OpenAI says its AI improved AMD's GPU design by up to 14%; we found no AMD statement

In short
  • On 18 November 2026 OpenAI said its AI models improved the design of AMD's flagship GPU by up to 14% on a combined power, speed and size measure.
  • AMD's role is OpenAI's account; we found no AMD statement. The 14% is OpenAI's own 'up to' figure, a best result, not an average.
  • If it holds up, chips could get faster or cheaper to design. OpenAI also holds a warrant for up to 160 million AMD shares, which is worth knowing.
  • Up to 14%PPA gain on AMD's GPU designOpenAI's own figure; "up to"
  • Under 5 moTime OpenAI says the work tookAbout a year of progress, it says
  • Single digitsGains in AMD CPU blocksAMD verified these, OpenAI says
  • Up to 160MAMD shares in OpenAI's warrantVest as milestones are met
Poster for OpenAI's chip design work with AMD: the OpenAI logo on black with the words 'AI designing chips'
Photo: TechSpotlight News on Cover card

An AI model has been rewriting the circuit code for one of AMD’s chips, and some of its changes are already in a design sent for manufacturing. That is what OpenAI says in a post published on 18 November 2026.

As OpenAI describes it, the work improved AMD’s “flagship GPU IP” (the design block for its top GPU, the kind of chip that runs AI) by up to 14% on a combined measure of power, performance and area. Every detail about AMD’s part comes from OpenAI. We found no statement from AMD.

In our reading, the real news is the speed OpenAI claims: a year of design progress in under five months, if it holds.

OpenAI AMD chip design: what the model changed

The measure is called PPA: power, performance and area. OpenAI reports one combined figure, not three, and says the gain is relative to “heavily optimized baselines”. It does not split the 14% across the three.

OpenAI says the model worked on the chip’s circuit code, written in SystemVerilog RTL. Tools from Synopsys, a chip-design software company, then confirmed that each rewrite left the block working the same way and measured the trade-offs. OpenAI calls the result “roughly a year of progress” achieved in under five months.

In OpenAI’s telling, AMD has already used some of the changes in a recent design handed to a factory (a step called tape-out) and stored others in its production files for later versions. It adds that AMD verified gains in the single digits across key blocks of its next flagship CPU. No product name or tape-out date is given.

A big claim with little outside checking

Start with the number. AlphaSignal, an AI newsletter, notes the 14% is measured on individual blocks, with no whole-GPU figure. Because OpenAI says “up to”, it is the best result, not the average. OpenAI also does not publish the baselines or the list of blocks, so outsiders cannot easily repeat the test.

Then the interest. OpenAI holds a warrant for up to 160 million AMD shares, announced in AMD’s 6 October 2025 release with shares vesting as milestones are met. That does not make the result wrong. It does mean the claim comes from a company with a stake in AMD’s success.

OpenAI’s own tests are also modest about limits. In its RISC-V tests (RISC-V is an open chip standard), OpenAI says some features were left out, there was no tight size limit, and timing was not checked at the final stage. GPT-6.2 Astra scored best there, but those are lab tasks.

For now, nothing changes for chip buyers; for engineers, watch the tools

For most readers, nothing changes this week. No product, price or launch is tied to these changes.

The longer-term hope is simple. If AI shortens design work, chip makers could release better chips more often, or spend less to do it. Whether savings reach buyers of phones, laptops and cloud services is a separate question that nobody has answered.

For engineers, the post describes a model that works inside experts’ tools and checks, not one that replaces them. OpenAI is also partnering with Synopsys on GPT-Synopsys, a model meant to use design tools the way expert designers do. AlphaSignal reports that GPT-Synopsys has no public product, price or date. Which jobs shift, and how fast, is not yet visible.

In our reading: encouraging, and not yet proven

The good first. Chip design is slow, and a tool that proposes checked improvements to hard blocks could free skilled people for harder problems. Verifying each rewrite with Synopsys tools is a sensible safeguard. And OpenAI says the first silicon of Jalapeño, its own custom inference chip, passed bring-up, the first round of testing, which gives the approach some real-world footing.

The caution is equally plain. A best-case, block-level, self-reported number is a starting point, not a result. In our view, the missing piece is AMD itself: a statement saying which changes it used and what it measured would settle much of this.

On balance, this looks like a meaningful step for AI-assisted design, and the likelier near-term outcome is gradual gains inside existing engineering teams, not a sudden change in how chips are made.

Watch for the first outside test, and for the day AMD says in its own words which changes made it into a chip it ships.

Prefer to watch? Follow us on YouTube

The tech stories that matter, explained in plain words. Subscribe so you never miss one.

Subscribe

Quick answers

Is OpenAI designing AMD's chips?

OpenAI says its model rewrote parts of the circuit code for AMD's flagship GPU design and that AMD used some of the changes. We found no AMD statement confirming this.

Can AI design computer chips?

AI can now help with parts of the job. OpenAI says its model rewrote circuit code and checked its own work with Synopsys tools. Whether this scales to whole chips is not shown.

Is the 14% for the whole chip?

No. OpenAI gives it as an 'up to' gain on its flagship GPU design block, with no whole-chip figure.

What does PPA mean in chip design?

It stands for power, performance and area: energy use, speed and the silicon a chip takes up.

What does tape-out mean?

It is the hand-off of a finished chip design to a factory to be made.

Words to know

RTL
Register-transfer level: the code that describes how a chip's circuits move and process data.
PPA
Power, performance and area: three things chip designers balance. Better PPA means less energy, more speed or a smaller chip.
Tape-out
The hand-off of a finished chip design to a factory for manufacturing.
Chip IP
A ready-made design block, such as a top GPU (the kind of chip that runs AI), that a company builds a product from.
Sources (4): every key fact checked against at least two

Related stories