OpenAI unveiled its first custom AI chip called Jalapeño on June 24, 2026, co-built with Broadcom and manufactured by TSMC. It runs large language model inference faster and cheaper than current alternatives, was designed in nine months using OpenAI’s own AI models and hits data centers before end of 2026. This is OpenAI’s clearest signal yet that it’s building a hardware company behind the AI company.
Every time ChatGPT answered a question for the last three years, it ran on Nvidia hardware. OpenAI paid Nvidia’s prices, worked within Nvidia’s constraints and competed for GPU supply with every other AI company on earth.
That just ended.
On June 24, 2026, OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom-designed AI chip, built for one job: running large language models as efficiently as possible.
What Is the Jalapeño Chip Exactly?
Jalapeño is OpenAI’s first custom AI inference chip, unveiled June 24, 2026, co-designed with Broadcom and manufactured by TSMC on a 3-nanometer process. Built from scratch for LLM inference rather than adapted from a general-purpose accelerator, it delivers better performance per watt than current leading AI chips. It was designed in nine months with help from OpenAI’s own AI models and will be deployed in data centers by end of 2026.
OpenAI was unusually blunt about what this chip is: “Jalapeño is not a general-purpose accelerator modified from an existing AI chip. It is a chip newly designed from the ground up for large language model inference based on our experience operating ChatGPT and Codex.“
That distinction matters more than it sounds. Most AI chips try to do everything. Jalapeño does one thing. And when you design a chip to do one thing, you can go significantly further on efficiency than any general-purpose design allows.
“Jalapeño is not a general-purpose accelerator modified from an existing AI chip. It is an AI chip newly designed from the ground up for large language model inference based on our experience operating ChatGPT and Codex. It can interoperate with other LLMs as well.”
— OpenAI, Official Announcement, June 24, 2026
Why Is Inference the Right Target for a Custom Chip?
Training gets all the press. It’s the part where you build the model from scratch on thousands of GPUs for months. Expensive, rare and genuinely complex.
But inference is where the actual money goes.
Inference is what happens every single time a user asks ChatGPT something.
That happens billions of times a day. Training happens a handful of times a year. So when you’re running a product at the scale OpenAI operates, the cost of inference dwarfs everything else.
Here’s where things get interesting.
Inference has completely different computational requirements from training. It’s more repetitive, more predictable and far more sensitive to memory bandwidth than raw compute. A chip designed specifically for inference doesn’t need to be good at training. It just needs to move data efficiently and run matrix operations quickly. That’s a much cleaner design target.
Nvidia’s GPUs are brilliant at both.
But they’re not built specifically for either. Jalapeño is built specifically for inference. That focus is the whole point.
How Did OpenAI Build a Custom Chip in Just Nine Months?
Custom chip development normally takes two to three years. Jalapeño went from initial design to tapeout in nine months.
Let’s look at why.
OpenAI used its own AI models to help design the chip. Greg Brockman explained it directly: “You take components that humans have already optimized, pour compute into it and the model comes up with its own improvements. It discovered optimizations that would have taken human engineers weeks to find.”
That’s not a throwaway line.
Those AI-discovered optimizations produced “massive area reductions,” meaning circuitry that takes up less space on the chip, which translates directly into better efficiency and lower manufacturing cost per unit.
The manufacturing partner is TSMC on a 3-nanometer process, the same generation used in Apple’s latest processors and Nvidia’s current AI chips. Broadcom handled the chip design collaboration. Celestica is lined up for deployment.
What Does the Jalapeño Chip Actually Do Differently?
Three technical design decisions define Jalapeño:
It Minimizes Data Movement
Moving data between memory and compute is where most power gets wasted in AI inference. Jalapeño co-designs compute, memory and networking as an integrated system rather than bolting together off-the-shelf components. The result is a chip where data moves less far and less often. That shows up as better performance per watt.
It Runs on TSMC’s Best Process
3-nanometer transistors mean more compute per square millimeter of silicon and lower power draw per operation. This isn’t unique to Jalapeño but it’s table stakes for competing with the latest generation of AI accelerators.
It Has Built-In Networking
LLM inference at scale doesn’t run on one chip. It runs across many chips coordinating in real time. Tight network integration reduces the latency that chip-to-chip communication normally introduces. This is especially important for the kind of large model inference that powers GPT-5.5 and future OpenAI products.
How Does Jalapeño’s Performance Stack Up?
OpenAI hasn’t published full benchmark numbers yet. The chip is still moving through testing before data center deployment.
But there is one claim worth noting. Early results show better performance per watt than current state-of-the-art AI accelerators. The unnamed comparison point is almost certainly Nvidia’s H100 and H200 series, which power the majority of large-scale LLM inference today.
Full independent benchmarks will follow as the chip reaches production. Until then, the performance claim is OpenAI’s own and should be read accordingly.
