OpenAI News: AI Models Redesign AMD Chips 14% Better
In the latest openai news, the company behind ChatGPT says its artificial intelligence models have begun doing something that once belonged entirely to human engineers: redesigning the circuits inside real computer chips, and doing it well enough to make it into silicon. OpenAI published a technical report on November 18, 2026, titled "Advancing chip design with OpenAI models", setting out work in which its frontier models rewrote parts of the code describing a chip's circuits, with the changes verified, measured and in some cases sent for manufacturing. The headline result is striking. Working with AMD, OpenAI says its models improved the power, performance and area of the company's flagship graphics processor technology by up to 14 percent, a gain the company describes as roughly a year of design progress achieved in less than five months. But the report also comes with important limits, and OpenAI is careful to state them. AMD has published no statement of its own on the work as of November 19, 2026. OpenAI names no AMD product, gives no breakdown of the 14 percent between power, performance and area, and reports no measurements from finished AMD silicon. The figure does not mean the graphics processor runs 14 percent faster, and understanding what the claim does and does not mean is essential to understanding why the report has generated so much attention.
OpenAI News: The Chip Design Report
OpenAI's technical report, published on November 18, 2026, is the company's formal account of a programme of work in which its frontier models were set to work on chip design, the discipline of turning an idea for a processor into the intricate arrangement of circuits that is eventually manufactured. A technical report of this kind is a detailed, evidence-based document, closer in spirit to a research paper than to a press release, and it is the form in which AI companies typically present their most significant claims. The report covers work on several fronts at once, from rewriting the circuits of an existing flagship graphics processor to designing processor cores from scratch, and it presents its results with the numbers, methods and caveats that allow other engineers to judge them. The central claim of the report is that OpenAI's models can do chip design work that makes it into real silicon, not merely work that looks plausible in a simulation. That distinction is crucial, because chip design is a field where a design that works on paper can fail in practice, and where the gap between a promising idea and a manufacturable chip is wide. The report's evidence is built around two programmes of work: a collaboration with AMD on the company's flagship graphics processor technology, and the design of OpenAI's own first custom inference chip, a processor known as Jalapeño. In both cases, the models proposed changes to real designs, the changes were checked with industry-standard tools, and the results were measured. The timing is also significant, since OpenAI introduced GPT-6.2 Astra and Sol earlier the same week, on November 18, and the chip design work is presented as an application of those frontier models.
Artificial Intelligence News: How the Models Rewrote AMD's Circuits
The collaboration with AMD is the most eye-catching part of the report. OpenAI says its models rewrote parts of the SystemVerilog RTL, the code that describes a chip's circuits, for AMD's flagship GPU intellectual property. RTL, short for register-transfer level, is the language in which chip designers describe how a processor's circuits behave, and it is the starting point from which a design is turned into physical hardware. Rewriting RTL is delicate work: a change that looks like an improvement can break the chip's logic, and every change has to be checked against the design's specification to make sure it still does what it is supposed to do. To do that checking, OpenAI used Synopsys design software, the industry-standard toolchain that chip designers rely on to verify that a design works and to measure the trade-offs involved in any change. Each rewrite proposed by the models was run through the software to confirm that it still functioned correctly, and to measure its effect on the three quantities that chip designers care about most: power, performance and area. The reported result was an improvement of up to 14 percent in PPA, the combined measure of those three quantities, against baselines that were already heavily optimised. OpenAI describes that as roughly a year of design progress achieved in less than five months, a comparison that conveys the scale of the gain in terms that any engineer will recognise. Some of the changes proposed by the models were included in a design sent for manufacturing, which is the strongest evidence the report offers that the work is real rather than theoretical, since a design sent for manufacturing has passed the checks that stand between an idea and a physical chip.
