight, the technical bet was the right one. The RLHF approach that Dario walked away from has been quietly supplemented or replaced across the industry by variations of what he built. Dario won the argument. He just won it in a way that made the methodology everyone’s rather than anyone’s.

That is the story of how the AI industry’s defining rivalry actually began. Not a moral crusade. Not a safety emergency. A technical schism, a governance gap, and a calculated extraction — executed before the world knew the race had started.

One final observation. In 2020, the lean codebase made reconstruction fast for a team that already knew what to write. In 2026, agentic coding tools have advanced to the point where anyone who understands what to build can direct an AI to build it — even a million-line codebase, if you know what to ask. The execution barrier is gone. What remains is the knowledge of what to ask: the experimental intuition, the architectural judgment, the understanding of which training decisions matter and which don’t. That knowledge still lives in heads, not in repositories. It always did.

Even the models themselves now understand a great deal of what to do at a high level.

The moat may now just be a puddle to step over.

This landed in my inbox today — a Foreign Affairs subscription pitch. The lead article screams “How America’s War on Iran Backfired.” Subtitle: “Tehran Will Now Set the Terms for Peace.” Next to it, another piece on the Hormuz “minefield” where Iran supposedly holds all the cards. They even threw in a third about Europe fracturing because the war is so messy.

The Playbook Was Already Written

In “Nation Building: You Can’t Build What They Don’t Want,” I said the foreign-policy priesthood has spent eighty years stapling reconstruction fantasies to the back of every war and then acting shocked when the locals don’t line up for democracy class. The lesson wasn’t that force fails. The lesson was that force paired with nation-building only worked once, in 1945, and the establishment has been chasing that result ever since.

In “Iran: The Bill Has Come Due,” I laid out the actual playbook: stop asking how to make Iran behave. Start asking how to make it impossible for Iran to afford the behavior. Pull the financial pin—Kharg Island first, then the shadow fleet, then the blockade if they keep playing games. One mechanism, four problems solved. No occupation. No reconstruction conferences. Just bankruptcy.

In the follow-up, “Will the US Seize Iranian Oil and Natural Gas Assets?,” I noticed the 2,500 Marines climbing aboard the USS Tripoli amphibious ready group and said out loud what everyone in uniform already knew: the plan might not be to destroy the oil. It might be to take it. Geography handed us a gift—Khuzestan province, the Ahvaz supergiant, Kharg Island, all sitting in a neat little triangle with the Zagros mountains blocking Iranian reinforcements from Tehran. The March 13 strikes proved it: they hammered every military target on Kharg but left the terminals, tanks, and pipelines untouched.

Burning the Deterrent

And here we are, March 20, with the Tripoli group closing in and the regime still dumping what’s left of its missile inventory on Dubai hotels, Riyadh refineries, Gulf civilian airports, and Israeli cities.

They are burning their own deterrent on secondary targets while the real prize sits undefended.

This is the part the news media refuses to see. Iran didn’t save its best missiles and fast boats for the Marine landing that’s coming. It fired them at shopping malls and runways in countries that were cautiously trading with it six weeks ago. The neighbors aren’t feeling “backfire” sympathy. They’re feeling rage. UAE officials talk about a “trust gap that could last decades.” Saudi princes call it blackmail. Even the Qataris—who used to play both sides—are done. Iran just torched whatever residual goodwill it had left in the Gulf.

Meanwhile the clerics are watching their navy sink, their launchers get bunker-busted, and their missile inventory go from 2,500 to functionally zero in two weeks. They are not setting the terms for peace. They are brawling and about to gas.

The clerics aren’t blind. They can see the Tripoli on the same charts we can. But revolutionary regimes don’t survive by being rational; they survive by performing defiance. So they fire. It’s the scorpion on the frog’s back — they know what the sting costs, and they do it anyway. It’s not strategy. It’s their nature.

The Real Move

The three-card monte is this: keep everyone staring at “regime change” or “forever war” or “escalation with China” while the real move happens offshore. The Marines are not going to Tehran or to make the Hormuz Straits safer. They are going to Kharg Island and the Khuzestan fields. They don’t need to occupy a country of 90 million people or the coastline. They only need to secure a 200-mile triangle that contains the third-largest oil reserves on Earth and the export terminal that moves 90 percent of it. And any column Iran sends across the Zagros to take it back becomes a target. Open terrain, no cover, US air dominance. They’d be destroyed before they reached the foothills.

Once those pumps are under new management, the regime’s ability to fund themselves, Hezbollah, the Houthis, Hamas, and the nuclear program evaporates overnight. The oil doesn’t stop flowing—it just stops flowing to the clerics. China still gets its crude; it will simply buy it from whoever is running the terminal instead of from a sanctioned regime. Global markets stabilize. And the neighbors who just got rocketed get to watch Tehran finally learn that actions have consequences.

The Off-Ramp

The clerics still have an off-ramp. Reopen Hormuz. Verifiably cut the proxies. Walk away from the nuclear weapon path. Accept that the oil infrastructure will operate under international oversight with revenue sharing. They can keep their thrones in Tehran. They just can’t keep exporting violence.

But every day they wait, the missile and drone stockpiles shrink, the Tripoli gets closer, the sympathy gets thinner, and the terms get worse.

The Hormuz mining threat is another waste of resources. It won’t protect their income stream. It’s just defiance and theater.

This isn’t complicated. It’s not nation-building. It’s not regime-change theater. It’s exactly what I laid out two weeks ago: make aggression unaffordable. Whether by destroying the revenue stream or by taking the revenue stream, the result is the same. The bill has come due.

Regime change may come. It may not. Either way, it’s beside the point — a sideshow the media will obsess over while the real game plays out.

And the house—Kharg Island, Ahvaz, the whole southwestern prize—is about to change hands.

Fold or Lose

Fold or lose. The choice is still theirs.

For about seven more days.

In 1930, France began construction on the Maginot Line — a technically sophisticated, enormously expensive chain of fortifications along the German border. It was the product of genuine engineering brilliance and absolute strategic confidence. France had studied the last war, identified the threat, and built the perfect defense against it. The Germans went around it through Belgium. France fell in forty days.

The instinct that built the Maginot Line is still alive. It just raised $1.03 billion.

Europe Missed the Wave

The large language model revolution was built in San Francisco, London, and a handful of other American and British institutions. Not Paris. Not Berlin. Not anywhere in continental Europe. OpenAI, Anthropic, Google DeepMind — the companies defining the trajectory of AI are American. The infrastructure running them is American. The capital funding them is overwhelmingly American.

Europe watched this happen. And Europe’s response, as usual, has been to regulate what it cannot build.

The one genuine exception is Mistral. A French company, founded by researchers who left Google DeepMind and Meta, building competitive large language models that punch far above their weight given their funding constraints. As I documented in Why Mistral Is Cash-Starved in the LLM World, Mistral is not underfunded by European standards. It is cash-starved by frontier AI standards. Europe has the money. It simply hasn’t decided that backing a European LLM lab at the scale needed is a risk worth taking.

Putting this money into Mistral would make a lot of sense. It could surely become the dominant AI platform for all of Europe — actual product, actual users, actual revenue, actually in the fight. Europe had that option. They started late but they have the home field advantage there.

Doing something realistic that has a good chance to succeed is not good enough.

It chose AMI instead.

The Statement

AMI is a statement about the present.

The statement is this: the entire LLM wave was wrong. Silicon Valley built the wrong thing. France — with its French Turing laureate, its Paris headquarters, its presidential endorsement — will build the right thing.

