Intel’s AI Acquisition Déjà Vu: Why SambaNova Feels Like Groundhog Day Under Lip-Bu Tan
The Acquisition Graveyard
Intel’s board has a storied history of greenlighting AI and machine learning acquisitions that promised to catapult the company into Nvidia’s league—only to fizzle spectacularly. Nervana Systems (acquired for ~$400 million in 2016) was hyped as a deep learning powerhouse but got shelved by 2020 in favor of the next big bet. Habana Labs (snapped up for $2 billion in 2019) was meant to dominate data center AI training and inference, yet it has captured minimal market share amid Nvidia’s unrelenting grip. Movidius, Mobileye’s pre-IPO AI components, and other smaller bets followed similar trajectories: big announcements, limited impact.
Enter SambaNova
As of December 17, 2025, Intel is in advanced talks to acquire SambaNova Systems for approximately $1.6 billion (including debt), a deal reportedly driven by CEO Lip-Bu Tan. SambaNova, founded in 2017, specializes in full-stack AI platforms with its reconfigurable dataflow architecture (SN40L RDU chips and SambaRack systems), excelling particularly in large-scale LLM inference. Independent benchmarks—like Artificial Analysis leaderboards and the 2024 academic LLM-Inference-Bench study—often show SambaNova outperforming Intel’s Habana Gaudi2 (and early indications of Gaudi3) in tokens-per-second and efficiency for massive models (e.g., Llama 3.1 70B/405B, DeepSeek 671B), often at full precision without quantization.
Why buy another accelerator when Habana already exists? The rationale centers on complementary strengths: Habana’s Gaudi series has focused more on training (with respectable MLPerf submissions) but lagged in enterprise inference dominance. SambaNova’s dataflow design offers greater flexibility for irregular, production-scale workloads, potentially giving Intel a stronger hand against Nvidia in the inference-heavy enterprise and data center markets where it has struggled.
The Benchmark Problem
Yet SambaNova carries a glaring credibility gap: it has never submitted to MLPerf, the industry’s peer-reviewed, standardized benchmark suite. Habana participates regularly; SambaNova, like Groq and Cerebras, does not. The company (and peers) cite resource intensity, mismatched workloads (MLPerf’s batch-oriented focus vs. their low-latency inference strengths), and Nvidia’s optimization dominance as reasons. However, these firms undoubtedly run exhaustive internal benchmarks mirroring MLPerf—they know their performance inside out. Skipping public submission leaves room for skepticism: if the numbers were decisively superior in standardized configs, participation would silence doubters. Instead, SambaNova relies on third-party validations, academic studies, and self-comparisons, which impress but lack MLPerf’s audited rigor.
Usually, companies don’t submit industry benchmarks because they ran them internally first and got the wrong answer, if you catch my drift. The excuses are many and come from the list we learned in school, which includes “the dog ate my homework.”
The Conflict of Interest
Even setting the technical validation questions aside, the deal’s optics are complicated by conflicts of interest that would raise eyebrows in any public-company boardroom. Lip-Bu Tan, Intel’s CEO since March 2025, is also SambaNova’s executive chairman. His venture firm, Walden International, was a founding investor—co-leading the 2018 $56 million Series A with GV (Google Ventures) and participating in later rounds (total fundraising over $1.1 billion, peaking at a $5 billion valuation in 2021 before cooling). Earlier in 2025, Tan reportedly pitched the board on acquiring Rivos (another company he chaired), only for it to be rejected over similar concerns. With SambaNova’s implied valuation down in a tougher market, this acquisition could provide a lucrative exit for Walden and early backers.
The Real Intel Problem
Lip-Bu Tan’s background—BS in physics from Nanyang Technological University, MS in nuclear engineering from MIT, decades founding Walden International and chairing Cadence—is formidable in semiconductor investing and dealmaking. But his strengths lie in capital allocation, restructuring, and brokering transactions rather than in building and nurturing new AI architectures post-acquisition. Intel’s problem isn’t just finding promising technology to buy—it’s knowing what to buy and then integrating that technology into a coherent product strategy that developers actually adopt.
Groundhog Day
Nothing suggests Intel’s board has fundamentally changed its pattern of approving expensive, overlapping AI bets without clear integration success. The SambaNova pursuit risks repeating history: bolting on promising tech that fails to displace Nvidia or generate meaningful revenue. If the deal closes (expected possibly early 2026, though it could fall apart), it may bolster Intel’s inference portfolio—but without SambaNova stepping up to MLPerf, transparent conflict resolution, and a proven track record of post-acquisition execution, it feels less like a breakthrough and more like Groundhog Day: another costly swing in Intel’s long quest to catch up in AI.
Disclosure: Cranky Old Guy was a founding member of SiMa.ai, an embedded AI processor company, and owns stock in it.