Picture a chip engineer at two in the morning, staring at a circuit diagram that has already been polished for years. Every transistor has been argued over. Every wire has been measured twice. Finding even one percent more efficiency feels like squeezing water from stone.
Now picture the engineer at that desk not being a person at all. The designer is the artificial intelligence itself — rewriting the circuits that will power the next generation of artificial intelligence. That scene stopped being a thought experiment on November 18, when OpenAI published a technical report showing its models doing exactly that, on real production hardware.
OpenAI chip design is no longer a research demo. According to the company’s report, its frontier models rewrote the circuit code of AMD’s flagship graphics processor and improved it by up to 14 percent. The chips in question are not toys. They are the same class of hardware that trains and runs the world’s biggest AI systems.
How OpenAI’s Models Rewrote AMD’s GPU Circuits
The work happened at the deepest level of chip design. OpenAI’s models rewrote SystemVerilog RTL, the code that describes a chip’s actual logic gate by gate. This is not a high-level suggestion or a sketch. It is the blueprint from which silicon is built.
Each rewrite was then checked with Synopsys design tools, the same software human engineers use, to confirm the block still behaved exactly as specified. Only then did the model measure the trade-offs: how much power the new version used, how fast it ran, and how much space it took up.
Engineers combine those three measures into a single score called PPA — power, performance and area. Against a baseline that AMD’s own engineers had already heavily optimised, OpenAI’s models delivered up to 14 percent better PPA on the flagship GPU intellectual property. That is a striking number on a design where the easy wins were taken long ago.
The models did not just tinker with one corner of the chip. They modified computation circuits, control logic and the pathways that move data around. The work touched the parts of the design where engineers normally spend months hunting for small gains. Unite.AI’s coverage of the announcement lays out the full figures behind the 14 percent claim.
A Year’s Progress in Under Five Months
OpenAI described the GPU result as roughly a year of design optimisation compressed into less than five months. For an industry where a single chip generation can take two or three years, that kind of acceleration changes the economics of the whole process.
The result is not a simulation or a benchmark demo. AMD has already folded some of the model-proposed changes into a chip design it has sent for manufacturing. The remaining improvements have been written into production design files, ready for future products.
The collaboration went further than the graphics processor. AMD also verified single-digit percentage PPA gains across key blocks of its next-generation flagship CPU, using designs written by OpenAI’s models. The company has not named the products involved, and it has not published its own figures — the numbers come from OpenAI’s report.
Context matters here. OpenAI and AMD already have a close commercial relationship. The two companies signed a GPU supply agreement in October 2025, and OpenAI holds rights to buy up to 160 million AMD shares. The first planned deployment uses AMD’s Instinct MI450 series GPUs. The chip industry has been moving fast all month — TSMC recently committed $1 billion to Anthropic and OpenAI, a sign of how tightly the AI and semiconductor worlds are now bound.
Jalapeño: OpenAI’s Own Chip Gets the Same Treatment
Before touching AMD’s hardware, OpenAI practised on its own. The company began using GPT-4o to help design Jalapeño, its first custom inference chip, in early 2025. In an August 2026 update, OpenAI said AI assistance had helped move Jalapeño from its initial design to tapeout — the point where the design is sent to the factory — in nine months.
The new report adds detail to that story. Model-proposed changes reduced the standard-cell area of Jalapeño’s vector compute unit by about 13 percent, and its matrix compute unit by about 10 percent. Smaller cells mean a smaller, cheaper chip that does the same work.
Jalapeño matters because it is the chip OpenAI will use to run its own models. If the company can design its inference hardware faster and cheaper, every ChatGPT answer gets a little less expensive to produce. That is the loop the industry has been waiting for: AI making the machines that run AI.
GPT-Synopsys: Teaching AI to Use the Tools of the Trade
OpenAI is not stopping at one-off experiments. On September 30, 2026, the company announced a multi-year preferred-partner deal with Synopsys, the giant of electronic design automation software. Together they are training a specialised model called GPT-Synopsys.
The goal is a model that can run industry chip design tools the way expert engineers do — not just writing circuit code, but verifying it, testing it and refining it inside the same software professionals rely on. The deal includes joint go-to-market plans and revenue sharing.
To measure how fast this capability is improving, OpenAI ran nearly ten thousand scored tests across two open-source 32-bit RISC-V processors, PicoRV32 and Ibex. It tested five model generations, from GPT-5.5 up to the newly released GPT-6.2 Astra. Under a budget of about 112,000 tokens per run, GPT-6.2 Astra scored highest on both processors.
The progression tells its own story. Each new model generation handled the chip design tasks better than the last. What looked like a curiosity two years ago is becoming a measurable, repeatable engineering skill.
The Caveats OpenAI Didn’t Hide
To its credit, OpenAI was open about the limits of the work. The 14 percent figure is a design-stage metric, not a speed increase you will see in a finished graphics card. OpenAI gave no breakdown of how much of the gain came from power, speed or area, and it reported no measurements from finished AMD silicon.
There is another wrinkle. When a model redesigns a processor from its specification, it can quietly drop features that were in the original chip, such as performance counters used for debugging. That may explain part of the higher scores. Timing was also not checked after physical synthesis, the step where the design is turned into a physical layout.
None of that makes the result unimportant. It means the headline number should be read carefully. The models proved they can find genuine improvements in production designs. They have not yet proved they can take a chip all the way from specification to a working product on their own.
What Happens Next
The obvious question is how far this goes. If models keep improving at this pace, the next generation could handle larger blocks, full subsystems and eventually whole chips. Google’s Gemini 4 Carbon is racing on the same frontier, and every lab now knows the hardware race and the model race are one race. The bottleneck may shift from writing the design to verifying it — the slow, careful work of proving a chip does what it should.
That is where GPT-Synopsys comes in. A model that can operate the verification tools itself would close the loop, letting AI design and check its own work the way a senior engineer does. OpenAI’s partnership with Synopsys suggests that is exactly the plan.
There is a bigger shift hiding underneath. For decades, chip design has been one of the most human-intensive jobs in technology, a craft passed down through teams of specialists. If AI can genuinely contribute at the circuit level, the pace of hardware progress itself could accelerate — and faster hardware means faster AI, which means faster chip design again.
The cycle is already turning. AMD’s next chips will carry circuits that an AI model rewrote. OpenAI’s own Jalapeño chip was designed with AI help from the start. And somewhere, a chip engineer is looking at a circuit diagram at two in the morning, with a model beside them pointing out the one percent nobody else could find.
The machines are learning to build the machines. This time, the proof is in the silicon.