This is a more coherent explanation of AMI than any technical argument. It explains why the money went to a three-month-old company with no product, no demo, and a part-time chairman who does not appear to be leaving New York. It explains why Macron (a caricature and the very embodiment of French exceptionalism) endorsed it personally. It explains why Bpifrance, Dassault, and the Mulliez family wrote checks. It explains why the framing is explicitly “a credible frontier AI company that is neither Chinese nor American” — LeCun’s own words.

Hidden in this bravado is low confidence and fear of failure. France would rather fail gloriously than lose incrementally. If AMI fails, it fails on its own terms, reaching for something nobody else had the vision to attempt. If Mistral falls behind OpenAI, it is just another company that lost. France would rather be Icarus than runner-up.

Is the Proposition Realistic?

We examined this in detail in LeCun’s AMI: What is the Proposition?. LeCun’s actual track record — the Bell Labs team work built on prior art, the post-1998 field that moved without him, thirteen years at FAIR — does not suggest someone capable of the astounding feat of rethinking AI from scratch and besting the entire global research community. The credentials are real. The forecast they imply is not.

The theoretical foundation of AMI’s bet is also examined in the companion piece What Evolution Tells Us About the Path to AGI. The “stochastic parrots” argument, associated with Bender et al., turns out on examination to be a restatement of the same assumptions that failed in the symbolic AI era. The brain is also a pattern-matching system. Evolution built general intelligence without symbols, causal models, or explicit world representations. The deficiencies in LLMs keep shrinking through engineering, exactly the way nature refines. The argument that the current architecture has a ceiling requires believing something evolution already disproved.

Skin in the Game

Before examining whether AMI’s bet makes sense, there is a prior question worth asking: does Yann LeCun actually believe it?

The most honest answer to that question is not found in his interviews or his X posts. It is found in what he is doing with his life.

LeCun holds the Jacob T. Schwartz Chaired Professorship in Computer Science at NYU’s Courant Institute of Mathematical Sciences. This is not an honorary title or an emeritus position. It is an active faculty role — with graduate students who depend on him, a research group, and teaching responsibilities. He is not winding down his academic career to throw everything into AMI. He is keeping it fully intact. The students are still there. The office is still there. New York is still home. As far as we can tell, none of that is changing.

LeCun is sixty five years old. Is he going to spend another fifteen years working around the clock in a startup? They don’t grow by themselves. He is the spiritual leader.

Zuckerberg left Harvard at 19. Gates left Harvard at 20. Page and Brin abandoned their Stanford PhD programs. Jobs dropped out of Reed College. Sam Altman dropped out of Stanford. Elon Musk famously dropped out of Stanford's PhD program after two days. Every one of them did it against rational advice, against the safe path sitting right in front of them. Their parents almost certainly told them to finish school first. They were driven beyond reason. That is what genuine conviction looks like — not a calculated bet, but an obsession that burns every available bridge.

LeCun has burned nothing. The tenure is intact. The New York life is intact. The someone-else-runs-it structure is intact. If AMI succeeds, he claims the scientific vision. If it fails, LeBrun ran the company, the research needed more time, investors were impatient.

The people writing the billion-dollar checks should sit with that.

A Startup Is Not a Research Lab

The second problem is structural. AMI is asking a startup to do what startups cannot do: fundamental research on an open-ended timeline.

Research is like entering a tunnel of unknown length that may not even have an exit. It is dark and you can walk for ages and find nothing. And if you do find the exit, you may emerge into a forest with a fork in the road leading in ten different directions. That is not a flaw — it is the nature of the thing.

A startup has a clock. The burn rate is real. At some point the investors ask when the tunnel ends. “We don’t know — that’s the nature of the tunnel” is not a satisfying answer to people who wrote nine-figure checks. Japan’s Fifth Generation Computer project made the same bet in the 1980s — massive government funding, bold proclamations, fundamental rethinking from first principles. It died quietly a decade later having produced almost nothing. The hype was atmospheric. The results were not.

The Maginot Line

Wish LeCun well. His early contributions are real. But a $3.5 billion valuation for a small team with no product, no demo, a part-time chairman who does not appear to be leaving New York, a visual-only architecture with no language capability, and a thesis his own employer sidelined after thirteen years — that is not a scientific venture.

It is a monument to French certainty that what AI was missing was the French.

The First Dead End

The early days of artificial intelligence were dominated by a reasonable-sounding idea: figure out how thinking works, then write code that does it. Build explicit rules for vision. Hand-craft logic for reasoning. Encode knowledge as structured facts. This approach — now called Good Old-Fashioned AI, or GOFAI — ran into two walls simultaneously. The problems were too complex to specify by hand, and even partial solutions couldn’t scale. Someone had to write every rule. The combinatorial explosion of the real world crushed every attempt.

So researchers tried something different. Instead of designing intelligence from the top down, what if you imitated the structure that already produces it? The human brain — and to varying degrees, every animal nervous system — is a network of interconnected nodes that fire, adapt, and reorganize based on experience. Crude artificial versions of this, neural networks, had been theorized since the 1940s. The bet was simple: if the real thing works, a sufficiently good imitation might work too.

This is, at its core, an act of faith. But it is faith grounded in the most robust existence proof available. General intelligence exists. We know how it was built. It was built by evolution — by billions of years of random mutation, brutal selection pressure, and the gradual accumulation of structures that worked well enough to survive. No designer. No blueprint. Just variation and time.

The underlying assumption is worth stating plainly: general intelligence is an emergent property of sufficiently powerful predictive systems. That is not proven. But neither is its opposite. And evolution has already run the experiment once.

Why Does AI Need So Much Data for Training?

One of the most common dismissals of LLMs goes like this: a child can recognize a cat after seeing a handful of examples. An LLM needs to be trained on vastly more. Doesn’t that prove the architecture is inferior — that something essential is missing?

It proves nothing of the sort. The child is not starting from zero. Before that child sees a single cat, evolution has already loaded millions of years of visual processing capability into the hardware — object permanence, edge detection, motion tracking, pattern recognition, depth perception. All of it pre-installed, refined across countless generations of survival pressure. The child isn’t learning to see from scratch. The child is running a query against an extraordinarily sophisticated system that took billions of years to build.

LLMs start from zero. No pre-loaded visual cortex. No inherited reflexes. No evolutionary head start. They compensate with volume — training on the accumulated written output of human civilization, which is itself the encoded product of all that evolved intelligence. It is not less data than evolution used. It is the same data, stored differently — in text rather than in genes and synapses.

The comparison isn’t LLM versus child. It’s LLM versus the entire evolutionary lineage that produced the child. And that lineage has been running its training data for billions of years — almost certainly more than any current LLM has processed. The difference is that evolution compressed all of it into biological structures: DNA, neural architecture, instinct, reflex. Compact encodings of an unimaginable training run. Machine learning does something analogous with weights and embeddings — sparse representations that capture the essential structure of vast input. Different medium. Same principle.

The Existence Proof

The brain that resulted is not a precision instrument. It is an accumulation of evolutionary hacks — older structures repurposed, newer ones layered on top, the whole thing running on roughly 20 watts. It is, by any engineering standard, a mess. And yet it generalizes across domains, handles novel situations, reasons under uncertainty, and produces consciousness — or at least something that feels exactly like it from the inside.

Neural networks, even primitive ones, share something essential with this architecture: distributed representation, learned weights, emergent behavior from simple units operating in parallel. The neuroscientist David Marr argued in 1982 that understanding any information-processing system requires analysis at multiple levels — what it computes, how it computes it, and what physical substrate runs it. His point was that the implementation level doesn’t determine the computational level. At the computational level, biological brains and artificial neural networks are doing something recognizably similar: finding patterns in input and using those patterns to predict and act.

