On November 18, OpenAI published a blog post breaking down how its own models take part in chip design: GPT-4o has been helping design Jalapeño, the company's first custom inference chip, since early 2025, and the changes the models proposed cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit — already proven in silicon. The same approach has also been applied to AMD's production lines, delivering up to 14% PPA (power, performance, area) improvement on flagship GPU IP against heavily optimized baselines. This time OpenAI is talking in concrete numbers at the circuit level, not concepts.
Both key improvements in Jalapeño sit in the compute cores. In the vector unit, the models worked from floating-point adders and multipliers, conversion functions, and reduction logic, trying different logic implementations and pipeline partitions to find chances to share hardware and cut intermediate storage. In the matrix unit, one multiplier redesign cut synthesized area by about 15% while dropping latency from two clock cycles to one, and an adder redesign cut area by about 15% while keeping its two-cycle latency. Added together, these changes deliver a double-digit percentage area saving — all verified on real silicon, not paper numbers from simulation.
The AMD collaboration goes further. OpenAI's models rewrite the SystemVerilog code that describes circuits directly; AMD has put some of those changes into a recent design sent for manufacturing and added others to production design files for future versions. On key blocks of AMD's upcoming flagship CPU, model-written designs also delivered single-digit percentage PPA gains. OpenAI describes this as roughly a year of progress achieved in less than five months. The same evaluation method was also applied to the open-source RISC-V cores PicoRV32 and Ibex: with a budget of about 112,000 tokens per run, GPT-6.2 Astra posted the highest average score on both cores, roughly twice GPT-6 Astra's score on PicoRV32.
Two things are worth watching next. First is OpenAI's partnership with Synopsys: the two companies plan to jointly train a model called GPT-Synopsys that uses Synopsys EDA tools the way expert chip designers do — that will decide whether model-led chip design becomes a routine workflow rather than a one-off. Second is the limits of the evaluation itself: when designing a core from written requirements, models can leave out features such as performance counters that contribute little to throughput, inflating scores, and the evaluation did not tighten area limits or check timing after physical synthesis, so it remains some distance from production standards. That models can change chips is now a fact; what comes next is proving they can do it reliably, and get it right the first time.