OpenAI designed a chip in 20 months with help from its own AI
OpenAI built its Jalapeño inference chip in 20 months, using its own language models at every stage of development alongside partner Broadcom.
On August 25, OpenAI unveiled Jalapeño, its first chip built for processing AI model queries. According to IEEE Spectrum, the device delivers up to 13.4 petaflops of compute at 4-bit precision, works with 232 gigabytes of next-generation HBM4 memory, and exchanges data with it at 15.4 terabytes per second. By the company's own measurements, the latency between a query and the last word of a response on Jalapeño is up to 3.6 times lower than on the Nvidia GB300 chip OpenAI currently uses. The new chip is also more energy-efficient.
The numbers are notable on their own, but the main point of the publication lies elsewhere. Less than 20 months passed from the first architectural idea to finished silicon. From writing the code describing the chip's logic to handing the design off to the factory took just nine months. For developing specialized processors, that is remarkably fast. According to IEEE Spectrum, the speed came not from a huge team but from engineers actively using OpenAI's own language models at every stage of design.
Readers who run a shop, a service business, or a small manufacturing operation have no direct stake in chips. But behind this story lie two questions relevant to any business that uses AI services: why cloud AI might get cheaper and faster over time, and what lesson from OpenAI's approach can be applied to processes that have nothing to do with electronics.
What OpenAI actually built
Jalapeño is not a general-purpose processor but a specialized chip for inference — that is, for running already-trained models rather than training neural networks from scratch. It responds to user queries instead of teaching AI systems from the ground up. Such chips are grouped into large clusters — according to IEEE Spectrum, Jalapeño is designed for deployment in clusters of 2,048 units. Inference is precisely the most massive and expensive line item for companies like OpenAI, because every ChatGPT query requires computation on the spot.
The development team averaged fewer than 100 people over the course of the project — roughly the same number, according to Richard Ho, OpenAI's vice president of hardware, that is now working on the chip's second and third generations. OpenAI did not build the chip entirely on its own: Broadcom served as its partner. The division of labor was straightforward: OpenAI's team handled the overall system architecture — the accelerator itself, memory organization, and networking between chips. Broadcom handled the physical implementation — everything “from the logic gates on down,” as Ho put it to IEEE Spectrum.
Experts consulted by the publication were divided in their assessments. David Chin and Ravi Krishna, founders of the startup Verkor.io, which automates chip design through AI agents, called OpenAI's pace impressive but noted that without an experienced partner like Broadcom, such speed would have been impossible. Professor Andrew Kahng of the University of California, San Diego, called OpenAI's result likely the best among comparable projects today.
Where AI actually helped
Chip design splits into two large parts. The “front end” is the path from concept and logic coding to verifying that the circuit will work correctly. The “back end” is the physical implementation: routing connections inside the die, calculating clock signals and power consumption, and preparing data for the factory.
According to IEEE Spectrum, most of the benefit from AI came on the front end. The team built its workflow around the open-source tool XLS, originally developed at Google. It lets engineers describe chip logic in languages resembling ordinary programming languages — DSLX (a dialect similar to Rust) and C++ — and then automatically converts that code into the Verilog hardware description language. OpenAI's Chris Leary explained the reasoning simply: language models perform far better on tasks resembling ordinary programming, and XLS is close in style to programming itself. Professor Ankur Srivastava of the University of Maryland added that automation in chip design has existed for decades, but language models are the first tool to understand both human language and code at once, rather than merely executing predefined rules.
A telling example from the report is software optimization after the first chips returned from the factory in May. The team pointed its internal models at tuning benchmark programs, including one for a key computational block from the DeepSeek model. The result rose from 0.31 percent of the chip's theoretical ceiling to 88.94 percent — in 40 hours. Ho stressed that the result is reproducible, and the company now factors that speed into planning for future projects.
The back end is a different story. Here Broadcom did most of the work, and it lacked access to OpenAI's internal, specially fine-tuned models — only to publicly available commercial versions. Even so, AI assistance produced a measurable effect: as Ho and Leary said at the Hot Chips 2026 conference, AI-driven optimization shrank the area of matrix multiplication blocks by 10 percent compared with a manually optimized version — meaning more circuitry fit on the same piece of silicon. The Verkor founders consider this back-end approach fairly conservative for the current level of models and believe the next generation of chips could automate this part of the process much further.
