Featured image

OpenAI just built its own chip. Not a GPU. Not a TPU. An ASIC designed from scratch for inference, co-developed with Broadcom, taped out in nine months. The industry standard for a chip like this is eighteen to twenty-four months. OpenAI did it in half the time, and they did it by using their own models to design it.

The chip is called Jalapeño. Engineering samples are already running GPT-5.3-Codex-Spark in the lab at production clock speeds. Initial data center deployment is slated for the end of 2026, in gigawatt-scale facilities with Microsoft and other partners. Broadcom CEO Hock Tan says it's "just as good" as Nvidia's Blackwell GPUs and Google's TPUs for inference workloads.

But here's what most coverage is missing: this isn't an Nvidia killer story. Nvidia's training monopoly is intact. What OpenAI is actually doing is something more interesting and more dangerous to the competitive landscape. They're building the economics layer.

Why inference is the real battleground

Training a frontier model costs hundreds of millions of dollars. But you do it once. Inference, the process of actually running those models for millions of users, is where the money flows continuously. Every ChatGPT query, every Codex completion, every API call burns inference compute. And right now, OpenAI buys that compute from Nvidia at market prices.

The numbers tell the story. Nvidia's data center revenue hit $115 billion last year. Most of that is inference, not training. When you're serving hundreds of millions of users across ChatGPT, Codex, and enterprise APIs, the per-token cost difference between Nvidia's general-purpose GPU and a custom ASIC designed specifically for your workload is enormous. OpenAI claims Jalapeño targets a 50%+ cost reduction compared to Nvidia GPUs for inference. If that's real, and if they can deploy it at scale, the margin improvement alone could fund the next generation of models.

This is the play Google figured out years ago with TPUs. Amazon built Trainium and Inferentia for the same reason. Microsoft has Maia. The pattern is clear: whoever controls the inference hardware controls the economics of AI deployment. OpenAI was the last major AI lab without its own silicon. Now they have it.

The nine-month miracle

The development speed is what should worry competitors most. A typical ASIC goes through architecture design, RTL (register-transfer level) coding, verification, physical design, tape-out, and manufacturing. Industry veterans said eighteen months was aggressive. OpenAI did it in nine.

How? They used their own models to accelerate the process. Architecture search, layout optimization, bug detection in HDL code, all partially automated by the same AI systems that will eventually run on the chip. Greg Brockman confirmed on CNBC that OpenAI's models were used "throughout the design flow." This is the AI-designed-AI-chip feedback loop that people have been speculating about for years. It just happened.

The chip itself is massive. Tom's Hardware analyzed the floorplan and estimates it at roughly 840 square millimeters, a reticle-sized die surrounded by six HBM modules. The architecture is highly repetitive, consistent with tiled AI accelerator designs. Richard Ho, who leads OpenAI's hardware program, said they optimized around "the kernels, memory movement, networking, and serving patterns that matter most for frontier AI models."

Broadcom provided the silicon implementation and the Tomahawk networking silicon that lets thousands of chips talk to each other efficiently. Celestica handles board and system integration. It's a three-way partnership, but the architectural vision is OpenAI's.

What changes for developers

If Jalapeño delivers on its promises, the downstream effects are concrete. Lower inference costs mean OpenAI can either improve margins or pass savings to API developers. Given the current pricing pressure from open-weight models and Chinese labs, expect both. The API pricing wars of 2025-2026 have been brutal, and Jalapeño gives OpenAI a structural cost advantage that pure software optimization can't match.

For the Codex team specifically, this is existential. Coding agents generate massive token volumes. A single Codex session can consume hundreds of thousands of tokens across multiple turns. If inference costs drop 50%, the economics of offering Codex at current price points improve dramatically. That's the "coding margin play" that one analyst called the real story behind the announcement.

There's also the multi-generation roadmap. Broadcom's Hock Tan described Jalapeño as "just the beginning." OpenAI plans to iterate on this architecture annually, each generation getting more optimized for their specific workloads. Within two to three years, they aim to own their compute economics entirely. That's vertical integration on a scale that even Google hasn't fully achieved with TPUs.

The catches

Don't celebrate too fast. The chip is still in the lab. Engineering samples running at production speeds is impressive, but moving from lab to gigawatt-scale deployment in six months is a different kind of challenge. Manufacturing yield, thermal management at scale, reliability over years of continuous operation, none of that is solved by a nine-month tape-out.

Also, the "just as good as Blackwell" claim from Hock Tan needs independent verification. OpenAI hasn't published benchmark numbers. Performance-per-watt claims are internal projections. The technical report is coming "in the coming months," which means we're being asked to trust the marketing before the data.

And there's the organizational elephant. OpenAI is in the middle of a restructuring that's seen most of its founding team depart. The hardware program's credibility rests on Richard Ho's team, but the broader company stability is an open question. Building custom silicon is a multi-decade commitment. OpenAI needs to exist in its current form for that commitment to pay off.

Still, the trajectory is clear. OpenAI is no longer just a model company. It's becoming a full-stack AI infrastructure company, across the entire stack. The question isn't whether this changes the competitive landscape. It's how fast.

Sources