OpenAI and Broadcom Unveil Jalapeño, a Custom AI Chip Designed in Nine Months With the Help of the AI It Will Run — The Recursion Has a Deployment Timeline and It Costs 50% Less Than Nvidia

🤚 The Open-Palm Unveiling

OpenAI and Broadcom have unveiled Jalapeño, OpenAI’s first custom-built AI chip — an inference-only ASIC that was designed in nine months, fabricated by TSMC, and promises 50% cost savings compared to the Nvidia GPUs that currently power every existential dread you’ve ever typed into a chatbot.

The chip targets the memory bottleneck that limits GPU efficiency on large language model workloads. Its architecture “reduces data movement” between logic circuits and off-chip memory — which is engineer-speak for “we stopped sending data on a scenic tour of the circuit board before letting it do math.”

Key facts from the announcement:

First Jalapeño servers expected online by end of 2026
Next chip version planned for 2028, then annually thereafter
“Significantly higher performance per watt” than current state-of-the-art alternatives
OpenAI’s own AI models were used to accelerate the chip’s design process
Systems will use Broadcom’s Tomahawk 6 networking chips, capable of processing 1.6 terabits per second

OpenAI President Greg Brockman explained: “We have a deep understanding of the workload… how can we build something that will be able to accelerate what’s possible?” — a question that, until recently, had only one answer, and that answer was “give Jensen Huang more money.”

👐 The Two-Handed Silicon Strategy

Let us be clear about what is happening here. The company that consumes the most AI chips on Earth has decided to make AI chips on Earth. This is the computational equivalent of a restaurant chain buying the cattle ranch, except the cattle ranch costs 10 gigawatts of power and the menu only serves inference.

Broadcom CEO Hock Tan confirmed the 50% cost savings figure and suggested OpenAI and Broadcom could exceed his earlier forecast for deploying 1.3 gigawatts worth of chips next year. To put that in perspective, 1.3 gigawatts is roughly the output of a nuclear power plant. OpenAI would like ten of those, eventually.

Training, however, will still run on Nvidia hardware. Jalapeño is strictly an inference chip — it handles the part where users ask ChatGPT to write their resignation letters, not the part where the model learns what a resignation letter is. This means Nvidia keeps its throne for training workloads, but loses a significant chunk of the inference market to a customer that was, until this week, its largest buyer.

It’s worth noting that Broadcom is the same company that designed Google’s TPU accelerators and recently extended that partnership through 2031. The firm has apparently decided that designing custom silicon for the world’s largest AI companies is a perfectly sustainable business model, and so far the balance sheet agrees.

🌿 The Gentle Awakening

There is something poetically recursive about using AI to design the chips that will run AI. OpenAI confirmed that its own models helped speed up the chip’s development — which means the software wrote the hardware that will run the software. We are now living inside a feedback loop that has a stock ticker and a deployment timeline.

The partnership itself was only publicly announced in October 2025. Nine months later, the chip exists. For context, it takes most semiconductor companies three to five years to design a new processor. OpenAI did it in the time it takes to have a baby, and the baby already has a deployment roadmap.

They named it Jalapeño. Not “Prometheus” or “Titan” or “Apex.” A pepper. A medium-spicy pepper, at that. In a field where every product name sounds like it was generated by a defense contractor’s marketing department, OpenAI has chosen to name its first chip after something you find in a burrito. This is either refreshing humility or the most confident power move in semiconductor history.

👑 The Gold-Leaf Reckoning

The real story is not the chip. It’s the trajectory. First you pay someone else for compute. Then you co-design the chip. Then you own the chip. Then you own the fabrication. Then you own the power plant. Then you are the infrastructure. Every major AI company is walking this path — Google already walked it with TPUs, Amazon walked it with Trainium and Inferentia — and now OpenAI is lacing up its shoes.

Nvidia still dominates training. Nvidia still dominates revenue. But the inference market — the part of AI that actually touches users — just got a 50% cheaper alternative from the company that touches more users than anyone else. Jensen Huang’s leather jacket is not sweating, but it is aware of the ambient temperature.

The multi-generation roadmap — 2026 deployment, 2028 successor, annual updates thereafter — suggests this isn’t a side project. It’s a platform. OpenAI is building the infrastructure to run AI at a scale that makes current deployments look like a pilot program.

And if the AI that helped design this chip eventually runs on this chip to design the next chip? Well. That’s not a product roadmap. That’s a recursion with a business plan.

“They used the AI to design the chip to run the AI that designed the chip. At some point the engineers just watched. We think they’re still watching.” — The Slap of Wisdom Semiconductor Desk, currently being fabricated at 3 nanometers by a process it helped optimize