OpenAI’s Jalapeño chip beat Blackwell on a SemiAnalysis benchmark — but the fine print on who ran the test, on what workload, against which chip, changes everything about what that actually proves.
Theo Browne — t3.gg
“OpenAI’s first-generation inference chip is industry leading in beating every NVIDIA, AMD, and Google chip we have been able to test.”
Theo Browne, reading a SemiAnalysis headline on stream — “OpenAI Jalapeño: Better Than Nvidia Blackwell” — lands on the line everyone will repeat. First silicon almost always loses. This one, per SemiAnalysis, wins. If that holds, it is a genuine event in the AI-hardware story.
The phrase “if that holds” is the entire piece. Start with what is solid, because plenty is. The SemiAnalysis writeup is real, the quote is verbatim, and CNBC and Forbes carried it. The figures are striking: up to roughly 1.9x more output tokens per watt and up to about 3.6x lower latency than Blackwell on SemiAnalysis’s InferenceX suite. OpenAI co-designed Jalapeño with Broadcom, and the disclosed specs — 216GB of HBM4, 700 watts, scaling to 2,048 chips — describe a serious part, not a lab curiosity.
The key insight: This is not an independent benchmark — OpenAI invited SemiAnalysis into its lab, and SemiAnalysis is explicit that the numbers were provided by OpenAI, making “OpenAI’s own numbers, verified as having actually been run” a different and weaker claim than “we benchmarked it and it won.”
The caveats are not editorializing — they are SemiAnalysis’s own. This is A0 first silicon, measured on narrow 8k-context, single-turn workloads rather than the agentic benchmark SemiAnalysis considers most meaningful, without speculative decoding, on smaller models than competitors’ published runs. Most tellingly, SemiAnalysis itself calls the Blackwell comparison “somewhat incomplete and unfair”: Jalapeño uses HBM4, so the like-for-like rival is Nvidia’s Rubin — which is already shipping — while OpenAI has only engineering samples.
The fair framing is not “Jalapeño beat Blackwell,” but “an OpenAI-supplied early look shows first silicon competitive with a prior-generation part, on OpenAI’s chosen workload.” Strip the horse-race and the real signal not only survives, it sharpens. OpenAI chose to reveal Jalapeño — on its own terms, with its own numbers — the same week Nvidia printed $60 billion. The goal is not to dethrone Nvidia this quarter; it is to prove to its own board, its capital providers, and Nvidia’s sales team that a second source exists.
This is the same bifurcation Marvell’s raised guide and Anthropic’s hardware standard are pricing: the largest buyers are building an exit from single-supplier margin, and a favorable early benchmark is how you announce the exit is coming before it has fully arrived.
Structural Signal
Jalapeño is real and it matters — as leverage, as a second source, as proof of intent.
OpenAI has a credible in-house inference ASIC. That changes its negotiating position with Nvidia regardless of whether the chip ever ships at scale.
IMPLICATION 1 — FIRST SILICON THAT ACTUALLY COMPETES
First silicon almost always loses. Jalapeño, by OpenAI’s own numbers on its own workload, does not. That is a genuine event in the AI-hardware story — even before it ships at scale.
IMPLICATION 2 — READ BENCHMARKS, NOT HEADLINES
The distance between “Nvidia lost their lead” and “Nvidia’s largest customer showed a first-gen chip that, by its own numbers on its own workload, is competitive” is the whole discipline of reading a benchmark. The caveats here come from SemiAnalysis itself, not editorializing.
IMPLICATION 3 — SUPPLIER CONCENTRATION IS THE REAL STORY
The largest buyers are building an exit from single-supplier margin. Jalapeño, alongside Marvell’s raised guide and Anthropic’s hardware standard, signals that the hyperscale bifurcation away from Nvidia dependence is a structural trend, not a one-off product bet.
Business Engineer Framework
Map of AI — Where Jalapeño Sits in the Stack
Understanding OpenAI’s hardware move requires mapping where custom silicon sits across the AI stack — and why the largest players at the model layer are now pushing down into infrastructure to escape single-supplier lock-in. The Map of AI framework traces exactly this kind of vertical integration event.
“Nvidia lost their lead” is the headline. “Nvidia’s largest customer showed a first-gen chip that, by its own numbers on its own workload, is competitive” is the fact. Jalapeño is real and it matters — as leverage, as a second source, as proof of intent. Just not yet as a Blackwell obituary.
Clip via Theo (t3.gg) (source). Analysis by FourWeekMBA.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
