OpenAI and AMD Find Up to 14% PPA Gain in Flagship GPU Block in Five Months
OpenAI has shown that its models can improve the chip itself. Working with AMD, they looked for headroom in a flagship GPU block that engineers had optimized by hand for years. In under five months, they found a PPA (power, performance, die area) gain of up to 14%. The same approach helped improve Jalapeño, OpenAI's first custom inference chip, where the gain has been verified. For the industry, this is a signal that AI in hardware design is moving from demos into the actual workflow.
AI-processed from X: @OpenAI; edited by Hamidun News
AI in development is usually discussed as an assistant that writes code or answers questions. OpenAI suggests looking at another side: models are starting to improve the hardware they run on. In its announcement, the company said that together with AMD it found up to 14% PPA improvement in a flagship GPU block (IP) in under five months. And this is a design AMD engineers had already worked on for years.
What PPA is and why 14% is a lot
PPA stands for Power, Performance, Area: power consumption, performance and die area. The three parameters pull in different directions. Want more speed, and power consumption rises. Want a more compact chip, and speed suffers. Chip designers have spent decades balancing them by hand and with specialized electronic design automation software.
That is why "up to 14%" on a mature design sounds serious. If a team has already polished a block, every percentage point is hard-won. Picture a kitchen where a head chef has spent years rearranging equipment to work faster. A newcomer walks in and finds a way to cut a noticeable share of the steps between the stove and the counter. That is roughly the scale of the story OpenAI describes.
An important caveat: "up to 14%" means the maximum, not an average across the whole chip. In the announcement, OpenAI does not disclose which of the three metrics hit the peak, which models were used, or how much human oversight was required. These details will determine how reproducible the method is.
Jalapeño: the first in-house inference chip
The second part of the announcement is about Jalapeño, OpenAI's first custom chip for inference, meaning running already-trained models. According to the company, the same approach helped improve this chip, and the gain has been verified. That is an important detail: the story is not limited to the partner project with AMD, as the method was also applied to OpenAI's own hardware.
For practitioners, the connection is direct. Inference is now the main cost item for companies that sell access to models. The cost of a token is built from the price of chips, electricity, and how many requests each accelerator serves. Improving power or area at the same speed means more requests per watt and per square millimeter of silicon. The announcement gives no figures for Jalapeño, so it is not yet possible to estimate the economics in dollars per token.
What this means for developers and founders
First, the chip design cycle, which stretches over years, could potentially get shorter. Five months to find improvements in an already optimized block is a benchmark that traditional teams will have to be compared against.
Second, labs with frontier models gain another lever. They can not only buy accelerators but also influence their design. The OpenAI–AMD partnership shows that chipmakers are willing to let such models into their projects.
Third, for startups building design automation tools, this is both a threat and a confirmation of demand. Large labs are entering this niche, but the task remains highly specific: it requires access to proprietary designs, simulators and verification of results.
Honest limitations
The source is OpenAI's own statement, and there are no independent measurements. The company claims the Jalapeño gain was verified, but the post itself does not describe the methodology. An improvement at the design level is not yet an improvement in finished silicon: between a design and a production chip there are many stages at which the result can change.
There are three things to watch: whether details of the methodology appear, whether AMD confirms the result in its products, and whether real figures for Jalapeño in operation are shown. If so, "AI designs hardware for AI" will stop being a catchy phrase and become a routine item in a chip development plan.
- Nov 19, 2026Amazon and Google will hold 28.4% of Anthropic after its IPO, but without voting rights
- Nov 19, 2026Tencent Open-Sources Hy4 flash: 12B of 224B Parameters Active, 1M-Token Context
- Nov 19, 2026Anthropic Heads for the Stock Market: $17 Billion in Quarterly Revenue and $895.4 Billion in Server Contracts
- Nov 19, 2026AI token prices fell 21.2% in six weeks, while corporate API spending dropped 6.3%
Key Questions Answered
What PPA is and why 14% is a lot?
PPA stands for Power, Performance, Area: power consumption, performance and die area. The three parameters pull in different directions. Want more speed, and power consumption rises. Want a more compact chip, and speed suffers. Chip designers have spent decades balancing them by hand and with specialized electronic design automation software. That is why "up to 14%" on a mature design sounds serious. If a team has already polished a block, every percentage point is hard-won.
What are the key details on Jalapeño: the first in-house inference chip?
The second part of the announcement is about Jalapeño, OpenAI's first custom chip for inference, meaning running already-trained models. According to the company, the same approach helped improve this chip, and the gain has been verified. That is an important detail: the story is not limited to the partner project with AMD, as the method was also applied to OpenAI's own hardware. For practitioners, the connection is direct. Inference is now the main cost item for companies that sell access to models.
What this means for developers and founders?
First, the chip design cycle, which stretches over years, could potentially get shorter. Five months to find improvements in an already optimized block is a benchmark that traditional teams will have to be compared against. Second, labs with frontier models gain another lever. They can not only buy accelerators but also influence their design. The OpenAI–AMD partnership shows that chipmakers are willing to let such models into their projects.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.