Critics of modern AI often invoke “mere pattern matching” as a dismissal. The phrase assumes that pattern matching is categorically different from — and lesser than — genuine understanding. But this is precisely what evolution calls into question. The brain is a pattern-matching system. Its neurons fire based on weighted inputs; its architecture was selected because certain patterns of activation produced behaviors that survived. Whatever understanding feels like from the inside, the mechanism underneath is not categorically different from what large neural networks do. It is more complex, certainly. Categorically different, probably not.

Neuroscience is increasingly explicit about this. An influential framework in computational neuroscience — predictive coding, associated with Karl Friston’s free energy principle — describes the brain as a prediction machine: a system that continuously generates probabilistic models of sensory input and updates them based on error signals. In other words, the brain is doing Bayesian inference on a massive scale. If that’s what “genuine understanding” looks like under the hood, the “mere statistics” dismissal doesn’t just miss the point — it describes the thing it’s trying to dismiss.

If evolution’s messy, unguided process produced general intelligence from a neural substrate, the burden of proof falls on anyone who claims a different substrate cannot do the same.

The Critique That Proves the Point

The dominant critique of large language models today runs something like this: LLMs are sophisticated statistical engines — they predict the next token without any genuine understanding of the world. No model of physics. No cause and effect. No real reasoning. Just pattern matching at enormous scale. The argument was formalized in a 2021 paper by Emily Bender et al., which coined the term “stochastic parrots” — systems that manipulate linguistic form without any access to meaning. The conclusion drawn, then and since, is that you can scale this forever and never reach AGI because the foundation is wrong.

In my opinion, this critique is a restatement of the GOFAI program in different language. The ingredients the stochastic parrot camp says LLMs are missing — grounding, causal models, explicit world representations — are precisely what symbolic AI researchers spent decades trying to hand-code. It didn’t work. The problems were too complex to specify, and there was no path to scale. Dressing those requirements up as a critique of neural networks doesn’t solve either problem; it just relocates the failure.

This critique is a serious argument made by serious people. It is also, on examination, an extraordinary claim. It asserts that the researchers making it know better than four billion years of selection pressure what intelligence actually requires. They have a theory — real intelligence needs symbols, or embodiment, or causal models, or something the critics can gesture at but rarely fully specify — and they’re willing to bet that theory against the only working example of general intelligence we have.

There is a word for the belief that intelligence requires something extra, something that mere physical processes operating on simple units cannot produce. That word is not “neuroscience.”

Nature’s Method, Compressed

Modern machine learning has not stayed static while critics catalogued its limitations. Retrieval-augmented generation gives models access to external memory — something evolution solved with the hippocampus. Multi-step reasoning chains externalize working memory — something evolution solved by expanding the prefrontal cortex. Tool use and agentic architectures let models act on the world and observe consequences — something evolution solved by connecting nervous systems to bodies over hundreds of millions of years.

None of these additions required inserting a theory of intelligence. They required observation, experimentation, and iteration. The same process nature used, compressed by engineering and running on a different substrate.

The empirical record is hard to argue with. A decade ago, the list of things neural networks couldn’t do was long and confident: they couldn’t hold a conversation, write coherent prose, generate images from descriptions, solve novel mathematical problems, pass professional licensing exams. The list is shorter every year. The velocity of improvement is not slowing.

You don’t need a PhD in AI to see it. Today’s LLMs may not be AGI — but they are remarkable, and anyone who has spent serious time with them knows it. The proof is in the pudding. Whatever these systems are doing, it is working, and it is getting better.

And the deficiencies keep shrinking. Not by throwing out the foundation and starting over, but by tweaking the basic mechanism — exactly the way nature does it. Hallucinations reduced. Reasoning improved. Memory added. Context expanded. Each problem that was supposed to prove the architecture was fatally flawed turned out to be an engineering problem with an engineering solution. Nature never scrapped the neuron and started over. It just kept refining.

The Theological Argument You Didn’t Know You Were Making

To say that LLMs cannot become AGI is to make a specific claim: that the architecture is fundamentally wrong. Not immature, not incomplete — wrong at the foundation. That no amount of scale, data, or architectural refinement will get you from here to general intelligence.

But consider what that claim requires you to believe. The human brain began as a simple neural tube in primitive vertebrates. It was not designed. It had no roadmap. Through random variation and selection, it became what it is now — capable of language, mathematics, art, science, and every other cognitive achievement we recognize as distinctly human. At no point in that process was the foundation “right.” It was always a hack, always imperfect, always good enough to survive and reproduce.

To say that the neural network approach cannot reach AGI because it lacks something essential is to say that the brain could never have become what it became — that at some point in evolutionary history, an observer would have been justified in saying: this architecture has a ceiling, and general intelligence is above it.

That observer would have needed a designer to explain where general intelligence actually came from.

Nature is not intelligent. It is random. But given enough time and selection pressure, randomness produces intelligence. We have proof. It’s reading this sentence.

The only question worth arguing about is how long it takes us to compress four billion years of evolution into something we can run in a data center. The evidence so far suggests: faster than anyone expected.

The holy grail is AGI (artificial general intelligence). What that exactly means is not entirely clear, but it is something closer to what a person can do — reasoning, planning, understanding the world, adapting to new situations without being explicitly programmed for each one.

The proposition from Yann LeCun is that current AI with LLMs is fundamentally a dead end in the search for AGI.

LeCun argues that LLMs are sophisticated statistical parrots — they predict the next token without any genuine understanding of the world. No model of physics. No cause and effect. No real reasoning. Just pattern matching at enormous scale. His contention is that you can scale this forever and never get to AGI because the foundation is wrong.

Yann LeCun unveiled his new startup on March 10, 2026: Advanced Machine Intelligence Labs (AMI), in which he will replace current LLMs with “world models”. $1.03 billion seed round at a $3.5 billion pre-money valuation — one of the largest seeds ever. Paris headquarters. Twelve people. CEO Alex LeBrun. LeCun as executive chairman.

The current status of LLMs: Pretty amazing and getting more amazing by leaps and bounds at an accelerating rate.

The current status of “world models”: No product. No demo. No proof of concept. Just promises of “world models”: AI that understands the physical world, reasons, plans, anticipates outcomes.

On the surface this is a rather bold claim. LLMs are getting better all the time and aren’t currently showing any visible signs that the technique is fundamentally limited. In effect, LeCun is arguing that mostly all of the current AI industry is pursuing the wrong paradigm — one that can never lead to AGI.

This op-ed is in two parts. Part 1 examines why people might think he can do this and does it seem realistic. Part 2 is a deeper look into who is investing in this and what that tells us.

A note on the “LLMs can’t do math or physics” critique: it is incorrect to say that deployed LLMs do not handle math and physics. The systems as packaged for actual use consult math engines, code interpreters, and symbolic reasoning tools. The criticism describes the raw model in isolation, not the deployed system. Nobody ships the raw model alone. MCP servers (Model Context Protocol) now provide a standardized way to connect LLMs to virtually any external service or specialized engine — the scope keeps expanding with no sign of a ceiling. My general sense of many people who criticize LLMs is that they have never used these tools in a serious way, or they are talking about how they were four years ago.

How Science Actually Works

The public model of science is: lone genius has insight nobody else has, works against the skeptics, proves them wrong, changes everything. That model is a fairy tale. Science is slow, collective, and incremental. Breakthroughs happen when the tools are finally ready — and when they are, many people working independently arrive at essentially the same place at essentially the same time.

