OpenAI has working inference silicon and no public benchmark

OpenAI has crossed the line from a chip plan to working silicon. The company and Broadcom say engineering samples of Jalapeño are running machine-learning workloads at their target frequency and power, including GPT-5.3-Codex-Spark.

That is substantial progress for a custom accelerator developed from initial design to manufacturing tape-out in nine months. It also leaves the question buyers care about unanswered: how does Jalapeño perform against a named alternative on a workload anyone can inspect?

The Broadcom release, issued jointly with OpenAI on June 24, calls Jalapeño OpenAI's first “Intelligence Processor.” OpenAI designed the architecture around the kernels, memory movement, networking, and serving patterns used by ChatGPT, Codex, the API, and future agentic products. Broadcom handled silicon implementation and networking. Celestica supports board, rack, and system integration.

The chip is built specifically for inference. OpenAI says the architecture reduces data movement and balances compute, memory, and networking resources so more of the chip's theoretical capacity becomes usable throughput. Engineering samples are already running at production target frequency and power.

Here is the public evidence as it stands:

Broadcom says early testing shows substantially better performance per watt than the current state of the art. The release supplies no benchmark, baseline accelerator, workload, batch size, latency target, or measured power figure. It also supplies no public percentage for cost reduction. Those omissions limit what can be concluded from the performance claim.

The company perspective is straightforward. OpenAI can tune silicon around its own serving stack, trade generality for utilization, and reduce dependence on merchant accelerators. It also gains control over a layer of infrastructure that shapes both product latency and gross margin. Broadcom gains a multi-generation custom-compute program plus networking demand. The companies say the platform is intended to reach gigawatt scale with Microsoft and other data-center partners, beginning with an initial deployment targeted for the end of 2026.

Operators have a different decision. Working samples do not change this quarter's architecture or API budget. The first economic benefit lands inside OpenAI's serving operation, where lower infrastructure cost can improve capacity, latency, margin, or some combination of the three. Customers receive savings only when OpenAI changes prices or competition forces the issue.

Jalapeño also changes the capacity conversation. A production deployment gives OpenAI another source of inference compute and a hardware path optimized for its own model families. That could matter during demand spikes or when new agent products produce longer-running inference workloads. The timing remains a forecast, and the release's cautionary language explicitly treats production scale and deployment as forward-looking.

The next useful artifact is the promised technical report. It should identify the model, workload, comparison accelerator, precision, batch size, latency target, measured power, and system boundary used for the performance-per-watt result. Without those fields, “substantially better” remains a supplier claim attached to a working sample.

Watch for that report before the first end-2026 deployment. It will show whether Jalapeño is a broad inference advantage, a narrow optimization for OpenAI's own stack, or an impressive chip whose commercial effect stays inside the data center.

The Signal is the public edge of a private practice. Sherpa points the same intelligence engine at one owner's business — competitors, suppliers, regulators, watched daily, graded and sourced. Work with a Sherpa →