OpenAI designed a chip to own its compute. It still rents everything that makes the chip run.

The story arrived pre-narrated, which is usually the first sign to slow down. On Wednesday, OpenAI and Broadcom unveiled a custom chip called Jalapeño, and within hours the verbs were everywhere. OpenAI "distances itself" from Nvidia. It "strikes" at the leader. It moves to "build the full stack." Every one of those framings tells the same flattering story: a company seizing control of the thing it most depends on. I want to read the announcement the other way, because the useful question about any piece of infrastructure is never what someone built. It is who can still turn it off.

Start with what was actually announced, stripped of the verbs. Jalapeño is an inference chip — a blank-slate design for the compute-intensive work of serving OpenAI's models to the people typing into ChatGPT and Codex and the API. It is not a general-purpose processor and it is not for training the next model; it is built for one job, done at enormous scale. OpenAI says it designed the chip from scratch around its own understanding of how large language models run, with Broadcom and the manufacturer Celestica, and that it used its own models to accelerate parts of the design. The company claims it went from initial design to manufacturing tape-out in nine months, which would be among the fastest cycles ever for an advanced chip. Early testing, OpenAI says, shows roughly 50% cost savings against typical AI graphics processors and substantially better performance per watt.

The framing was set on the stage. Broadcom's chief executive, Hock Tan, and its president, Charlie Kawwas, physically handed the first chip to OpenAI's Sam Altman and Greg Brockman. Tan called the collaboration "a fundamental commitment to scaling the physical infrastructure required for the next decade of AI" and "just the beginning of a multi-generation roadmap." Jalapeño is the first step toward a stated goal of 10 gigawatts of OpenAI compute by 2029, with gigawatt-scale data centers built alongside Microsoft and other partners starting this year. Tan, separately, described the demand from Broadcom's handful of large customers as "simply insatiable." Hold onto that word.

What OpenAI actually owns now

A design. That is the honest noun for it. OpenAI now owns an architecture — a set of decisions about how to lay out silicon for serving its own models, an instruction set, a blueprint refined against its real workloads. That is genuinely valuable, and genuinely hard, and the company deserves the credit for doing it fast. But a design is a drawing. Drawings do not compute. To turn Jalapeño into a running chip inside a powered building, OpenAI needs a long list of things it does not own.

• The foundry. OpenAI has no fab. Jalapeño is etched on someone else's leading-edge line — the same scarce nodes every advanced-chip customer on earth is queuing for.
• The partner of record. Broadcom designs and builds it, and Broadcom's own chief executive just told you its other customers — the ones whose demand is "insatiable" — are in the same line.
• The systems integrator. Celestica assembles the machine around the chip.
• The capital and the power. "Gigawatt-scale data centers with Microsoft and other partners." The electricity, the land, and the balance sheet wear someone else's name.

So read the dependency map again, before and after. Before Wednesday, OpenAI rented its compute mostly from one landlord, Nvidia, and resented the rent — the margin Nvidia takes on every GPU is the most-quoted grievance in the industry. After Wednesday, OpenAI co-owns a blueprint and rents the fab, the manufacturer, the integrator, the capital, and the power from a slightly larger set of landlords. That is diversification. It is a shrewd way to negotiate down the Nvidia tax. It is not independence, and the gap between those two words is the whole column.

Notice, too, what Jalapeño is for. Inference — serving the model — is the half of the compute bill that is cheapest to optimize and most repetitive, which is exactly why it is the right place to start with custom silicon. Training the next frontier model, the work that actually decides who leads, still rides on accelerators OpenAI buys from others. The company has carved off the optimizable half of its compute and left the frontier-defining half where it was. "Build the full stack" is a sentence about ambition. The stack on display is a floor of it.

A design is a drawing, and drawings do not compute. OpenAI now owns the blueprint and rents the fab, the factory, the integrator, the capital, and the power. That is a better lease, not a deed.

The counter-argument that nearly changed my mind

Let me make the strongest version of the case for reading this the way I was told to, because it is a serious case. Vertical integration here is not vanity. It is the single most rational move a company operating inference at OpenAI's scale can make. Google proved it years ago with the TPU: design your own chip for your own workload and you win a durable advantage in cost and control that renting can't match. Amazon followed with Trainium, Meta with its MTIA line, for the same reason. A 50% saving on the largest line item in the business is not a rounding error — at OpenAI's volumes it can be the difference between a unit economic that works and one that doesn't. And owning a design buys leverage even over the suppliers you keep: the credible threat to build your own is how you get a better price from everyone you still buy from.

I will go further and admit a place I have been wrong. I used to argue that the model was the moat — that the company with the best system would own the category. Models leak, get distilled, get cloned; I have since said in this column that the compute is the asset, not the model. A custom inference chip is OpenAI agreeing with that correction in silicon. So I take the bullish reading seriously. It is not foolish. It is, on its own terms, correct.

And here is where it stops

The binding constraint on artificial intelligence was never the cleverness of a chip design. It is megawatts and money. Read OpenAI's own headline number again: 10 gigawatts by 2029. Read the partner clause again: "with Microsoft and other partners." Those two phrases name the real owners of OpenAI's compute, and neither of them is OpenAI. You can design the most elegant inference chip on the planet and you still cannot energize it without a power-purchase agreement you do not control and a capital stack you cannot fund alone. The chip is the visible asset — the thing you can photograph and hand across a stage. The power contract and the balance sheet are the invisible ones, and invisible is where ownership actually lives.

Two weeks ago in this space I argued the opposite-shaped version of this story: that Apple, by renting its flagship intelligence from Google, had leased away the one capability a device company is supposed to own. OpenAI is the more interesting case because it is trying to do the reverse — to own the substrate instead of renting the brain. But it can only own the part of the substrate that is intellectual property: the design. The physical substrate — the fab, the factory, the grid — belongs to others, and there are startlingly few others. Apple rents the intelligence; OpenAI owns a chip it cannot manufacture, power, or finance by itself. Both companies are arriving, from opposite directions, at the same wall: in this industry you can own the idea or you can own the iron, and almost no one owns both.

This is the part the keynote was built to keep you from asking. Suppose it is 2029, and the people who own the fabs and the gigawatts decide OpenAI is no longer their best customer — a rival bids higher, a government leans on the foundry, a partner's own insatiable demand comes first. What does owning the Jalapeño design get you then? A drawing, and a place in the queue. Custom silicon is how you renegotiate the rent. It is not how you stop having a landlord.

The companies spending the most to look independent are telling you, with their balance sheets rather than their slogans, that they know exactly who they still answer to. The compute is still the company. OpenAI just commissioned a more flattering map of who holds it.