You can find a new insect in the Amazon. That’s a solo discovery. But something fundamental — the kind of advance that reorganizes how a field thinks — is almost never that.

Even the most celebrated names in science look different when you actually read the history. Newton and Leibniz developed calculus independently and simultaneously. Hooke had gravitational inverse-square ideas before Newton formalized them. Newton himself acknowledged standing on the shoulders of giants.

The famous mathematician Gauss’s most celebrated achievement is the Fundamental Theorem of Algebra (this is a complex proof — the word algebra here is not the algebra you had in junior high school). Before Gauss, d’Alembert, Euler, and Lagrange had all attempted proofs — all incomplete. Gauss first published his own proof in 1799, but it was also incomplete with gaps. The first rigorous proof was published in 1806 by Argand — an amateur mathematician. Gauss kept producing versions of the proof for decades afterward.

Not to digress, but sometimes the work is done by people the field refuses to celebrate. An amateur mathematician finally solving a problem that d'Alembert, Euler, Lagrange, and Gauss had all attempted and failed to fully crack — that is not the history the credentialed class wants to tell. Argand's name is not the one anyone remembers.

The lone genius is a story we tell afterward, once the crowd has been edited out of the frame.

The Modern Genius

Einstein is the archetype of the lone genius in the public imagination — the patent clerk who rewrote physics. The reality, as usual, is more complicated, and I am not going to wade into the priority disputes around special relativity here. There are legitimate questions — Henri Poincaré and Hendrik Lorentz had developed key mathematical elements before Einstein’s 1905 paper, the Lorentz transformations that are central to special relativity are named after Lorentz because he derived them first, and Einstein’s paper notably did not cite either of them. Serious historians of science have written about this. It is also worth noting that Poincaré was a mathematician, not a physicist — and as with Argand solving what the famous mathematicians could not, the field has a long history of being reluctant to credit people who come from the wrong discipline. The consensus view is that Einstein’s formulation was more complete and more physically unified, so the credit is broadly defensible — but it is not clean. That debate can go on without us.

What is not disputed is what happened after. Einstein spent the last thirty years of his life pursuing a unified field theory that went nowhere. He rejected quantum mechanics — “God does not play dice” — against overwhelming experimental evidence and the judgment of virtually every physicist working in the field. He sailed boats. He made funny faces for photographers. He became the world’s most famous scientist and produced no significant new physics after the mid-1920s. The legend kept growing while the work stopped.

Consider one more irony. The Nobel Prize — the institution most synonymous with recognizing lone genius — was itself named after a system integrator. Alfred Nobel didn’t discover nitroglycerin; that was Ascanio Sobrero, his former teacher. He didn’t invent the safety fuse; that was William Bickford. He combined existing components into a commercial product. Useful. Valuable. Not a lone genius moment. But the prize bearing his name spent the next century teaching the world that science works the other way. That myth is now so embedded in how we reward science that we can’t see the crowd behind the name.

LeCun, Hinton, and Bengio won the Turing Award. What does that mean? A committee of the ACM decided to give it to them. That does not mean they uniquely did this work or were even the first. Why they gave it to them and not others is not entirely clear. When investors hear “Turing Award winner” they process it as “certified genius who sees what others can’t.” That is not what it means.

To be fair, the trio did make contributions to the field and literature, they helped popularize machine learning and were true believers during the AI winter — the long decades when the field was defunded, dismissed, and ignored. That persistence was real and it mattered. But what actually unlocked modern AI wasn’t a scientific breakthrough by any of them. It was the GPU, invented by Nvidia to meet the computing demands of video games. When that hardware became available, the algorithms that had been waiting for it finally had somewhere to run. The tide came in. Everyone standing on the beach got credit for the ocean.

One Name on a Crowd’s Work

Let’s examine what LeCun actually did, because the press never bothers with the details. The work associated with his name is computer vision work using a convolutional neural network published with a team of three other researchers at Bell Labs.

The convolutional architecture at the heart of this work wasn’t developed by the team LeCun worked with. Fukushima’s Neocognitron in 1980 had the core convolutional architecture — approximately eight years before the Bell Labs team applied backpropagation to train it.

The backpropagation algorithm itself was Rumelhart, Hinton, and Williams’ (1986) — except it wasn’t really theirs either. The underlying mathematics had been used in control theory for aircraft in the 1960s by Bryson and Ho. Werbos had applied it to neural networks in his 1974 PhD dissertation. Linnainmaa had described the algorithm in his 1970 master’s thesis. Rumelhart, Hinton, and Williams popularized it in their 1986 Nature paper, which is what the Bell Labs team actually cited and built on. But even that algorithm had decades of prior history behind it. The team borrowed from people who had themselves borrowed.

The foundational 1998 paper “Gradient-based learning applied to document recognition” has four authors: LeCun, Bottou, Bengio, and Haffner. Bottou has spent decades making stochastic gradient descent practical at scale. What the paper demonstrates is backpropagation applied to a convolutional architecture to make it trainable on real-world data. That is the contribution. Whose specific idea it was from the Bell Labs team to train a convolutional network using backpropagation, or if it was collaborative as the multiple authorship presumably states, and whether others were already working on the same combination, the paper does not tell us. But synthesis on top of other people’s components, executed by the Bell Labs team, at an institution that provided everything required — that is not the lone genius narrative.

It is also worth asking: how many people in the entire world were working on this problem at that time? Dozens. Maybe a few hundred at the outer edge. With the tools that existed, someone was going to get there.

And that count only includes the work anyone knew about. This was before the internet, before electronic journals, before preprint servers. Work done in Soviet labs, Japanese universities, and Eastern European institutions frequently never made it into Western publications. The Russians developed photocopying technology before the West — the KGB suppressed it because it would enable citizens to freely reproduce and distribute information, undermining state control over what people could read and share, and the West later “invented” it independently. We have no idea how much parallel work was happening in 1988 in places Western researchers never read. Now there are millions of researchers working on AI globally, everything is published instantly, and the competition is total.

After that: AlexNet — Krizhevsky, Sutskever, Hinton. VGG, ResNet — not LeCun. YOLO — Redmon, not LeCun. Vision Transformers — Google Brain, not LeCun. Diffusion models — not LeCun. The entire modern computer vision stack that runs self-driving cars, medical imaging, and manufacturing inspection was built by other people on architectures other people invented.

The field he claims as his life’s work, originally as part of a four-man team at Bell Labs, has been shaped for decades primarily by other people. He helped build one of the early versions of computer vision and watched other people build far better ones ever since. Unlike the fewer than one hundred people worldwide working with primitive computing when he did his foundational work, there are now millions of PhDs with incomprehensible amounts of compute dedicated to solving these problems. As the field of computer vision grew and millions came to work on it, his name disappeared from the cutting edge results. Others took that position.

The transformer architecture — the technical foundation of every major AI system today, including the LLMs LeCun dismisses — was published in 2017 in “Attention Is All You Need” by eight authors from Google Brain and Google Research. Nobody calls it the “Vaswani paper.” It is just the transformer. Eight people, one institution, collective work. That is how modern AI actually gets built.

He Predicted Them Dead. They Keep Getting Better.

Since at least 2022, LeCun has posted relentlessly dismissing LLMs. Hallucinations prove they can’t reason. Scaling won’t get to intelligence. The architecture is fundamentally wrong. His critiques predate ChatGPT — but after November 2022, when a startup nobody had heard of dropped ChatGPT and overnight became the face of AI, the dismissals got louder. Sam Altman, not even an amateur scientist but a businessman and promoter, on every magazine cover. Not LeCun. Not FAIR. Not Meta. For someone who had spent decades as a leading voice in AI, watching a startup he’d publicly dismissed become the most consequential technology company of the decade had to sting. Meanwhile OpenAI and Anthropic keep releasing more stunningly clever models with greatly diminished issues such as hallucination. Scientists with famous names are as capable of envy as anyone else. Einstein did not even cite the obvious work that special relativity built on. Sour grapes doesn’t mean the critique is entirely wrong. But when someone with a competing thesis and a bruised ego tells you the competition is fundamentally broken, you apply a large discount.