Models not yet publicly available
One notable detail from IEEE Spectrum: the project began using the o3 model and finished using predecessors of GPT-6 Astra, which had not yet been released to the wider public at the time of publication. The newer model can work directly with Verilog, skipping the intermediate translation step through XLS, and, according to Leary, is close to being able to operate specialized design tools on its own.
Ho confirmed that the team also had access to internal models specially fine-tuned for chip-design tasks that are not publicly disclosed. He stressed that the goal is to fold this accumulated experience into the company's commercial products, and that future model versions, including Astra, should handle hardware design tasks reasonably well “out of the box.”
It's worth pausing here rather than concluding that engineers have been fully replaced by robots. Ho himself directly pushed back on that interpretation in conversation with the publication: the team is not claiming that anyone can walk in and, using the Codex coding tool, build a top-tier accelerator. The point is that a small team with a properly structured process and current models delivers results noticeably faster than before. Final decisions still rest with engineers.
Why this matters for the cost and speed of cloud AI
For businesses paying for ChatGPT subscriptions or using AI APIs in chatbots or advertising tools, there is one practical takeaway from this story: the faster and cheaper companies like OpenAI can build their own inference hardware, the lower their potential cost of processing requests. Inference — serving millions of everyday user queries, rather than training models — is what consumes the bulk of computing capacity at large AI companies.
The claimed reduction in latency and energy consumption compared with Nvidia chips — if it holds up in practice at scale, something IEEE Spectrum itself treats cautiously — is one of the few real levers such companies have to cut costs without raising prices for customers. The report makes no direct promises about lower subscription prices, and it would be premature to predict one. But the broader trend is clear: AI infrastructure is becoming more efficient, and its refresh cycle is shortening. That means the quality and speed of services businesses rely on could improve faster than it seems today, and competition among cloud AI providers could intensify.
What this means for small businesses in Russia and the CIS
The Jalapeño story carries no direct impact on small businesses in Russia, Belarus, or Kazakhstan — access to Jalapeño itself and similar infrastructure is closed off to local companies for geopolitical and technical reasons. But the lesson in how OpenAI organized its work applies directly, and it goes like this.
The company did not hand the entire chip design over to AI. It broke a complex task into understandable pieces, identified the parts that resembled ordinary programming, and gave the models maximum freedom exactly there. Where years of engineering experience and physical intuition were needed, it left the decisions to people and to a partner with the right expertise. The same principle works for any company thinking about where to integrate AI into its processes:
- identify tasks in your work that are structurally similar to text or code — customer correspondence, drafting proposals, analyzing reviews, initial sorting of requests — and hand those to language models first, since that's where returns come fastest;
- do not try to automate everything at once — OpenAI kept its team under a hundred people and deliberately kept some work in-house and with its partner, because AI is still weaker than humans in certain areas;
- track a measurable result, as the Jalapeño team did with its benchmark test: a jump from 0.31 to 88.94 percent in 40 hours is easy to explain to management and to replicate on the next project, whereas vague impressions of “it got faster” don't help with decisions;
- watch how the cost and speed of the AI services available to you change — if major providers lower the cost of inference, you gain a chance to get more capability for the same money, and it's worth periodically reviewing the plans and tools your company uses;
- don't expect AI to fully replace a specialist on complex technical or legal tasks — even in a project with an almost unlimited budget, OpenAI left final decisions to its engineers.
Bottom line
The Jalapeño story is less a hardware breakthrough than a working example of how to embed AI into a real, complex, multi-stage process in a way that delivers a measurable time savings. For a small business owner, the useful part isn't the petaflops or the terabytes per second, but the underlying logic: break the work into pieces, find where the model is strong, trust it there, and check the results with numbers rather than impressions. Everything else comes down to how quickly the cloud AI a business relies on gets cheaper and faster — and that no longer depends on any individual entrepreneur, but on how fast companies like OpenAI keep updating their own infrastructure.