OpenAI's First Chip, Jalapeño, Reportedly Beats Nvidia's Blackwell
Independent benchmarks from SemiAnalysis show OpenAI's custom inference chip beating Nvidia's GB300 on performance per watt. Here's what that does and doesn't prove.
OpenAI has been talking about building its own chips for a while. This week, the talk turned into numbers. Research firm SemiAnalysis got hands-on access to OpenAI’s first custom silicon, codenamed Jalapeño, and published benchmarks showing it beating Nvidia’s current flagship inference hardware, the GB300 Blackwell system, on performance per watt in nearly every test they ran.
The headline figures: 1.5 to 1.9 times more AI work per watt at peak throughput, with latency cuts of 1.7 to 3.6 times, depending on the workload. Jalapeño’s package draws 700 watts against the GB300’s 1,400 watts — half the power budget before you even get to the throughput numbers.
What Jalapeño actually is
Jalapeño is an inference chip, not a training chip. That distinction matters more than it sounds. Training a frontier model is a small number of enormous, sustained runs — you buy the best chip you can get and run it for months. Inference is different: it’s the chip that answers your ChatGPT query, over and over, at a scale where power cost is a real line item. A company running hundreds of thousands of inference requests a day cares enormously about work-per-watt, because that number turns directly into a power bill and a data-center footprint.
That’s the pressure point OpenAI is aiming at. It doesn’t need to beat Nvidia everywhere. It needs to beat Nvidia on the metric that determines whether it’s cheaper to serve its own models on its own silicon than to keep renting Nvidia GPUs at Nvidia’s margins.
The caveat SemiAnalysis put on its own numbers
To its credit, SemiAnalysis flagged the comparison as “somewhat incomplete and unfair” in OpenAI’s favor. Jalapeño uses newer HBM4 memory. The GB300 doesn’t. Nvidia’s actual like-for-like answer to Jalapeño is Rubin, its next-generation platform, which also moves to HBM4 and isn’t shipping in volume yet. So this isn’t “OpenAI beat Nvidia’s best chip” — it’s “OpenAI’s chip beats Nvidia’s current shipping chip, using a memory generation Nvidia’s current shipping chip doesn’t have.” Whether Jalapeño still wins against Rubin is the number that actually matters, and nobody has published it yet.
It’s also one benchmark run by one outside firm, on hardware OpenAI chose to show them, running workloads that likely favor Jalapeño’s design. That’s not a knock on SemiAnalysis’s rigor — independent, hands-on benchmarking is exactly the kind of scrutiny custom-silicon claims should get, and it’s a lot more than a press release. But “beat” is doing real work in these headlines, and the honest framing is narrower: on the specific tests run, using the specific hardware compared, Jalapeño came out ahead.
Why OpenAI is doing this at all
The economics are the whole story. Nvidia’s data-center margins are famously wide, and every company buying at that scale — OpenAI, Google, Amazon, Meta — has been moving toward custom silicon for the exact same reason: cut Nvidia out of your own inference costs, even if you keep buying Nvidia GPUs for training and for everyone else’s workloads. Google’s Marvell deal from earlier this month was the same move from a different angle — a financing structure that gets Google more custom TPU capacity without leaning on Broadcom’s near-monopoly on the packaging side. Jalapeño is OpenAI doing the equivalent thing for its own inference stack.
None of this means Nvidia is in trouble soon. OpenAI still needs enormous volumes of Nvidia GPUs for training, and Jalapeño is a first chip from a company that has never shipped silicon before — yields, supply chain, and software maturity are all unknowns that benchmark slides don’t capture. But analysts quoted by CNBC framed it correctly: this is pressure on Nvidia’s inference-market margins specifically, not an existential threat to its training dominance. It’s also a data point in the broader financing story — the $500 billion Nvidia has been lining up from Wall Street to fund AI buildout assumes continued demand for Nvidia chips at Nvidia prices. Every customer that successfully builds an inference alternative chips away at that assumption, even at the margins.
The number to watch next
Jalapeño versus GB300 was always going to favor OpenAI — it’s newer silicon with newer memory. The real test is Jalapeño versus Rubin, Nvidia’s actual next-generation answer, once both are benchmarked on equal footing. Until then, treat this week’s numbers as evidence that OpenAI can build competent inference silicon, not evidence that Nvidia’s moat has a hole in it.
Sources: