TL;DR
– OpenAI and Broadcom pushed Jalapeño from empty whiteboard to manufacturing tape-out in nine months. The fastest high-performance ASIC cycle anyone has publicly claimed.
– It’s inference-only. No training. Engineering samples already executing GPT-5.3-Codex-Spark workloads.
– Gigawatt-scale data centers coming online with Microsoft by end of 2026. Broader rollout follows.
– OpenAI ran its own AI models through the design and verification loops. Broadcom handled silicon. Celestica took the raw chip and turned it into rack-mountable hardware.
Nine months. That’s what OpenAI and Broadcom needed to go from zero to a working chip.
A reticle-sized inference ASIC that pushes right up against the physical limits of what you can etch onto silicon.
The industry typically budgets 18 to 24 months for this kind of chip.
OpenAI and Broadcom cut that in half. Maybe more than half, depending on which part of the timeline you start counting from.
Broadcom’s Hock Tan and Charlie Kawwas hand-delivered the first samples to Sam Altman and Greg Brockman.
Yeah. In person. That’s not a courtesy call. That’s a joint statement about how much both companies have riding on this.
The numbers from announcement day: #1 on Hacker News, 800+ points, 457 comments by June 24.
If you’re building on the OpenAI API, this matters. Straight up.
What Exactly Is Jalapeño?
It’s a reticle-sized ASIC. Single massive die. The kind of chip that makes foundry engineers nervous because you’re playing at the edge of what the equipment can actually print.
Not a GPU.
Not a training chip. Built for one job: running inference on massive language models at scale.
What does that mean in practice? They’re not chasing raw compute numbers. They’re chasing the bottlenecks that actually hurt when you’re handling millions of concurrent API calls.
Data movement, the balance between compute and memory, how fast chips can talk to each other across a rack.
OpenAI’s been writing enormous checks to Nvidia for years.
Everyone knows this. The GPU giant’s margins are absurd. Jalapeño is OpenAI’s way of saying: we don’t want to keep paying the toll forever.
Officially, they’re calling it the first step in a multi-generation compute platform. Not a one-off. And here’s the detail that doesn’t show up in the spec sheets: they built it to work with LLMs across the industry, not just their own. That leaves the door cracked open for third-party access.
Which would rewrite who controls inference infrastructure.
That’s not a given.
But it’s on the table.
Nine Months. How Is That Even Possible?
Let’s be clear about the timeline. Traditional advanced ASIC work: 18 to 24 months minimum. OpenAI and Broadcom hit tape-out in nine.
Both companies are claiming “fastest ever” for a high-performance chip at this scale.
Given the reticle size, that’s not marketing fluff. That’s genuinely aggressive scheduling.
One reason it happened: OpenAI put its own AI models to work on the design and verification tasks. This isn’t a throwaway line. Using LLMs to help with chip verification is something Google has experimented with internally for years. The fact that OpenAI applied it to their first chip. Their actual first chip. Tells you something about where their model capabilities sit behind closed doors.
Broadcom brought the Tomahawk networking silicon to connect multiple Jalapeño chips at scale.
Celestica handled board design, rack integration, and system assembly. Turning the raw silicon into something you can actually slot into a data center. Every layer of the stack has a named partner. That’s unusual for a first-generation design.
What This Means for Your API Bill
Here’s where it gets practical.
OpenAI’s early numbers show meaningfully better performance-per-watt versus current alternatives. They specifically called out real-time coding models as a target workload. If those numbers hold at production scale, the cost structure for serving inference shifts.
When your cost per token drops, you have choices. Keep the savings as margin. Pass them to customers.
Reopen projects where the token math didn’t previously work.
My agency runs agentic products on top of the OpenAI API. We track token costs on every client project. There are automations we’ve passed on since the inference bill made the economics ugly. Cheaper inference changes which conversations are worth having.
But here’s the thing nobody’s talking about enough.
OpenAI now owns the models, the API pricing, the consumer products, and — with Jalapeño.
The silicon those models run on. That’s vertical integration. Full stack. The question isn’t whether they can lower prices. They clearly can. The question is whether they pass the savings down or keep the margins. I’m not betting against eventual price cuts. But I’m watching the actual pricing announcements, not the chip launches.
That’s the honest answer.
The Nvidia Problem Nobody Wants to Admit
Every major AI lab is building custom silicon now.
Google has TPUs. Amazon has Trainium. Now OpenAI has Jalapeño.
Nvidia still owns training. That doesn’t change soon. The training market runs on raw GPU muscle that custom ASICs aren’t built to match. But inference is where the volume lives. Every API call you make lives in inference land.
Broadcom confirmed Jalapeño goes into gigawatt-scale data centers with Microsoft starting end of 2026. The plan is deployment by year-end, broader expansion in following years. Whether it serves third-party models or only OpenAI’s own workloads — that’s still unclear. They say it was built for all LLMs. That answer is not the same as “we will offer it to third parties.”
Here’s what I’m doing in my shop. Keeping client work model-agnostic where the use case allows it. Using OpenAI when it’s the best tool. But maintaining the ability to pivot to Anthropic or open-source alternatives when the economics shift.
Jalapeño makes OpenAI’s infrastructure story stronger. But it also sharpens the lock-in risk. Build on their platform. Keep your exit ramp visible.
The move I’d make right now: start tracking inference costs by model and provider. Not next quarter. Today. When pricing shifts hit. And they will — you want to already know your numbers so you can move fast. The companies that win the next phase of AI adoption will be the ones who treated inference cost as a line item they actively managed, not a black box they paid without questioning.
Side note: their joint announcement blog post is 1,200 words and somehow says less than this summary. Worth a read if you want the official framing.
Sources
OpenAI announcement | TechCrunch coverage | Tom’s Hardware analysis
