Analysis
OpenAI published its first real benchmarks for Jalapeño, the custom AI chip it unveiled in June built with Broadcom, and the numbers are the strongest evidence yet that frontier labs can meaningfully compete with Nvidia on their own silicon. Presenting at the Hot Chips conference, OpenAI said the chip delivers 1.5 to 1.9 times more AI work per watt and 1.7 to 3.6 times lower latency than Nvidia's Blackwell GB200 and GB300 systems across three public models. OpenAI's VP of Hardware, Richard Ho, said the chip "achieves high throughput and low latency simultaneously, a first in the industry."
What makes this different from a vendor's own marketing slide: SemiAnalysis, an independent semiconductor research firm, sent engineers to OpenAI's labs and ran the workloads itself using InferenceX, its own public inference benchmark, rather than taking OpenAI's numbers on faith. That verification is the actual news two months after the June unveiling -- Jalapeño went from an announced project to a chip with third-party-checked performance data.
The caveats that matter
Analysts were quick to flag that the comparison is not entirely clean. Jalapeño uses newer HBM4 memory, which Blackwell does not, making the fairer long-term matchup Nvidia's upcoming Rubin systems -- also HBM4-based and starting to ship to customers now, while Jalapeño remains at the engineering-sample stage. Jalapeño also handles inference only; Nvidia's dominance in training hardware is untouched. OpenAI has said deployment inside its own infrastructure begins in "very small volumes" by the end of 2026, with meaningful scale not arriving until 2027, and it's already working on Jalapeño's second and third generations.
OpenAI joins a growing list of frontier labs and hyperscalers building custom silicon specifically to reduce Nvidia dependency: Google's TPUs, Amazon's Trainium, Microsoft's Maia, and Meta's MTIA all follow the same logic, and Anthropic is separately building out its own in-house chip design team. The pattern is now the norm rather than the exception among the companies large enough to afford it -- Nvidia remains the default supplier for everyone else, and its own record $96.2 billion quarter this week shows demand for its chips isn't going anywhere regardless.
The counterweight: a benchmark win at engineering-sample scale is not the same as a production win. Nvidia has repeatedly out-executed custom-silicon challengers on the actual manufacturing ramp, and OpenAI's own admission that volume production waits until 2027 means Blackwell and Rubin will be shipping at scale for at least another year before Jalapeño is a real alternative rather than a proof of concept.
OpenAI's own chip costs and cash burn -- the company isn't expected to reach profitability until around 2030 -- make Jalapeño as much a cost-control story as a performance one; the next checkpoint worth tracking is whether OpenAI's 2027 production ramp holds to schedule once Rubin is actually shipping in volume.
