OpenAI published research on 18 November 2026 showing that its AI models are already improving real chip designs, including its first custom inference chip, Jalapeño, and processor designs from AMD. According to OpenAI’s write-up, engineers began using GPT-4o on Jalapeño in early 2025, and some of the changes its models suggested are now proven in working silicon. The work matters beyond OpenAI: if AI can shorten the years-long cycle of designing chips, it changes the economics of the hardware that AI itself runs on. This article explains what OpenAI did, the results it reports, the caveats, how it fits into the history of AI-assisted chip design and why it matters.
Key results at a glance
| Project | Reported result |
|---|---|
| Jalapeño warp engine | About 13% less standard-cell area |
| Jalapeño matrix unit | About 10% less standard-cell area |
| Jalapeño multiplier | About 15% less cell area; latency cut from two clock cycles to one |
| Jalapeño adder | About 15% less cell area at the same two-cycle latency |
| AMD flagship GPU IP | Up to 14% improvement in power, performance and area (PPA) |
| AMD next-generation flagship CPU | Single-digit percentage PPA gains across key blocks, verified by AMD |
| RISC-V benchmark | GPT-6.2 Astra scored highest on both test cores |
Background: what is Jalapeño?
OpenAI and Broadcom unveiled Jalapeño on 24 June 2026. OpenAI calls it its “first Intelligence Processor”, an accelerator designed from a blank slate for large language model inference, which is the work of running trained models to answer user requests, rather than a general-purpose chip adapted from training hardware. According to OpenAI’s announcement, engineering samples were running workloads in the lab at production target frequency and power, including a version of OpenAI’s Codex model.
Engadget reports OpenAI’s claim that Jalapeño was co-developed “from initial design to manufacturing tape-out in just nine months”, a very fast cycle for a large custom chip. Tom’s Hardware describes it as a reticle-sized ASIC, meaning it is about as large as a single chip can be manufactured. OpenAI says the architecture reduces data movement, one of the main bottlenecks in inference, and claims performance per watt substantially better than current state-of-the-art hardware. Initial deployment is planned for the end of 2026, scaling to gigawatt-class data centres with Microsoft and other partners, with Celestica handling board, rack and system integration.
The new research explains part of how that nine-month timeline was achieved: OpenAI’s own models were used to optimise the chip’s circuits during development.
What OpenAI’s models did on Jalapeño
Chip design moves through several stages. Engineers describe the logic of each block in a hardware description language such as SystemVerilog, a level known as register-transfer level or RTL. Tools then turn that description into a physical layout of standard cells, the basic building blocks of a chip. The size, speed and power use of the final chip depend heavily on how efficiently the RTL is written.
OpenAI’s models rewrote RTL for key blocks of Jalapeño. In the warp engine, savings came from floating-point adders and multipliers, conversion functions, reduction logic and the special-function unit, and the models also found ways to share hardware and reduce intermediate storage. In the matrix unit, which performs the multiplications at the heart of neural networks, a redesigned multiplier used about 15% less area and finished its calculation in one clock cycle instead of two. These results were confirmed during chip bring-up, the stage when first silicon is tested, so they are measured in real hardware rather than only in simulation.
Why do these numbers matter? Area translates directly into cost and capacity: a smaller block leaves room for more compute on the same die or lets the same design be made more cheaply. Cutting a multiplier’s latency in half speeds up every operation that depends on it. Across millions of chips in a data centre fleet, single-digit or low double-digit percentage gains add up to large savings in money and power.
The AMD collaboration
OpenAI’s work with AMD tested the same approach on one of the industry’s most heavily optimised designs. The model rewrote SystemVerilog for AMD’s flagship GPU intellectual property and, according to OpenAI, reasoned at the level of an expert microarchitect. Results reached up to a 14% improvement in power, performance and area over baselines that AMD’s engineers had already tuned extensively.
Every rewrite was checked with Synopsys tools to confirm that the new logic behaves exactly like the original, a step known as equivalence checking, and to evaluate the trade-offs. AMD included some changes in a design recently sent for manufacturing and added others to production files for future versions. AMD also verified single-digit percentage gains across key blocks of its upcoming next-generation flagship CPU. OpenAI describes the GPU progress as “roughly a year of progress achieved in less than 5 months.”
The RISC-V benchmark: how the models compare
To compare its models, OpenAI tested them on two open-source 32-bit RISC-V processor cores: PicoRV32, a compact core common on FPGAs, and Ibex, an embedded processor used in the OpenTitan security chip. The study ran 9,881 scored attempts across two tasks.
