Skip to content

Tech

OpenAI's Jalapeño chip is spicy, but it ships in "very small volumes"

OpenAI and Broadcom built an inference chip and published benchmarks against Nvidia Blackwell. The catch: real deployment starts tiny at the end of 2026, with the bigger push in 2027.

Cover photo: Jernej Furman from Slovenia / CC BY 2.0

TL;DR

  1. OpenAI published benchmarks for Jalapeño, its Broadcom-built inference chip, compared against Nvidia Blackwell systems.
  2. OpenAI says it serves more work per unit of power, with faster responses.
  3. Rollout: very small volumes at end of 2026, more significant deployment in 2027.

What happened

OpenAI has published benchmarks for Jalapeño, its custom inference chip built in close collaboration with Broadcom. First announced last October, it is meant for serving models to users, not training them.

The pitch, per OpenAI: Jalapeño does more AI work for each unit of power and returns answers faster than Nvidia Blackwell systems. That is a two-for-one on the metrics that make inference bills hurt.

Richard Ho, OpenAI's head of hardware, called the results a "very, very significant" advance over the state of the art during a press call. Vendor claim, vendor stage, so we file it under "promising, pending outside verification".

One detail that sounds like a plot twist: OpenAI's own models assisted in the development of the chip. The robots are now helping design their own housing, sapiens.

The company also describes Jalapeño as a multigenerational platform. The idea is a full-stack approach where products, models, chips, and memory are developed together. That lets the team target specific phases of inference that tend to create friction, including delays around prefill.

Why it matters

Inference is where the money is spent every time someone asks a chatbot to rewrite an email. A chip tuned for power efficiency and low latency goes straight at serving costs and response speed.

There is also a strategy angle. A major AI lab designing its own silicon with Broadcom signals that relying on off-the-shelf GPUs is no longer the only plan.

Now the fine print. The comparison is against Blackwell, and the competition may move a lot by the time Jalapeño reaches full deployment. Ho's own estimate: a trickle of deployment at the end of 2026, then a much larger rollout in 2027. In other words, Nvidia keeps its crown for now, but the challenger has a date on the calendar.

In other words, this is a benchmark slide today and a data-center fixture later. Nvidia is not standing still, and a chip in a press deck serves zero customers.

What to do with it

  • Treat the numbers as a vendor claim.Wait for independent tests before rewriting any infrastructure plan around Jalapeño.
  • Watch your API bill, not the chip.If serving gets cheaper for OpenAI, pricing and latency are where developers would feel it, though nothing is promised.
  • Keep your stack portable.Custom silicon from labs means more hardware variety, so avoid hard-coding assumptions about a single vendor.
  • Mark your calendar for late 2026 and 2027.That is when small volumes should turn into something measurable.

Ciao, sapiens. The chip is hot, but the pepper is not out of the jar yet.