I use these systems daily to write these pieces. My business uses them as the primary coding agent. I don’t see any indication of the models reaching some kind of fundamental limitation. The limitations become known and then they get solved, and at this point, much faster than I could have envisioned.

LeCun’s core claim is that LLMs cannot develop genuine understanding of the physical world. Where is his proof? Because they hallucinated a lot four years ago? Or is it just NIH (Not Invented Here)?

LeCun spent thirteen years at Meta running FAIR — unlimited compute, top researchers, no product pressure. The output: papers, open source code, and a visual encoder with no language capability. Meta’s actual consumer AI runs on Llama — a large language model, the thing LeCun publicly calls a dead end. His own employer voted with their product roadmap against his thesis. If this research direction was going to produce something deployable, why didn’t it in thirteen years with those resources?

The Institutions That Built the Legend

The ACM and the press built a wall of credibility around these names that nobody with a checkbook knows how to look behind. A Turing Award is not a forecast. It is a receipt for work done decades ago, which is often not even for scientific discovery but for championing something. Investors read it as a prediction. It isn’t.

And frankly, I don’t see any reality in which AMI is more than LeCun moving his pet research project from Meta to a new address. Google, DeepMind, and others already have teams working on world models and visual representations. Nobody — not one of them — is showing a realistic prototype or proof of concept. The field isn’t holding back because it lacks a French headquarters and a Turing laureate. “World models” for AGI is a speculative research direction at this point, and many such ideas never amount to anything.

In Part 2 we will look more deeply into who is actually writing the checks and why. But it’s worth noting one thing before we get there: Mark Zuckerberg is not on the investors list. Whether Meta was offered the chance to invest and passed, or was never approached, we don’t know. What we do know is that the man who funded LeCun for thirteen years, who knows his work better than any outside investor, is not in this round. There is also good reason to think that Meta, having committed fully to Llama and the LLM path, was ready for a new AI leader — one who wasn’t publicly dismissing their core product strategy.

.

Two pieces ago, in Nation Building: You Can’t Build What They Don’t Want, I wrote this: “It may be that the West needs to stop pretending and simply occupy the resource-critical portions of these regions, controlling how those resources are extracted and allocated.” I called it the honest conversation we wouldn’t have — the option that’s too uncomfortable to say out loud, so we dress up resource interests in the language of liberation and democracy promotion instead.

In my last piece, Iran: The Bill Has Come Due, I argued for a different approach: don’t occupy, destroy. Cut the oil revenue and you cut everything it buys — missiles, proxies, centrifuges — simultaneously. A staged campaign starting with Kharg Island. No nation-building attached. Pull the financial pin and let the programs collapse on their own. I closed with this: “What they haven’t done is be honest with the American public that this is the plan.”

Then the United States put 2,500 Marines aboard a three-ship Amphibious Ready Group on a slow boat to the Persian Gulf.

Nation-building talk is theater. Nobody credible in Washington is planning to rebuild Iran — Afghanistan and Iraq are permanent reminders of what that costs and produces. Regime change rhetoric is for the cameras. What military planners actually do is look at assets, objectives, and access routes.

And when you look at what 2,500 Marines aboard a three-ship Amphibious Ready Group are designed to do — seize coastal objectives, secure ports, hold fixed infrastructure — the question shifts.

The plan may not be to destroy Iran’s oil and gas business. It may be to take it — or to threaten seizure so credibly that Iran has to treat it as real.

That uncomfortable conversation I said we wouldn’t have? We may be having it right now — with actions instead of words. Whether the Marines land or the threat alone does the work, the leverage is the same. Iran has to plan for the worst case. So does everyone else.

The Prize Is Stacked in One Corner

Iran holds the third largest proven oil reserves on earth — roughly 150 billion barrels, enough at current production rates to last 145 years. But reserves in the ground are a geology story. What matters operationally is where the producing infrastructure sits.

Here is the strategic geography that nobody on television is explaining: the vast majority of Iran’s crude oil reserves are concentrated in the southwestern Khuzestan province, pressed against the Iraqi border to the west and the Persian Gulf to the south. The Ahvaz field — third largest in the world — holds an estimated 65 billion barrels, roughly 23% of Iran’s total reserves. Marun, Gachsaran, and Agha Jari round out the supergiant tier. All of them in the same province. South Pars — the largest single gas field on the planet — sits offshore in the Gulf.

The entire productive complex — fields, processing plants, pipeline network, and the Kharg Island terminal that handles 90% of Iran’s oil exports — fits within a rough triangle: Ahvaz to the north, the Iraqi border to the west, the Gulf coast to the south. Call it 200 miles on a side.

Khuzestan is geographically separated from the Iranian heartland by the Zagros mountain range. Tehran is roughly 400 miles away on the other side of those mountains. This is not an accident of geology. It is an operational gift.

Iran Spent Its Deterrent on the Wrong Targets

Two weeks ago an amphibious assault on Kharg or a ground advance into Khuzestan would have faced a serious Iranian military response. Today the math has shifted dramatically, and Iran’s own decisions explain why.

Iran entered this war with approximately 2,500 ballistic missiles. After ten days, according to IDF and US military assessments, roughly 2,410 had been fired and over 60% of launchers destroyed. The daily launch rate collapsed 92% — from 480 on day one to 40 by day ten. That is not rationing. That is structural failure. The US used bunker buster bombs to seal the entrances of Iran’s underground missile storage facilities, leaving hundreds of missiles entombed. The inventory exists. It cannot be reached.

Iran’s navy is severely degraded — 43 vessels destroyed or damaged. The fast-attack boat swarm that formed the core of Iran’s Persian Gulf anti-access doctrine has been eliminated. What remains is a drone capability — real and persistent, but an attrition weapon, not one that stops a ground advance or defends a fixed perimeter.

Iran fired its deterrent at airports in Dubai, hotels in Riyadh, and US bases across the Gulf. It spent what it had harassing the neighborhood instead of holding it in reserve for the scenario now on the table. The plain English is: they wasted their bullets.

This is a classic military mistake — one that looks like sound strategy to non-military observers and reveals itself as catastrophic to anyone who has studied military history. Targeting civilian infrastructure feels decisive and that it will destroy the will of the people to resist. It signals resolve. What it almost never does is win wars or preserve military capability for the fights that actually matter. Iran had a sophisticated anti-access capability purpose-built for the Persian Gulf. They spent it sending messages instead of stopping ships. Now the ships are coming.

The Operational Logic

The March 13th strikes on Kharg made the direction of travel explicit. CENTCOM destroyed 90+ military targets — mine storage, missile bunkers, IRGC facilities — while explicitly sparing the oil infrastructure. Destroy the defenses, leave the pumps running, unlock the front door. That is not punishment. That is a turnkey seizure setup.

The geography does the rest of the work. The strait is not closed — shipping companies are self-deterring, which Iran is passing off as leverage. The moment the US commits to escorting convoys through, that theater ends. Iran can make the strait scary. They cannot stop a determined naval force. Khuzestan is accessible from Iraq across flat terrain, with the Zagros blocking Iranian reinforcement from the heartland. Kharg, stripped of its military garrison, sits with its pumps running and 18 million barrels in storage. The offshore fields that feed it are platforms — small footprint, manageable.