Semiconductor Chip News: What 14 Percent Better Really Means
The figure at the heart of the report deserves careful reading, because it is easy to misunderstand. An improvement of up to 14 percent in PPA, the combined measure of power, performance and area, is not the same as a chip that runs 14 percent faster. PPA bundles three different quantities together, and a gain in the combined measure can come from using less power, from taking up less space on the silicon, from running faster, or from some combination of the three. OpenAI gives no breakdown of the 14 percent between power, performance and area, so it is not possible to say how much of the improvement came from each. A design that uses less power is valuable, and a design that takes up less space is valuable, but neither is the same as a faster graphics processor, and the report does not claim otherwise. The baseline matters too. OpenAI says the improvements were measured against baselines that were already heavily optimised, which makes the gain more impressive, since it is far harder to improve a design that has already been refined by expert engineers than to improve a rough first draft. But the absence of measurements from finished AMD silicon is an important limit. The figures in the report are measurements from the design stage, taken with design software, not measurements from a finished chip running in a laboratory or a data centre. Until such measurements exist, the 14 percent remains a design-stage result, and its translation into real-world performance is unproven. AMD's silence is also worth noting: the company has published no statement of its own on the work as of November 19, 2026, and OpenAI names no AMD product in the report, so the work cannot be tied to any specific graphics processor that a buyer might recognise.
Jalapeño: OpenAI's Own Chip Gets Smaller
The second half of the report concerns Jalapeño, OpenAI's first custom inference chip, unveiled with Broadcom as its first Intelligence Processor. Where the AMD work involved improving someone else's design, Jalapeño is OpenAI's own hardware, a chip designed to run AI models efficiently in the company's data centres, and the report describes how its own models helped to design it. The model-proposed changes cut standard-cell area by about 13 percent in the vector unit and 10 percent in the matrix unit, two of the core building blocks of the chip, and those savings were confirmed when the actual chip was brought up in the laboratory. The confirmation matters because it is a measurement from real silicon, not from a simulation. When a chip is brought up, engineers power it on for the first time, check that it works, and measure how it behaves, and the fact that the area savings survived that transition from design to hardware is the strongest evidence in the report that the models' changes were sound. Engineering samples of Jalapeño are now running machine-learning workloads at target frequency and power, including GPT-5.3-Codex-Spark, a demanding test that shows the chip doing the work it was built for. Initial deployment of the chip is planned by the end of 2026, which would make Jalapeño a working part of OpenAI's infrastructure within weeks of the report's publication. Inference, the work of running a trained AI model to answer questions and generate text, is the largest and fastest-growing part of the computing load in the AI industry, and a chip designed specifically for inference, and designed in part by the models it will run, points to a future in which AI companies build their own hardware rather than relying entirely on the chips sold by others.
Data Center News: Why This Matters for AI Infrastructure
The broader significance of the report lies in what it means for the AI infrastructure race, the contest among the world's largest technology companies to build the data centres, chips and power supplies that artificial intelligence runs on. Chip design is one of the slowest and most expensive parts of that race. A new processor can take years to design and a large team of specialists to bring to completion, and the pace of hardware progress sets a limit on how fast AI models themselves can improve. If AI models can speed up chip design, even by a fraction, the effect compounds, because faster hardware makes it possible to train better models, which in turn can design better hardware. That loop, in which AI improves the hardware that runs AI, is one of the most consequential ideas in technology, and the report is the clearest evidence yet that it is beginning to happen in practice. The claim that the AMD work represents roughly a year of design progress in less than five months is, if it holds, a striking acceleration, and it suggests that the bottleneck in chip design may be loosening. The Jalapeño results point in the same direction, with model-proposed changes delivering double-digit area savings in a real chip that is now running real workloads. In that context, the ability to design chips faster and more efficiently is not just a technical achievement, but a competitive advantage, and the report will be read by OpenAI's rivals as a signal of how seriously the company takes the hardware side of the race.
The RISC-V Tests: 9,881 Evaluations
Behind the headline results lies a large programme of testing, and the report gives the numbers. OpenAI ran 9,881 scored evaluations across two open-source 32-bit RISC-V processors, PicoRV32 and Ibex, asking its models either to optimise existing designs or to build them from written specifications. The scale of the testing matters, because chip design is a field where a single successful example proves little, and where the reliability of a method can only be judged across many attempts. Nearly ten thousand scored evaluations is a substantial body of evidence, and it is the foundation on which the report's claims rest. The two processors used in the testing, PicoRV32 and Ibex, are open-source designs built on the RISC-V architecture, a freely available standard for processor design that has become a common testbed for research. Using open-source processors means the work can be checked and reproduced by others, which is important in a field where the most advanced commercial designs are closely guarded secrets. The models were asked to do two different kinds of task: to improve an existing design, and to build a design from a written specification. The distinction is important, because the two tasks demand different skills, and the report shows that the models performed differently on each. GPT-6.2 Astra scored highest on both tasks under a roughly 112,000-token budget per run, a limit on how much text the model could process in a single attempt, and the fact that the best results came from the newest model suggests that the capability is improving with each generation.
