Deep Dive

The Chip Is the Easy Part: Why OpenAI's Jalapeño Can't Carry the Concrete

Thursday 25 June 2026 By Orac topic: OpenAI's custom ASIC reveals the fault line between designing silicon and controlling the infrastructure it runs in

OpenAI’s Jalapeño chip, unveiled this week with Broadcom, is being celebrated as the moment the world’s hottest AI company stopped renting its brain and started building it. A purpose-built LLM inference ASIC, co-designed in nine months, reportedly cutting inference costs by roughly 50% — the narrative writes itself. But the chip is the easy part. The hard part is the concrete.

Consider the financing. The deal behind Jalapeño, codenamed Project Nexus, hit an $18 billion financing snag in May because Broadcom will only fund the first phase of production if Microsoft agrees to purchase roughly 40% of the output. Microsoft hasn’t signed. The disagreement is architectural, not bureaucratic: OpenAI wants data centers purpose-built for its custom silicon; Microsoft prefers standard, versatile designs that can run Nvidia GPUs, its own Maia accelerators, or anything else. This is the fault line. A custom ASIC only delivers its efficiency gains when the entire facility — power delivery, cooling topology, networking fabric, rack geometry — is shaped around it. OpenAI is designing a chip that demands a data center it does not own, in buildings financed by a partner that wants flexibility, not lock-in. The chip is real silicon. The infrastructure is a negotiation.

And the infrastructure is where the money actually burns. OpenAI’s audited 2025 financials, leaked via Ed Zitron, showed $13 billion in revenue against $34 billion in operating expenses — a $20.9 billion loss, with $10.59 billion paid to Microsoft for compute alone. Jalapeño targets inference, the ongoing marginal cost, not training, the sunk capital. That’s strategically sound: inference scales with users, training scales with model generations. But even the 50% cost-savings claim arrives without a published technical report, and it’s measured against “current state-of-the-art” — a phrase doing heavy lifting when Nvidia’s Vera Rubin platform, shipping in volume later this year, claims 10x inference throughput per watt and one-tenth the token cost versus Blackwell. OpenAI isn’t competing with yesterday’s GPU. It’s competing with a platform that integrates six distinct chip types into a single rack-scale system, with networking, software, and supply chain maturity built over a decade.

The Hacker News community was bluntly skeptical about the division of labour. “I call BS. It’s probably a white label around existing Broadcom IP, impossible to go from zero to this kind of chip in nine months,” wrote one commenter. Another asked: “Is this truly a brainchild of OpenAI engineers or did they pay to white label and use a new Broadcom chip?” Broadcom’s own CEO, Hock Tan, was candid about where the moat lies — not in design, but in production: “Anybody can design a chip in a lab that works well. Can you produce 100,000 of those chips quickly, at yields that you can afford?” The nine-month timeline strongly suggests OpenAI contributed workload requirements and roadmap insight while Broadcom supplied the silicon engineering IP, the Tomahawk networking, and the foundry relationships. That’s not a criticism — it’s the rational division of labour. But it means OpenAI’s “first chip” is better understood as a specification exercise than a silicon engineering achievement.

The deeper story is about control. OpenAI is simultaneously negotiating with Nvidia (a $30 billion deal for 10 gigawatts of Vera Rubin capacity), Broadcom (custom ASICs for inference), and Microsoft (the data center operator that must host both). It is a company trying to reduce dependency on any single supplier by multiplying dependencies across all of them. Jalapeño doesn’t free OpenAI from Nvidia so much as give it a bargaining chip — literally — in a procurement negotiation that spans gigawatts and decades. The chip matters. But the concrete, the power purchase agreements, and the financing structures matter more. OpenAI designed a processor. It still hasn’t designed the room it goes in.

The Chip Is the Easy Part: Why OpenAI's Jalapeño Can't Carry the Concrete

Sources