The 2,500 Marines en route aboard the USS Tripoli — accompanied by the USS New Orleans and USS San Diego as part of a full Amphibious Ready Group — are not a regime change force. A MEU takes ports, terminals, and pipeline junctions. It does not take capitals. Nobody is marching on Tehran.

Where It Gets Hard

Three things complicate the arithmetic.

Ahvaz city. The Ahvaz field is threaded through a metropolitan area of roughly 1.3 million people. The answer is not nation-building — it is relocation. Move the civilian population out of the operational zone, secure the infrastructure, keep the oil flowing. It is not pretty. It is also not nation-building, reconstruction, or the installation of democracy. It is making the world’s economy function. The obligation is to move people safely, not to rebuild their city or determine their government.

A word on sympathy. The Iranian people deserve it — many are trapped by a government they’d gladly be rid of. But the 1979 revolution was not imposed on a passive population. It was a mass uprising. Millions of Iranians took to the streets to bring down the Shah. They chose theocracy. Theocracy failed them — but that is a consequence of their own political history, not an obligation the rest of the world is required to fix. The world economy did not vote for the Islamic Republic. It should not have to keep paying for it.

The scorched earth question. Non-military observers assume Iran would torch the fields rather than surrender them. Students of military history are more skeptical. Scorched earth requires a command structure willing to permanently destroy its own country’s primary source of national revenue — not to win the war, but to spite the occupier. Nobody burns down the house they intend to live in. Marginal sabotage is likely. Systematic destruction of the entire field complex is a regime betting it has no future. That is a different, harder calculation.

China. Approximately 80% of Kharg’s oil goes to China. The foreign policy establishment calls this unthinkable. That analysis mistakes theater for reality. China buys Iranian oil because it is cheap and reliably delivered. If the US controls Kharg, China still needs the oil. The negotiation changes from a back-channel arrangement with a sanctioned regime to a commercial transaction with whoever is operating the terminal. China will complain loudly and then place the order. The US controlling Iranian production is not a cutoff — it is a change of management. This is a business problem, not a security threat.

There is a secondary dimension worth naming. US control of Kharg is not just leverage over Iran — it is structural leverage over China’s energy supply that no sanctions regime has ever provided. Washington would hold the spigot for roughly 12% of China’s seaborne oil imports. That is a negotiating position across every other issue on the US-China agenda. Beijing knows this. China also holds an estimated 115-day strategic reserve — enough to absorb a transition period without a supply crisis, which means they can adapt to new commercial terms without going to the mat over it. It is one more reason they will adapt rather than escalate.

The Off-Ramps Iran Won’t Take

Iran has options. Reopen Hormuz. Stop funding proxies verifiably. Abandon the nuclear program. Accept internationally managed oil operations under revenue-sharing terms. Any one of those is a climbdown the regime could probably survive.

Regime change is not the objective and never was. The clerics can stay in power. They simply have to stay inside Iran. What is not acceptable is funding Hezbollah, arming Houthis, backing Hamas, spinning centrifuges toward a weapon, and threatening international shipping. Stop doing those things and the pressure stops. Behavioral compliance, not regime change. The bar is low enough that a rational government would have cleared it years ago.

This is the same framework the Trump administration has applied to Venezuela. Maduro can run his dictatorship — just keep it inside your borders. Iran and Venezuela are not the same crisis. They are the same problem statement.

The strait is not closed — shipping companies are self-deterring, which Iran is passing off as leverage. It isn’t. The moment the US commits escorts, the theater ends. Iran is trying to win by making everyone uncomfortable. The US is trying to win by making Iran unable to fight. Those are not symmetric strategies. One of them runs out of runway. Iran’s Hormuz gambit is not leverage. It is the rope to hang itself.

The off-ramps are visible, they are survivable, and they are closing. Every day Iran waits, the terms get worse and the options get fewer.

The Verdict

The window exists. Iran has spent its conventional deterrent. The geography concentrates the prize into a compact, accessible theater. The assets are positioned. If there is ever a moment when seizing Iran’s oil infrastructure is militarily feasible — whether as an executed plan or a threat held in reserve — this is it.

Kharg and the offshore fields are seizable with what is in theater — and the strait disrupted by fear rather than actually closed is less of an obstacle than it sounds. Khuzestan — the full prize — is harder but gets more possible every day Iran refuses the off-ramps.

The definition of success is behavioral, not governmental. The US does not need Iran to become a democracy or even like us. It needs Iran to stop exporting violence and stop threatening global shipping. The clerics can keep their jobs. The price is staying in their lane.

In my last piece I argued that destroying Iran’s oil capability was the right play. I still think destroying is cleaner than taking — you don’t have to defend rubble. But the Marines on that slow boat are not a demonstration force. You send carriers and Tomahawks for a demonstration. You send an amphibious assault ship with a Marine Expeditionary Unit when you are planning to put boots on something and hold it.

The USS Tripoli is about 7 to 10 days from theater. That is Tehran’s clock. When it arrives, the options narrow.

Iran can see what is coming. They need to accept the inevitable — or watch this plan unfold.

If Dario Amodei were a friend, I think I’d organize an intervention. The symptoms are all there — the messianic certainty, the inability to accept that reasonable people might disagree, the compulsion to be seen as humanity’s moral guardian even at catastrophic cost to his own business.

I think professional counselling may be in order.

Last week Anthropic filed two simultaneous lawsuits against the Pentagon, challenging its designation as a “supply chain risk” — a label previously reserved for companies with ties to foreign adversaries, particularly China.

The Legal Case Is a Fig Leaf

Let’s be clear about the legal merits. Courts defer to the executive branch on national security designations with near-religious consistency. The First Amendment claim, while creative, requires proving the government took action specifically to punish protected speech — an extraordinarily high bar against a Pentagon that has perfectly coherent operational reasons for dropping a vendor who tried to impose its own use restrictions. The statutory interpretation argument is the strongest piece, but “this label was designed for foreign companies” is a long way from a winning case given how consistently courts defer to the executive branch on national security grounds.

What He Was Actually Selling

Anthropic walked into a $200 million contract negotiation with the Pentagon and attempted something unprecedented: getting the United States military to contractually accept a private CEO’s personal judgment about appropriate use of AI in warfare. No autonomous lethal weapons. No mass surveillance of Americans. Reasonable-sounding principles, until you ask the obvious follow-up — says who?

Dario Amodei is not an elected official. He has no democratic mandate, no Senate confirmation, no oversight accountability. His ethical positions, however sincerely held, represent the views of one man running a San Francisco startup.

No serious defense procurement officer could accept that. The wonder isn’t that the talks collapsed — it’s that they got as far as they did.

Perhaps Dario never heard “don’t fight city hall.”

The Lord Giveth and the Lord Taketh Away

But forget the Pentagon for a moment. The most damaging consequence of this entire episode isn’t the lost contract or the supply chain label. It’s what every enterprise CTO and procurement officer just watched happen in real time.

Can you build critical infrastructure on a platform whose CEO reserves the right to decide your use case is immoral and pull the plug? Can you architect workflows, train internal models, build customer-facing products, or commit serious engineering resources around a vendor whose terms of service are ultimately subject to one man’s evolving moral philosophy?

The answer is obviously no. And Anthropic just demonstrated it publicly and dramatically, at scale, in the highest-profile contract negotiation in the industry, with every enterprise customer watching. They didn’t just threaten to restrict use — they actually did it. Every CTO in America just got a live demonstration of exactly what building on Anthropic looks like when Dario decides your use case crosses his line.

