OpenAI designed its Jalapeño chip in under 20 months using its own LLMs and Google's open-source XLS tool

IEEE Spectrum reports that OpenAI has taped out its first internal accelerator, Jalapeño, and that the team built it from concept to silicon in under 20 months, with 9 of those months from RTL to tape-out. The reporter is Matthew S. Smith, writing on 14 September. The point that matters for a software team is what the team chose to feed the models.

The tool the models could read

Instead of writing Verilog by hand, OpenAI used Google's open-source Accelerated Hardware Synthesis, known as XLS. XLS lets a designer write in DSLX, a language that looks like Rust, or in C++, and generates the low-level hardware description from that. Chris Leary, a member of technical staff on the project who previously worked on XLS at Google, told IEEE Spectrum that "XLS in some ways looks like software, so it got that benefit."

The benefit, Leary said, is that LLMs did well because the work stayed in what he called the linguistic domain. A model that can write and edit code is being asked to write and edit code, not to solve a problem it has no training data for. The team started the project on the o3 model, released in April 2025, and finished it on precursors to GPT-6 Astra.

The optimisation number

After first silicon came back from the foundry in May, OpenAI ran its own models over the benchmark software that measures how well the chip actually runs. IEEE Spectrum reports that on DeepSeek's attention kernel, the code moved from 0.31 percent of the theoretical peak to 88.94 percent in about 40 hours of model time. Richard Ho, OpenAI's VP of hardware, told IEEE Spectrum that "the models are giving superpowers to our engineers."

The article notes that a physical-design pass driven by AI reduced the area of the matrix-multiplication unit by 10 percent. Jalapeño accesses 232 gigabytes of HBM4 memory at 15.4 terabytes per second, and the team building it averaged about 100 people over the project.

Change the tool before you add the model

A team rebuilt its process around what the models are good at. When the team could stay in a language a model can read, it stayed there. When it needed to place transistors on silicon, that stayed with the humans and the physical-design tooling, and AI only came in afterwards to tune the software. The order matters. If a team keeps the old tooling and asks an LLM to work in it, the LLM does badly and the team blames the LLM.

One warning applies to every number here. All of them come from OpenAI itself, through the reporter, and IEEE Spectrum is careful to attribute each claim to a named engineer. Nobody outside OpenAI has measured Jalapeño yet, so the 88.94 percent figure and the 10 percent area reduction are what OpenAI says, in a piece where OpenAI is the subject and the source.

Source

Primary report: IEEE Spectrum, "How OpenAI Used Its Own LLMs to Design Its Jalapeño Chip" by Matthew S. Smith, 14 September 2026. Discovery: Hacker News thread.