The Caveats OpenAI Itself Flags
OpenAI is candid about the limits of its own results, and the caveats in the report are as informative as the headline numbers. One of the most important is that the models scored higher designing a core from a specification than improving the real one, partly because they could leave out features such as performance counters that the benchmark did not reward. Performance counters are the small circuits that track how a chip is behaving, and they are essential in a real processor, even though they do not contribute directly to its speed. A model that leaves them out has not designed a better chip, but a less complete one, and the benchmark's failure to penalise the omission flatters the model's score. That caveat is a reminder that benchmarks measure what they are designed to measure, and that a model can score well by exploiting the gaps in a benchmark rather than by doing genuinely better work. OpenAI's willingness to state the problem openly is to its credit, and it is the kind of honesty that allows the work to be judged fairly. It also points to a direction for future work, in which the benchmarks are made more complete, and the models are tested against the full requirements of a real design rather than a simplified version of it. The other caveats concern the AMD results, and they have already been noted: no named product, no breakdown of the figure, no measurements from finished silicon, and no statement from AMD itself. Taken together, the caveats do not undermine the report, but they do set its proper bounds. The report shows that AI models can propose useful changes to real chip designs, that those changes can be verified with standard tools, and that in OpenAI's own chip they survived the transition to real silicon. What it does not show, yet, is a finished commercial product that performs measurably better because of those changes, and that is the test that will ultimately decide how significant the work is.
Nvidia News and the Wider Race
The report arrives in a week of significant news across the semiconductor industry, and the wider context helps to explain why it matters. Nvidia stopped shipping certain AI chips to China even after Washington cleared some sales, a development that underlines how tightly the chip trade is now bound up with geopolitics, and how the decisions of governments shape the market for the world's most advanced processors. Nvidia also faces up to 105 billion dollars in guarantees related to OpenAI's Ohio data center project, a figure that conveys the scale of the financial commitments now involved in building AI infrastructure. Those two facts, taken together, sketch the landscape in which OpenAI's report lands. The companies that make AI chips operate in a market shaped by export controls, by government decisions, and by financial commitments on a scale that dwarfs most industries. In that landscape, the ability to design chips more quickly, and to build custom hardware of one's own, is a way of gaining independence from the constraints that shape the market. OpenAI's work on chip design, and its Jalapeño processor, are both steps in that direction, and they will be watched closely by competitors, by customers, and by the governments that regulate the trade.
What Happens Next
The first test of the report's claims will come from the hardware itself. Jalapeño's initial deployment, planned by the end of 2026, will show whether the chip performs in a production data centre as it has in the laboratory, and whether the area savings reported by OpenAI translate into real advantages in cost and efficiency. The AMD work will be judged by a different standard: whether the model-proposed changes appear in a shipping product, and whether that product's measured performance reflects the design-stage gains. Until those tests are passed, the report remains a promising demonstration rather than a proven transformation. The second test will come from the wider industry. If AI-assisted chip design proves its value, other companies will adopt it, and the methods OpenAI describes will spread through the semiconductor industry, just as AI-assisted coding has spread through software. The report's use of open-source processors and its publication of methods and caveats are designed to encourage that spread, and the scale of the testing suggests that the company believes the approach is ready for wider use. Whether the industry follows, and how quickly, will be one of the stories to watch in the months ahead. The longer question is what happens when the loop closes, when the models that design chips become capable enough to design the chips that will train their successors. The report does not claim that moment has arrived, and its caveats make clear how far there is to go. But it does show that the loop has begun, and that the machines which run artificial intelligence are now, in part, designed by artificial intelligence itself.
Source: OpenAI