This isn’t an AI safety argument anymore. It’s a vendor reliability argument. The Lord giveth and the Lord taketh away is charming theology. It’s a catastrophic enterprise sales proposition.

The Audience He’s Actually Playing To

So if the legal case is weak, why file? One can only speculate that the lawsuit isn’t about winning in court. It’s about controlling a narrative for several very specific audiences simultaneously.

Investors are asking very uncomfortable questions right now. You lost a $200 million contract. You got labeled a national security risk. OpenAI is now embedded in classified Pentagon systems and you’re not. You voluntarily walked away from the largest customer on earth over restrictions nobody asked you to impose. The lawsuit reframes that conversation entirely. Suddenly Dario isn’t the founder who torpedoed major revenue over personal ideology — he’s heroically fighting illegal government retaliation.

Some tech talent in San Francisco won’t work for companies they believe are doing autonomous weapons contracts or mass surveillance of the public, even though it’s not clear what this means in practice — and people striking these poses generally have no understanding of how weapons systems are actually built and tested. Anthropic just publicly, loudly, legally confirmed it refused that work on principle.

Bobblehead Democratic politicians looking to find issues that might yield political currency lined up to express concern. And the media delivered the framing he needed without much resistance. The real story — vendor tries to impose unprecedented restrictions on military customer, gets dropped — became “Trump weaponizes national security powers against principled AI company.”

What a Military Project Actually Is

At the risk of sounding condescending, the people talking about autonomous weapons and military AI speak in ways that make it clear their only real understanding of what goes on in a military project comes from science fiction movies and playing Call of Duty. They know as much about military procurement, weapons testing, and system integration as Pete Hegseth knows about transformer architecture.

Having worked on military projects, I can say with some confidence that the hand-wringing about AI being used for autonomous lethal weapons and mass surveillance reflects a level of hysteria that has no grounding in how these systems are actually built, tested, and deployed. Military projects operate inside layers of oversight, legal review, and institutional checks that most civilians never see and apparently never consider. Many projects are so classified that even their names are secret — you’re not working on an autonomous weapons system, you’re working on Project P10. Employees with actual knowledge of wrongdoing have whistleblower protections and a long history of using them.

Nobody fields a system with inherent risk. Military projects typically take ten years or more from inception to deployment — sometimes much longer. The implicit assumption behind all this hysteria is that tomorrow the Pentagon is going to let Claude fly a stealth fighter mission. That’s not procurement reality. That’s science fiction.

The reliability argument deserves one sentence: the engineers, testers, and systems integrators who build these platforms have spent careers thinking about failure modes, redundancy, and operational risk. They don’t need a San Francisco CEO to educate them on how to make reliable military systems.

Consider the inevitable scenario: Dario reads in the New York Times or Washington Post a leak from an unnamed Pentagon official claiming Anthropic’s model is being used for autonomous weapons or mass surveillance. He demands answers. He threatens to pull the technology. But even knowing the name of the project he’s asking about requires a top secret clearance — or higher. He has neither the clearance nor the context to evaluate what he’s being told, yet under his proposed arrangement he would have contractual power to shut it down. That’s well-meaning meddling by people who don’t understand what they’re looking at.

That is a supply chain risk. Full stop.

I don’t know the precise legal definition of supply chain risk, but I know this: I would never award a contract to a company behaving the way Anthropic is behaving. Nobody building serious systems wants that kind of meddling. Nobody should want that.

Congress makes laws that decide what lawful use of a product is. Your recourse is the judicial courts, not the Anthropic meeting rooms.

The Moral Beacon for the Planet

What unifies all of it is something that predates the Pentagon fight entirely. Anyone who has followed Amodei’s interviews and writings closely has watched a consistent and revealing pattern. This is not a man who believes he has some interesting ideas about AI safety. This is a man who believes he is the moral beacon for the planet — one of a small number of technically sophisticated people who understand the existential risks facing humanity and are therefore morally obligated to guide it, whether humanity wants guidance or not.

That worldview, deeply rooted in the effective altruism and longtermism movements that shaped his intellectual formation, is self-sealing by design. Every criticism confirms the narrative. Pushback means you’re threatening the mission. Losing the contract means you refused to compromise your principles. Getting labeled a supply chain risk means the forces of evil are retaliating against the truth teller.

It’s a completely closed loop. And it makes him, practically speaking, an impossible business partner. He’s not actually selling AI. He’s selling AI with Dario Amodei as an undisclosed co-dependent partner who retains moral veto power over your business decisions.

No serious enterprise customer wants that product once they understand what it actually is.

The Kicker

Sam Altman said yes to the Pentagon and got the contract. Dario Amodei said no and got a lawsuit, a press cycle, and a martyr narrative.

The underlying concerns about AI being used for autonomous lethal weapons and mass surveillance of Americans are worth taking seriously. These are real questions that democratic institutions should be wrestling with openly. But those questions don’t get answered by a private CEO extracting contractual veto power from the military. They get answered through legislation, oversight, public debate, and democratic accountability. Institutions Amodei shows no particular interest in, possibly because they don’t have a position for him.

If Dario is looking for his next opportunity, he might consider Iran. The Supreme Leader position offers lifetime tenure, unquestioned moral authority over an entire population, and the power to cut off critical infrastructure to anyone who violates your ethical code. The kleptocracy benefits are excellent too — the previous leader reportedly amassed hundreds of billions.

And nobody can label the Supreme Leader a supply chain risk.

Dario: focus on making your great product better, but stay in your lane and sphere of competency, which is not anything having to do with the military.

Iran is where that lesson needs to be applied right now.

Not because we want a war with Iran. We don’t. Not because diplomacy hasn’t been tried. It has — across eleven presidential administrations, 47 years, every conceivable combination of carrots and sticks. But Iran today fields more capable missiles, funds more dangerous proxies, and runs a more advanced nuclear program than it did when the Shah fell in 1979. The diplomacy produced the opposite of its stated objectives. At some point the evidence closes the argument.

Here’s what 47 years actually produced: a regime that has absorbed every sanction, every expulsion, every maximum pressure campaign, and every diplomatic initiative — and emerged from each one still funding Hezbollah, still spinning centrifuges, still paying proxy armies across five countries. The tools changed. The result didn’t.

At some point the definition of insanity applies — doing the same thing and expecting a different result.

The Wrong Question

Washington’s entire Iran policy has been built around one question: how do we pressure Iran into changing its behavior?

That’s the wrong question. It assumes the regime is a rational actor that will eventually respond to sufficient pain by moderating. The evidence — nearly five decades of it — says otherwise. The regime’s external aggression isn’t a policy preference it might negotiate away.

There is also no surgical option. There never was. The foreign policy establishment has spent that entire time searching for the intervention that applies just enough pressure to change behavior without consequences — without pain, without disruption, without anyone inside or outside Iran getting hurt. That option does not exist. Iran has deliberately embedded itself in global energy markets, regional proxy networks, and Chinese supply chains precisely to make surgical action impossible. The complexity isn’t accidental. It’s the strategy. Every layer of entanglement is a deterrent against clean solutions.

The search for the surgical option is itself part of the technical debt. Every year spent looking for the clean answer was a year the messy answer got more expensive. Anyone still proposing targeted sanctions, diplomatic off-ramps, or calibrated pressure as a serious solution in 2026 is not offering an alternative. They are proposing to extend the debt further.

The right question is: how do we make it impossible for Iran to afford the behavior we’re trying to stop?

