OpenAI says its frontier models rewrote the circuit code for one of AMD's flagship GPUs and delivered up to 14% better power, performance and area. The company also says the same approach cut circuit area on its own Jalapeño inference chip. Both claims come from a company report, and no independent party has verified them.

The report, titled "Advancing chip design with OpenAI models" and dated November 18, describes AI systems rewriting SystemVerilog register-transfer level code, the text that defines how a chip's logic behaves. The details below come from press coverage of the report, not the report itself. Tom's Hardware summarized the GPU gains as up to 14% in PPA, the shorthand for power, performance and area. The comparison point matters. The gains were measured against a design AMD had already tuned heavily, not a rough first draft.

How OpenAI says its models rewrote AMD's GPU circuits

The timing claim draws the most attention. OpenAI reports that the work compressed roughly a year of design progress into less than five months, as AlphaSignal reported. TheNextWeb framed the same result as 14% gains in AMD's flagship GPU design.

Some of the changes have already moved beyond a test environment. Wccftech reported that some AMD changes are already in a chip AMD sent to manufacturing, while others sit in production design files for later versions. That split is the most concrete signal in the report. It suggests model-written logic reached a real production pipeline.

OpenAI also reports single-digit gains on a key part of AMD's next flagship CPU. According to TheNextWeb, AMD has reportedly confirmed that result. The CPU gain is smaller than the GPU figure, but it suggests the method is not confined to one block of one chip.

What the numbers do and do not cover

Claim Reported figure Main source Status
AMD flagship GPU blocks Up to 14% better PPA Wccftech Company-reported, block-level
Design progress About a year of progress in under five months AlphaSignal Company-reported
Jalapeño standard-cell area About 13%, and about 10% in two key blocks Inside AI News Outlets disagree; one says up to 15%
Jalapeño design to tape-out Nine months with AI help Perplexity AI Magazine Company-reported, secondary coverage
AMD next-generation CPU key block Single-digit gains TheNextWeb AMD reportedly confirmed

The 14% figure is a best case for specific, unnamed GPU blocks. It does not describe whole-GPU performance, and the published numbers do not break down power, speed and area separately. Readers should treat it as a claim about a few blocks, not a faster GPU.

Which AMD products include OpenAI's changes

AMD has not named the products that contain the work. PC Gamer framed the outcome as AMD using some of the tweaks in future products. That means the changes may eventually reach silicon customers buy, but no buyer can check that today.

For people shopping for current AMD GPUs and AI accelerators, that gap is the central point. Efficiency gains, if they arrive, will show up in product specifications and third-party tests, not in a research post.

Jalapeño, OpenAI's own inference chip

Jalapeño is the more unusual part of the story because OpenAI is designing its own silicon. According to Cryptopolitan and Complete AI Training, OpenAI began using earlier models to help design the chip in early 2025. Jalapeño is OpenAI's first custom inference chip, meaning hardware built to run trained models rather than train them.

Perplexity AI Magazine reported that the chip went from design to tape-out in nine months with AI assistance, starting with GPT-4o. Model-proposed changes cut circuit area, and those changes held up in silicon during bring-up, according to Tom's Hardware. Inside AI News put standard-cell area cuts at about 13% and about 10% in two key blocks.

Perplexity gave "up to 15%" for the same chip, and the block names differ across sources. Until OpenAI publishes a clear breakdown, the exact Jalapeño figure remains unsettled.

OpenAI's broader chip push adds context. Reuters reported earlier that OpenAI was building its first chip with Broadcom and TSMC and sourcing from several chipmakers, according to a MarketScreener copy of that report.

Why the 14% figure needs a cautious read

The limits are substantial. AlphaSignal's coverage notes that no independent audit of the GPU result or public design files has been released, and that OpenAI has commercial interests tied to AMD. An outside test would carry more weight than a result published by a company that also sells AI services to AMD's customers.

Complete AI Training described the study as an admitted limited benchmark. Designing from a written specification lets models drop features such as performance counters, and timing was not checked after physical synthesis. Those gaps matter because a chip can meet a spec on paper and still miss its timing targets in production.

The report also said all five tested models scored higher when designing a RISC-V core from a written specification than when optimizing an existing design, according to Inside AI News and Complete AI Training. If that holds, it suggests models do better building from scratch than squeezing gains from a mature design. That is a narrower lesson than the headline suggests.

Synopsys and the GPT-Synopsys effort

Every rewrite was checked with Synopsys electronic design automation tools, the software engineers use to design and verify chips, to confirm the circuit still behaved as specified. That check appears in DEV Community's summary and in AlphaSignal's coverage.

OpenAI is also working with Synopsys on a model called GPT-Synopsys, meant to operate chip-design software the way expert engineers do. Cryptopolitan and Tom's Hardware both describe the project. The available reporting does not give a timeline for it.

Synopsys sits in the middle of this story. If models can drive design tools directly, value could shift toward whoever controls that workflow. EDA vendors have a strong reason to build that layer themselves.

What this means for chip designers, AMD buyers and investors

Chip designers and EDA vendors have the most direct stake. If the results hold, design cycles could shorten, which is the core claim in the coverage. The AMD result also shows AI applied to production chip work, not only software. Earlier AI-for-chip efforts mostly produced benchmarks rather than shipping silicon.

Buyers face a slower path. Because AMD has not named the affected products, any efficiency gains remain speculative for now.

Investors in AMD, Nvidia and OpenAI-linked supply chains will watch whether these claims appear in product launches. Nvidia Posts Record $115.8 Billion Quarter Without Jensen Huang shows how much money already flows into AI hardware, so even modest efficiency gains draw attention. The Federal Reserve has also warned that the AI buildout is now an inflation and stability risk, as USA 24 reported, which gives chip efficiency a wider policy relevance.

What to check before trusting vendor chip claims

  • Ask whether a PPA figure covers one block or a whole chip, and request the breakdown across power, performance and area.
  • Wait for a named product and a published specification before adding any efficiency gain to purchasing or capacity plans.
  • Look for independent benchmarks run on shipping hardware, not on design files or sample silicon.
  • Check whether a claim refers to chips already sent to manufacturing or to designs slated for later versions.

Consumers have no steps to take now. These results concern future chips, not anything you can buy this week.

What to watch next

The next signals will come from AMD's product announcements, which should show whether any of these changes reach shipping parts. Watch for an independent audit or published design files, a clear reconciliation of the Jalapeño area figures, and a timeline for GPT-Synopsys. Each of those would turn a company claim into something buyers and engineers can test, and the proof that matters most will come from silicon customers can actually buy.