- Task 1, optimise existing RTL: models tried to increase throughput. Designs scored zero if they failed simulation checks or grew beyond 16 times the original area. At about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores, roughly double GPT-6 Astra’s score on PicoRV32. GPT-6.1 Sol also outscored the older GPT-5.6 Sol.
- Task 2, design from a specification: all five models scored higher than on Task 1. At the same budget, GPT-6.2 Astra averaged about 0.7 on PicoRV32 and 0.75 on Ibex.
The pattern is consistent with other benchmarks: newer, larger models given more tokens to reason produce better hardware designs.
Synopsys and GPT-Synopsys
Synopsys is one of the two dominant suppliers of electronic design automation software, the tools almost every chip company uses. OpenAI is working with Synopsys to connect its models directly to those tools, and the companies are training a new model, GPT-Synopsys, to use them the way expert chip designers do. If that succeeds, AI assistance could reach chip teams through software they already use rather than through separate research projects.
The caveats OpenAI lists
OpenAI is explicit about the limits of the RISC-V tests. Designs written from a specification could leave out features of the original cores, such as performance counters and extra instructions, which may have inflated scores. The tests did not impose a tight area limit or check timing after physical synthesis, both of which real production designs must pass. Scores include zeros for failed attempts. And the Jalapeño and AMD results come from OpenAI and its partners; they have not been independently replicated.
It is also important to be clear about what the models did not do. They did not design entire chips on their own. They redesigned specific circuits and blocks, with human engineers directing the work and verification tools checking every change.
A short history of AI in chip design
Using AI to design chips is not new, but the approach has changed. Google researchers published work in 2021 on using reinforcement learning to place components on a chip’s floor plan, a method Google later named AlphaChip and used in its own TPU accelerators. Synopsys and Cadence have offered machine-learning tools since around 2020 that automatically tune the settings of their design flows. Those systems optimise specific steps.
What is different about OpenAI’s work is that a general-purpose language model is reading and rewriting the design itself, at the RTL level where architects work, and explaining its reasoning. That is closer to having an expert engineer suggest improvements than to tuning a tool’s parameters. It also means the same model that writes software can contribute to hardware.
Why it matters: the AI hardware feedback loop
The most important implication is a feedback loop. Better models help design better chips; better chips make it cheaper to train and run better models. OpenAI’s latest system card tracks AI self-improvement as a risk category for this kind of reason, and rated GPT-6.2 Astra as the first externally released model to reach its High threshold.
There is also a competitive dimension. Nvidia dominates AI hardware, and the largest AI companies are building their own chips to reduce cost and dependence. Microsoft introduced its Maia 200 inference chip in January 2026, according to VentureBeat, and Google has designed TPUs for years. If AI-assisted design shortens development cycles, companies with the best models gain an edge in building the hardware those models run on, and the gap between them and everyone else could widen.
For the chip industry, the near-term effect is practical: faster iteration, smaller and more efficient designs, and verification-heavy workflows where AI proposes and tools prove. AMD’s decision to put some of the model’s changes into designs headed for manufacturing is the strongest signal that the approach is moving from research into production.
What it means for engineers
- Verification becomes more important, not less. AI-generated RTL is only useful if equivalence checking and timing analysis prove it correct.
- Architects shift toward review. Engineers spend more time directing and judging AI proposals than writing every block by hand.
- Tool integration is key. Partnerships such as OpenAI and Synopsys will decide how quickly these capabilities reach everyday design teams.
How a language model improves hardware code, step by step
OpenAI’s write-up describes the results rather than every step of its workflow, but the process it outlines follows a loop familiar to chip engineers, with the model taking on the drafting role:
- Start from working RTL. The model receives the existing SystemVerilog for a block, such as a multiplier, along with the goal: less area, lower latency or higher throughput.
- Propose a rewrite. The model restructures the logic, for example by sharing hardware between operations, simplifying arithmetic or reducing intermediate storage, and explains its reasoning.
- Check that behaviour is unchanged. Verification tools, in this case from Synopsys, confirm the new design produces exactly the same outputs as the original for every input.
- Measure the trade-offs. Synthesis tools estimate area, timing and power so engineers can compare the rewrite with the baseline.
- Repeat or accept. Engineers keep improvements that pass verification and meet their targets, and feed failures back to the model.
The key point is that the model never needs to be trusted blindly. Hardware design already has rigorous tools for proving that two designs are equivalent, which makes it an unusually good fit for AI assistance: the model can be creative, and the tools catch mistakes before anything reaches silicon.
