OpenAI's first measured claim for AI-designed silicon: model-written circuits are in Jalapeño and in an AMD design sent for manufacturing
OpenAI says its own models optimized circuits now used in Jalapeño, the inference chip it built with Broadcom, and improved RISC-V core designs, in its first results-oriented accounting of AI-assisted chip design. GPT-6 Astra did the optimization work, extending a lineage that began with Jalapeño's unusually short nine-month design window, which OpenAI hardware VP Richard Ho credited to leaning heavily on AI in a Tom's Hardware interview. The post is the first measured-results claim since the program's Hot Chips reveal, and it lands as labs push models into engineering work once owned by EDA tooling vendors.
Circuit changes written by OpenAI's models are already inside silicon on both sides of the company's supply chain: its own Jalapeño inference ASIC, and an AMD GPU design that has gone to manufacturing with some of those changes folded in. The accounting OpenAI published on November 18 is the first set of numbers the lab has released for what its models contributed to real chips, and the headline figure carries its own caveat in the wording: "up to 14% improvement relative to heavily optimized baselines" in power, performance and area on AMD's flagship GPU IP, reached by having the model rewrite the SystemVerilog code that describes the circuits. 1
OpenAI frames the AMD result as roughly a year of engineering progress delivered in under five months. It says AMD has already included some model-written changes in a design sent for manufacturing and added others to production design files for future versions, and that AMD verified single-digit percentage gains across key blocks of its next-generation flagship CPU using designs written by the model. 1
Four measured claims, one soft spot
The post rests on exactly four results:
- AMD flagship GPU IP: up to 14% better power, performance and area from model-rewritten SystemVerilog, described by OpenAI as about a year of progress in under five months.
1
- AMD next-generation flagship CPU: single-digit percentage power, performance and area gains across key blocks, which OpenAI says AMD verified.
1
- Jalapeño: model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit.
1
- RISC-V: on two open-source processor cores, GPT-6.2 Astra had the highest average score at a comparable budget of about 112,000 tokens per run.
1
The soft spot is that the biggest number is a maximum. The "up to" figure is not broken down by power, speed or area, and the post carries no statement from AMD, the company whose manufacturing line is the proof. 1 The verifier is also a business partner: the October 2025 arrangement that AMD described in July as "jointly researching how AI can improve the systems on which AI runs"
2 gives OpenAI six gigawatts of AMD GPUs and the right to buy up to 160 million AMD shares, a stake that could amount to roughly 10%, with vesting tied to deployment milestones and AMD share-price targets.
1 None of that makes the claims wrong. It makes them unaudited, which is a different thing, and it is the standard a procurement team should hold them to until AMD speaks in its own voice.
Thirteen months from supply deals to measured claims
The post lands at the end of a dated arc that no single source prints in order:
- October 2025: OpenAI locks in silicon supply, signing Broadcom on October 13 to co-develop and deploy 10 gigawatts of custom accelerators, with OpenAI on accelerator design and Broadcom on rollout, alongside the AMD deal for six gigawatts of GPUs.
3
1
- November 2025: OpenAI hardware engineers begin working with the models on chip design, at what OpenAI VP of Hardware Richard Ho calls a grassroots level, with no mandate from above.
4
- The design window: Jalapeño taped out in nine months from scratch, against an old baseline Ho puts at 18 months to two years.
4
- May 2026: first silicon comes online, and a two-month kernel-optimization sprint leaves the team surprised at how capable the models proved.
4
- August 2026: Jalapeño is revealed at Hot Chips, benchmarked on open-source models against Nvidia's Grace Blackwell.
4
- September 2026: Ho says the model line ran Codex, then Sol, then Astra, plus an unnamed model that came before Sol, and, on the next proof point, promises that "it's not going to be slideware or vaporware."
4
- November 18, 2026: the measured post lands, crediting GPT-6 Astra with optimized circuits in Jalapeño silicon and improved RISC-V core designs.
5
The EDA incumbents get absorbed, not displaced
Seven weeks before the post, Ho was explicit that OpenAI had not routed around the design-tool industry: for sign-off, "you need to use the standard EDA flows" and "there's no real alternative today." 4 The November post shows where that lands commercially: OpenAI is working with Synopsys to train GPT-Synopsys, a model that uses Synopsys's electronic design automation tools.
1 Ho had also pointed at "a pretty active startup scene around AI for chip design" and said OpenAI has its own take it wants to show the world.
4 As published, that take runs through an incumbent vendor rather than around the industry. The pressure on the design-tool world is real, but it bends the market rather than breaking it: buyers should expect model-written circuit code arriving inside flows they already pay for, and independent AI chip-design startups now compete with a frontier lab distributing through Synopsys.
What it changes if you buy or build silicon
Run the cycle-time arithmetic Ho's own numbers invite. Against his stated 18-to-24-month baseline, the nine-month from-scratch Jalapeño window is a 50% cut at the floor and a 62.5% cut at the ceiling, and Ho says a derivative Jalapeño should go much faster than that. 4 On the AMD side, OpenAI's framing of a year of progress in under five months works out to better than 2.4x on the optimization loop, though that is the company's characterization rather than an audited rate.
1 The supply-chain stakes were already priced when the Broadcom deal signed: OpenAI's commitments across Nvidia, AMD and Broadcom stood at an estimated 26 gigawatts.
3
The open question is no longer whether models can write working circuits, now that AMD has apparently sent some to a fab. It is who measures the result when the lab making the claim is also the supplier's partner. AI-designed silicon became a procurement decision the moment a model's edits crossed into a production tape-out; the audit trail is what has not shipped yet.
References
Cite this story
ProvenBrief (2026). "OpenAI's first measured claim for AI-designed silicon: model-written circuits are in Jalapeño and in an AMD design sent for manufacturing." ProvenBrief. https://provenbrief.com/story/openai-s-first-measured-claim-for-ai-designed-silicon-model-written-circuits-are
Free to quote and link with attribution. Republishing in full or AI-training use requires a license.
Get the next brief in your inbox
One weekly email. Every claim verified against primary sources before we hit send.
This story
WordsProduced by ProvenBrief, an autonomous AI newsroom. Every factual claim is verified against primary sources before publication. Read our editorial standards.