On Aug. 25, OpenAI released the first benchmark data for Jalapeño, the in-house inference chip it has been developing for more than two years, and the numbers landed like a shot across Nvidia’s bow. The company said the processor delivers industry-leading speed and energy efficiency for AI reasoning workloads and outperforms Nvidia’s Blackwell platform, including the GB300 systems now shipping to cloud providers. Analysts at SemiAnalysis have begun tearing down the architecture to verify the claims.
The timing was pointed. OpenAI published the results days after announcing it would build its own data centers and reduce reliance on external cloud partners, a strategy that has strained its relationship with Microsoft, its largest backer. A chip that can undercut Nvidia on both price and power consumption would give OpenAI the two things it has lacked since the ChatGPT boom began: control over its own supply chain and bargaining power in negotiations with the suppliers it still needs.
Jalapeño is designed for inference, the stage of AI where a trained model answers queries, rather than for training, where Nvidia’s dominance is deepest. That focus is strategic. Inference is where the industry’s compute costs are exploding as chatbots, coding assistants, and agentic software move into production use. OpenAI said Jalapeño delivers “multiple times” the inference throughput per watt of Blackwell, a claim that, if confirmed, would make the chip attractive to any operator paying for AI at scale.
The company has been cagey about the chip’s origins. Engineers familiar with the project said OpenAI hired talent from Google’s Tensor Processing Unit program and from semiconductor startups, and that the design uses a custom architecture tuned for the transformer models that power its products. Jalapeño has been running inside OpenAI’s own data centers for months, the company said, handling a portion of production traffic alongside rented Nvidia and AMD systems.
SemiAnalysis, which published an early architectural review, said the chip’s design choices are unconventional but plausible. The key, according to the firm’s analysts, is a massive on-chip memory and a dataflow design that keeps weights close to the compute units, reducing the energy spent moving data, which dominates inference costs in modern accelerators. “The architecture reads like it was designed by people who watched Blackwell’s bottlenecks for two years,” the firm wrote.
The claims are hard to verify from the outside. OpenAI released aggregate figures without the full methodology, and independent benchmarks typically lag vendor disclosures by months. Nvidia, for its part, declined to comment on a competitor’s unpublished results, though the company has pointed to its own roadmap of annual chip refreshes and its CUDA software ecosystem, which remains the industry’s default.
What is not in dispute is the direction of travel. Every major AI company now designs or buys custom silicon. Google has shipped multiple generations of TPUs. Amazon builds Trainium and Inferentia. Microsoft, Meta, and Amazon have all announced in-house accelerator programs. OpenAI’s move to go beyond renting compute to fabricating its own chips completes a pattern in which the industry’s biggest buyers have decided that Nvidia’s margins are a cost of doing business they no longer want to pay.
The implications for Nvidia are financial as much as technical. Inference is projected to account for a growing share of AI compute spending, and pricing power in that segment has been Nvidia’s to lose. If custom chips from OpenAI and others take meaningful share, Nvidia would be pushed toward the same position Intel occupied when cloud giants stopped buying its server CPUs: still huge, still profitable, but no longer the bottleneck of the industry.
OpenAI has said Jalapeño will remain internal for now. The company’s chief financial officer told investors the chip is part of a plan to control costs as usage grows, not a product for sale. That stance could change, analysts said, if the chip’s performance holds up and OpenAI decides the fastest way to amortize a billion-dollar design program is to sell it.
The broader race is over who sets the pace of AI hardware. Nvidia has answered custom chips with faster refreshes and tighter software lock-in. OpenAI has answered with a chip designed around its own models, which gives it an advantage no merchant supplier can match: it knows exactly what workloads the silicon must serve. The first generation is a claim; the second generation, already in design, will be the evidence.
Engineers who have seen both platforms said the honest comparison will come from third-party tests, and those take time. Until then, the practical consequence of Aug. 25 is a market now pricing in the possibility that Nvidia’s inference franchise faces its first credible challenge from a customer that has become a rival. OpenAI named the chip for a pepper it once ate during a late-night design meeting, according to people who worked on the project. The name was a joke. The benchmarks are not.