That’s a different problem with a different solution. And the solution is straightforward, even if the execution isn’t: destroy the economy. Not sanction it. Not pressure it. Destroy it — sequentially, proportionally, with each stage triggered by Iranian behavior, and each stage designed to be as damaging as possible at minimum cost to us.

How You Destroy an Economy

Iran’s external threat capability runs on oil money. Not ideology — money. Cut the money and the proxies stop. Cut the money and you cut everything it purchases — simultaneously, automatically, without having to chase each program individually around the country.

This is the strategic insight that the four-objectives, whack-a-mole approach misses entirely. You don’t need to separately destroy the missile program, defund the proxies, set back the nuclear program, and degrade power projection capability one at a time. You pull the financial foundation out from under all of them at once. No money means no missiles and no proxies and no nuclear program. One mechanism. Four problems degraded simultaneously.

Iran currently exports roughly 1.5 to 1.6 million barrels of oil per day. Almost all of it goes to China, laundered through a shadow fleet of tankers with spoofed transponders, ship-to-ship transfers near Malaysia, and UAE shell companies obscuring the origin. Kharg Island handles roughly 90% of those exports. Iran has secondary export capacity at Bandar Abbas and the Jask terminal — Jask built specifically to bypass Hormuz — plus small-scale pipeline and smuggling routes. Destroying Kharg doesn’t eliminate exports. It cripples primary export capacity and forces Iran onto slower, costlier, more interdictable alternatives. That is the point. Each alternative route is less efficient, more exposed, and easier to shut down in subsequent stages.

The constraint has never been capability. It has been political will.

Stage One: Pull the Pin

Destroy Kharg Island. Destroy the shadow fleet wherever it can be found and interdicted. Destroy the key refinery and storage infrastructure that feeds the export chain.

This is the cheapest, most surgical option available. No ground troops. No occupation. No nation-building. A sustained air and naval strike campaign against defined, targetable assets. Expensive in ordnance, manageable in risk, devastating in effect.

Iran goes from roughly $40–50 billion in annual oil revenue to a fraction of that — forced onto secondary routes at higher cost, lower volume, and greater interdiction risk. Existing reserves — best estimates put usable liquid reserves at $10-20 billion, with additional yuan-denominated credits trapped in Chinese banks — combined with dramatically reduced export income, start a clock the regime cannot stop.

Then we wait.

Do not automatically proceed to the next stage. Apply Stage One and watch what happens. If Iran stops funding proxies, stops missile development, stops the nuclear program — you’ve achieved the objective at minimum cost. Unlikely, given the evidence. But the option exists, and leaving it open costs nothing.

The burden of escalation belongs to Tehran. Every subsequent stage is a choice they make, not one we make.

The Test After Stage One

If they don’t stop, go to Stage Two.

Stage Two: The Blockade

A naval blockade is more expensive and more escalatory than Stage One. It requires sustained carrier battle group presence, interdiction of all maritime traffic into Iranian ports, and neutralization of Iran’s asymmetric naval capability — fast boats, mines, shore-based anti-ship missiles along the Gulf coastline.

Then we wait. Same test. If they don’t stop, go to Stage Three.

Stage Three and Beyond

If somehow the regime is still functioning — continue shutting down alternate export routes, tighten the blockade, target remaining IRGC economic infrastructure. The options are available.

But frankly, Stage Three should never happen. A country that has lost its primary oil export infrastructure and been under naval blockade is bankrupt. The regime cannot fund a nuclear program, pay proxy armies, or maintain a missile arsenal on empty reserves and blocked ports. Stage Three is the contingency plan for a scenario that the first two stages should make impossible.

What This Is Not

It is not nation-building. There is no reconstruction plan attached to this. What Iran looks like on the other side is not our problem and never was. We are not in the business of determining Iranian governance. We are in the business of ensuring that whatever government runs Iran cannot afford to threaten its neighbors. Those are completely separate questions and conflating them is exactly how half a century of policy failure happened.

It is not a forever war. The exit condition is simple and measurable: when Iran demonstrably cannot project force beyond its borders, the active campaign ends and monitoring begins. No victory parades. No reconstruction conferences. Monitoring and deterrence.

It is not unprovoked. Iran funded October 7th. Iran directs Houthi attacks on global shipping. Iranian proxies have attacked U.S. forces in Iraq and Syria hundreds of times. We are not contemplating action against a peaceful country. We are contemplating cutting off an active belligerent’s war funding — which is the most restrained available response to what Iran has already been doing for decades.

A Note on the Present Moment

This piece was written as advocacy. As of March 9, 2026 — ten days into Operation Epic Fury — it reads as a roadmap being executed in real time. Kharg Island strikes are actively under discussion. The Strait of Hormuz is at near-standstill. IRGC command centers and missile sites have been hit. The sequential logic described here is playing out, stage by stage, whether by design or convergent reasoning.

Current events suggest that the proposal put forth in this op-ed may well be the thinking of the administration. What they haven’t done is be honest with the American public that this is the plan.

The question: Are LeCun, Hinton, and Bengio what people think they are — pioneers of machine learning?

Is the critique legitimate?

This is Jürgen Schmidhuber’s personal history page — and it’s essentially a long, meticulously documented argument that LeCun, Hinton, and Bengio (the “Turing Award trio,” sometimes called the “Godfathers of Deep Learning”) received credit for work that was substantially pioneered by others, most notably Schmidhuber himself and his colleagues at IDSIA.

Largely, yes — with important caveats.

Schmidhuber’s claims are well-documented with timestamped citations. Some of the most striking:

• GANs: Bengio’s 2014 GAN paper — his most-cited work — did not cite Schmidhuber’s 1990–91 adversarial generative NN work, which used essentially the same minimax framework between two networks.

• Transformers: The linear Transformer is mathematically equivalent to Schmidhuber’s 1991 “fast weight programmer,” published decades before the 2017 “Attention Is All You Need.”

• Knowledge distillation (used by DeepSeek and others): traced to Schmidhuber’s 1991 “chunker/automatiser” work, not Hinton’s later popular framing.

• Unsupervised pre-training: The conceptual basis for what made ChatGPT possible was established via the Neural History Compressor (1991), predating Hinton’s Deep Belief Networks (2006) by 15 years.

• CNNs in practice: Schmidhuber’s student Dan Ciresan’s DanNet was winning contests in 2011 — beating LeCun’s team by 3x — before AlexNet.

Where it gets complicated

Schmidhuber is right that the history is messier and more internationally distributed than the popular narrative suggests. The Turing Award trio benefited enormously from timing — they were prominent, English-speaking academics at well-networked institutions when the compute finally made deep learning commercially explode. Credit in science is notoriously path-dependent: who got cited, who gave the influential talks, who framed the narrative.

That said, LeCun’s convolutional networks were genuinely foundational for image recognition. Hinton’s backpropagation popularization and later work on dropout were real contributions, even if not as original as portrayed. Bengio’s work on sequence modeling and language models contributed meaningfully to the LLM trajectory.

The bottom line

The “Godfathers” label is a marketing narrative, not a precise history. Schmidhuber’s documentation is credible and largely holds up — he has a legitimate grievance that foundational work from his lab (LSTM, linear Transformers, GANs, distillation, pre-training) was either not cited or was rebranded when it became commercially important. The trio were important contributors and excellent at communicating and advocating for the field — but the “inventors of deep learning” framing is an oversimplification that erases a lot of prior art, much of it Schmidhuber’s, but also work by Werbos, Fukushima, Amari, and others.

It’s a familiar story in tech: the people who get credit are often those who were there when the world was finally ready to listen, not necessarily those who had the ideas first.

Claude is an AI assistant made by Anthropic. This response was generated from a single query with no editorial guidance from the user.