OpenAI's first chip has benchmarks now, and OpenAI ran them

OpenAI put the first measured performance numbers for its own inference chip on the table on Tuesday, in a talk at the Hot Chips conference and a post published the same morning. The headline claim is that Jalapeño, the accelerator OpenAI designed with Broadcom, does two things at once that inference hardware normally makes you choose between: more work per unit of power, and faster responses for the individual user sitting in front of the product.

A companion piece from OpenAI CFO Sarah Friar went up five minutes later, arguing the strategic case for owning the stack down to the silicon. The chip itself is not new — OpenAI and Broadcom unveiled it on 24 June, when the companies said final performance was still being measured and that a detailed technical report would follow "in the coming months." This is that report, two months on.

What Jalapeño is

Jalapeño is what Broadcom's June announcement called OpenAI's first "Intelligence Processor": a blank-slate design for large language model inference rather than a general-purpose AI accelerator adapted from older workloads. OpenAI designed the architecture around its own kernels, serving systems and model roadmap; Broadcom handled silicon implementation and networking, including its Tomahawk switching parts, and Celestica does board, rack and system integration.

Two details from that June announcement are worth carrying forward. First, the program went from initial design to manufacturing tape-out in nine months, which Broadcom described as possibly the fastest ASIC development cycle achieved in high-performance semiconductors — partly, both companies said, because OpenAI's models helped with the design work. Second, engineering samples were already running production workloads in the lab at target frequency and power, including GPT-5.3-Codex-Spark.

What OpenAI measured, and against what

OpenAI ran Jalapeño on InferenceX, a benchmark maintained by the analyst firm SemiAnalysis that measures end-to-end request handling rather than isolated kernel throughput. On that harness, OpenAI reports both more tokens per user and more throughput per kilowatt than currently available state-of-the-art inference processors. The comparison point is an NVIDIA Blackwell system.

The bottom line is that the results show a very, very significant performance advance over state of the art. Jalapeño can serve more AI work per unit of power, while also returning responses more quickly. It's very efficient to serve a lot of customers, but it can also be very low latency.

The architectural argument behind that is about data movement, not peak FLOPs. OpenAI says the design targets the prefill and communication phases specifically, which is where a lot of real serving time disappears. In its post, the company describes keeping model state — including the KV cache built up while generating a response — explicitly placed and local, then activating the right mix of compute, memory and networking for each phase of inference.

If that holds up, it is the interesting part. Throughput per watt and per-user latency normally pull against each other: you batch harder to fill the accelerator, and the individual request waits longer. A design that moves both at once is a claim about serving architecture, not about a faster multiplier.

Three things to hold onto

• OpenAI ran the benchmarks. These are vendor-measured results on a third-party harness, not an independent evaluation. SemiAnalysis maintains InferenceX; it has not published a review of these runs.
• The comparison is a generation behind what NVIDIA is selling next. Blackwell is what you can buy today. NVIDIA spent the same conference week promoting its successor.
• Almost nothing ships this year. Ho said Jalapeño deploys at the end of 2026 in very small volumes, with meaningful deployment arriving in 2027.

That second point deserves its own paragraph, because the timing is not subtle. On 24 August — the day before OpenAI's talk — NVIDIA published measured results claiming its Vera Rubin NVL72 racks deliver up to 30x higher throughput per megawatt than GB300 NVL72 on agentic coding workloads. Those numbers were measured by NVIDIA, on a different SemiAnalysis harness called AgentX, and NVIDIA notes they are pending SemiAnalysis review. So the week produced two large efficiency claims, both self-measured, both on benchmarks from the same analyst firm, and neither independently confirmed. Read them as vendor positioning that will be adjudicated later, because that is what they are.

What actually changes for people building on these models

This week: nothing. No API price moved, no model changed, no capacity opened up. A chip that ships in very small volumes at the end of 2026 does not touch your token bill.

What it changes is the 2027 forecast, and specifically who controls it. OpenAI is one of the largest buyers of inference compute in the world, and the cost of serving a Codex session or a long-context request is currently set by hardware it rents rather than builds. Jalapeño is the beginning of an alternative — not a replacement for NVIDIA capacity, which OpenAI continues to contract at gigawatt scale, but a second source with different economics. If a meaningful share of OpenAI's own traffic eventually runs on silicon OpenAI designed, the floor under API pricing becomes partly an internal decision.

The workload the chip is tuned for is also worth noticing. Prefill cost, communication overhead and KV cache locality are the three things that dominate exactly the pattern agentic coding tools produce: long accumulated context, many turns, tool calls in between. This is hardware aimed at the workload that has been growing fastest, not at chat. Whether any of the efficiency reaches developers as lower prices is a separate question, and a commercial one rather than a technical one.

A note on sourcing

OpenAI's own results page serves a bot challenge to non-browser clients, so we could not open it directly. We verified the post's existence, exact title and publication time through OpenAI's official news feed, and the figures, quotes and framing above come from OpenAI's press call and its post as reported by TechCrunch, plus the June announcement issued through Broadcom. Specific performance multiples circulating in secondary coverage are not repeated here, because we could not open a source that states them.