Why Did OpenAI Actually Build Its Own Chip?
There are four real reasons. Not the marketing ones.
- Supply chain independence. Nvidia has dominant pricing power over AI chips right now. Every time OpenAI scaled up, it had to compete for GPU supply with Amazon, Google, Meta and every other AI company simultaneously. Custom silicon ends that dependency.
- Cost per query. ChatGPT runs at billions of queries per day. Even a small improvement in inference efficiency compounds into enormous savings at that scale. Inference-optimized hardware cuts the cost of every single answer the product gives.
- Full-stack control. OpenAI has been building toward owning the entire technology stack. Models, software, infrastructure and now silicon. When you control the hardware, you can co-optimize the model and the chip together in ways that are simply impossible when you’re running on someone else’s general-purpose hardware.
- Everyone else already did this. Google has TPUs. Amazon has Trainium and Inferentia. Meta is building its own chips. Microsoft has the Maia chip. OpenAI being the last major AI company running entirely on Nvidia hardware was becoming a structural disadvantage. Jalapeño closes that gap.
“The chip was built to improve efficiency and lower costs, advancing OpenAI’s strategy to build out a full stack behind its models and products.”
— The Wall Street Journal, June 24, 2026
How Does Jalapeño Compare to the Competition?
| Company | Custom Chip | Focus |
|---|---|---|
| OpenAI | Jalapeño | LLM inference only |
| TPU v5 | Training and inference | |
| Amazon | Trainium 2, Inferentia 3 | Training and inference |
| Meta | MTIA 2 | Inference |
| Microsoft | Maia 100 | Training and inference |
| Groq | LPU | LLM inference |
| Nvidia | H100, B200 | General purpose |
The pattern here is that every hyperscaler narrows their chip focus differently. OpenAI chose the narrowest target of all: inference only, LLM only. That’s a confident design bet that their inference workload is distinct enough and large enough to justify a chip that does nothing else. Given that ChatGPT is the highest-traffic AI product in the world, that bet isn’t unreasonable.
What Do the Deployment Numbers Actually Tell Us?
OpenAI plans to deploy up to 10 gigawatts of custom compute between 2026 and 2029.
To put that in context: a single gigawatt of AI compute infrastructure is roughly the scale of a large hyperscale cloud data center. Ten gigawatts isn’t an experiment. It’s a complete rebuild of how OpenAI provisions compute at the infrastructure level.
Jalapeño starts hitting real-world data centers before end of 2026. The chip will not be sold to anyone else. It runs OpenAI’s products and nothing else.
What Actually Changes for ChatGPT Users?
Nothing immediately visible. Jalapeño rolls out gradually into data centers over the coming months. You won’t see a notification in ChatGPT announcing that your query just ran on an OpenAI chip.
But the effects build over time:
- Faster response times as inference latency drops on optimized hardware
- Lower operating costs for OpenAI which gives the company more room on pricing and free-tier capacity
- Better product reliability as OpenAI stops competing for scarce third-party GPU supply
- Tighter model and hardware co-optimization as future models get designed knowing exactly what hardware they’ll run
The 10-gigawatt commitment by 2029 is the signal to take seriously. That’s not a hedge. That’s a full infrastructure rebuild and OpenAI is clearly willing to fund it.
Frequently Asked Questions
Jalapeño is OpenAI’s first custom AI chip, unveiled June 24, 2026. It was purpose-built for LLM inference to run ChatGPT and Codex faster and more efficiently than general-purpose GPUs.
Broadcom co-designed the chip with OpenAI. TSMC manufactures it on a 3-nanometer process. Celestica is the confirmed deployment partner for data center rollout.
Nine months from design to tapeout. OpenAI used its own AI models to accelerate the process and discover optimizations that would have taken engineers weeks to find manually.
Early results show better performance per watt than current leading AI chips. Full benchmarks are pending as the chip moves from testing into production deployment.
Data center deployment begins before end of 2026. OpenAI plans to roll out up to 10 gigawatts of custom compute between 2026 and 2029.
No. It is built exclusively for OpenAI’s internal products including ChatGPT, Codex and the API. There are no plans to sell it commercially.
Yes. OpenAI confirmed it can interoperate with other large language models though its primary purpose is running OpenAI’s own products and API.
Conclusion
Jalapeño is OpenAI saying out loud what it has been building toward for years: it wants to own the entire stack from the model to the silicon it runs on. Nine months to design. Billions of queries to run. Ten gigawatts of compute by 2029. This isn’t a chip announcement. It’s a hardware company announcement dressed up as one.
Disclaimer
This article is based on publicly available information from OpenAI’s official announcement, Reuters, TechCrunch, The Wall Street Journal, AP News and other sources as of June 25, 2026. Performance benchmarks are preliminary. Deployment timelines and scale targets may change. This article is for informational purposes only.