Why OpenAI started with an inference chip
Training a frontier model is a large, one-off cost. Serving it to hundreds of millions of users every day is a cost that grows with usage and never stops. As ChatGPT, Codex and the API have scaled, inference has become the larger long-term expense for AI companies, which is why OpenAI’s first chip targets it. A chip designed specifically for the patterns of LLM inference, such as moving model weights and attention data efficiently, can deliver more answers per watt than a general-purpose GPU built to handle training as well.
OpenAI emphasised low operating costs when running real-time coding models, according to TechCrunch. Axios noted that OpenAI has said it is considering whether to extend its own chips to training as well. If AI-assisted design keeps shortening development cycles, building a training chip becomes a more realistic option.
What it means for Nvidia and Broadcom
For Broadcom, which turns customers’ designs into manufacturable chips and supplies networking, OpenAI’s success is a strong reference for its custom-silicon business. For Nvidia, it is a reminder that its largest customers are also becoming competitors in parts of the market. Nvidia’s flagship Rubin graphics cards can run both training and inference workloads, as SiliconANGLE noted, and their flexibility and mature software keep them in demand. But if custom inference chips deliver meaningfully better performance per watt for specific workloads, AI companies will route more of their serving traffic to their own hardware over time.
The AMD collaboration adds another twist: OpenAI is not only building its own chips but helping a major chipmaker improve its products. That gives OpenAI influence over the hardware roadmap of a key alternative supplier to Nvidia.
Questions that remain
As VentureBeat pointed out when Jalapeño was unveiled, important questions remain unanswered: how the chip performs against direct competitors on real workloads, what it costs and how well it can be manufactured at scale. Independent benchmarks will only arrive once the chip is deployed in data centres, which OpenAI expects to begin by the end of 2026. Similarly, the AI-assisted design gains will be most convincing when other chip companies report similar results with their own teams and tools.
Key takeaways
- Proven in silicon: OpenAI’s model-suggested changes to Jalapeño were confirmed in working chips during bring-up, not just in simulation.
- Meaningful gains: area savings of roughly 10% to 15% on key blocks, and a multiplier that runs in one clock cycle instead of two.
- Beyond OpenAI: AMD has put some of the model’s changes into a design sent for manufacturing, the clearest sign the approach is reaching production.
- Newer models do better: GPT-6.2 Astra led OpenAI’s RISC-V tests, roughly doubling GPT-6 Astra’s score on one core.
- Verification is the safety net: every rewrite was checked with Synopsys tools, which is what makes AI suggestions usable in hardware.
- Caveats apply: the results are self-reported, and the RISC-V tests skipped some production checks such as post-synthesis timing.
The broader significance is speed. Chip development has traditionally taken years from architecture to silicon, which limits how quickly hardware can adapt to new AI workloads. OpenAI’s claim that Jalapeño went from design to tape-out in nine months, with its own models helping along the way, suggests that timeline is compressing. If that holds across the industry, the pace of AI hardware improvement could start to track the pace of AI software improvement more closely, with consequences for cost, competition and how quickly new capabilities reach users.
FAQ
What is OpenAI’s Jalapeño chip?
Jalapeño is OpenAI’s first custom chip, built with Broadcom and designed specifically for LLM inference. It was unveiled in June 2026, with initial deployment planned by the end of 2026.
Did AI design the whole chip?
No. OpenAI’s models redesigned specific circuits and blocks, with engineers directing the work and verification tools checking every change.
How much did OpenAI’s models improve the chip?
OpenAI reports about 13% less area in Jalapeño’s warp engine, about 10% in its matrix unit and about 15% in individual multiplier and adder designs.
Which OpenAI model performed best?
In OpenAI’s RISC-V tests, GPT-6.2 Astra had the highest average score on both cores tested.
What is GPT-Synopsys?
A model OpenAI and Synopsys are training to operate Synopsys’s chip-design software the way expert designers do.
Sources
- OpenAI: Advancing chip design with our models
- OpenAI: OpenAI and Broadcom unveil LLM-optimized inference chip
- TechCrunch: OpenAI unveils its first custom chip
- Tom’s Hardware: Jalapeño’s nine-month development cycle
- VentureBeat: Jalapeño and OpenAI’s chip strategy
- The Neuron: AI news digest, 19 November 2026
Published
Spotted an error? Read our corrections policy or email editorial@perplexityaimagazine.com.