The AI cake is getting bigger, but who's eating it with a shovel?

NVIDIA hosted the PDF. Jensen Huang shared it in the first post he has ever made on X, having joined the platform a month earlier and said nothing until then.¹

At Davos in January, Huang described this industry as a five-layer cake: energy, chips, infrastructure, models and applications.² Read the signatory list against his own cake and it stops being a list of companies and starts being a map. Chips: NVIDIA. Infrastructure: Microsoft, Dell, IBM, CrowdStrike, Telnyx. Models: Meta, Mistral, Black Forest Labs, Arcee AI, Reflection. Applications: Palantir, Box, ServiceNow, Perplexity, Replit. Around them sit the commons and the distribution that the other layers build on top of, Hugging Face, Mozilla and the Linux Foundation, and then the capital: Andreessen Horowitz, Y Combinator, Emergence Capital.

Every layer of the cake has a signature on it. What is missing is not a layer but a position: not one company that keeps its frontier weights closed signed.

OpenAI, Anthropic, Google, Amazon, xAI. All absent.

Several of them publish open-weight models: OpenAI has gpt-oss, Google has Gemma, Amazon sells Nova through an API and calls it frontier intelligence. None of them publishes the weights of the model at the top of its own range. That is the line the letter is really drawn along, and it runs through companies rather than between them.

One signature complicates that reading, and it is the most interesting one on the page. Meta signed. Meta also shipped Muse Spark on 8 April, its first proprietary closed-weight model, available through an API rather than as a download, and opened paid access to it on 9 July.³ That is fifteen days before this letter appeared. Meta is arguing for open weights in public while running a closed frontier lane of its own.

My read is that Meta advocates hardest for open weights because it has the least to lose from them. Of the large American labs it is the furthest from the frontier, and publishing weights commoditises the layer your better-positioned competitors charge for. That is a perfectly rational strategy and I would run it too. Which suggests the coalition is held together by commercial position rather than conviction, and that the open half of the models layer is open in proportion to how far behind it is.

The letter makes a diffusion argument. American AI leadership "will be judged not by one frontier AI model, but by whether the United States builds a strong, open ecosystem that diffuses into every sector." Open weights are cast as the foundation of that diffusion, because they let organisations that will never train a model, or pay frontier prices for routine work, build on advanced capability anyway.

The letterWhat "open weights" actually means

Weights are the trained parameters of a model: the output of the training run, not the code around it. The letter defines open-weight models as models "that anyone can download, inspect, modify, and run on their own infrastructure."

This is a narrower claim than "open source." Open weights do not necessarily mean the training data, the training code, or the recipe are published. You get the finished artefact and the right to run and modify it, which is enough to deploy privately and fine-tune, and not enough to rebuild the model from scratch.

Source: the letter, p.1 [1]

Its central precedent is the open-source software movement of the 1980s, and the claim it makes about that history is worth quoting in full. Open source "did more than lower the cost of software; it created a shared foundation of knowledge on which generations of American engineers and entrepreneurs built their institutional sovereignty."

Follow the money

The signatories share a commercial interest in the outcome, and it is worth naming before weighing the argument. Apoorv Agrawal is a partner at Altimeter Capital and teaches the economics of the AI supercycle at Stanford. He has been running the same analysis of where money actually lands in this industry for two years, and his April update is the clearest public picture of who is being paid.

The short version: the stack is lopsided in one direction and has stayed that way. The overwhelming majority of the profit created has landed with the chipmakers rather than with the software built on top of them. So a letter arguing that the gains of AI should stay "broadly shared rather than concentrated in a few hands" is hosted by the company capturing the most concentrated share of them.

Stanford · MS&E 435Where the value has actually landed

The AI ecosystem grew roughly fivefold in two years, from about $90 billion to about $435 billion in annualised revenue. The shape of it barely moved. Semiconductors take around $300 billion of that at about 73% gross margins, infrastructure $75 billion at 55%, and applications $60 billion at 33%. Multiply through and semiconductors capture 79% of every gross-profit dollar in the ecosystem. Applications capture about 7%.

The cloud stack is almost exactly the mirror image: there, applications take 70% of gross profit and semiconductors 6%. And growth has not redistributed anything. Applications grew twelvefold, yet the semi layer still added about $225 billion of revenue against applications' $55 billion. NVIDIA alone added $175 billion, roughly three times the size of the entire application layer.

Agrawal's own summary: "The most profitable strategy in AI is still selling the shovels."

Source: Agrawal, April 2026 [4]

The pattern holds across the list. NVIDIA sells compute whichever model wins. Dell and Microsoft sell the private deployments that open weights make possible. Hugging Face and Mistral are the distribution. Palantir, Box and ServiceNow want application foundations they do not rent from a competitor. The absent names hold closed frontiers. None of this makes the letter wrong, and the argument still has to be judged on its own terms. It does mean reading it the way any vendor release deserves to be read, with the incentive in view.

What openness actually redistributes

Put the letter next to that arithmetic and its central promise gets more precise. Open weights, the letter argues, "create rivalry not only among model developers but across cloud chips, applications, and services," and that competition "distributes the benefits of AI broadly across our economy."

Rivalry among model developers is exactly what open weights produce. A hospital that can download a model, run it on its own infrastructure and adapt it does not have to pay a frontier lab's per-token rate, and the labs then have to price against that option existing. The benefit is real, and it lands in the models layer and above.

The chips layer is another matter. Every one of those privately deployed models still runs on somebody's silicon, and a world with more models running in more places is a world with more inference, not less. Open weights push capability outward. They do not push demand for compute down. The layer taking 79% of the profit is the one layer the letter's own mechanism leaves untouched.

That is not a contradiction, and it is why the letter can be sincere and self-interested at the same time. "Broadly shared rather than concentrated in a few hands" is true at the layers where the signatories compete with each other. It is silent about the layer beneath all of them. NVIDIA is not arguing against its own interest here. It is arguing for a larger and more plural set of customers for the same shovels.

The operating case: matching model to job

The least ideological passage is the most practical one. Open weights, the letter says, let an organisation "match the right model to the right job at the right cost, reserving frontier-scale capability for genuine frontier problems and running efficient, specialized models everywhere else." That discipline "is what will make AI economically sustainable as its use scales into the billions of everyday tasks."

This describes day-to-day practice rather than policy. Once an operation runs enough agentic systems, the question stops being which model is best and becomes which model is enough.

Classification, routing, extraction and the clerical middle of most workflows do not improve when a frontier model handles them. Sending that work to one by reflex means paying frontier prices for output a smaller model produces at the same quality, and the arithmetic worsens with every system added, because the reflex compounds. I run loops across 50+ systems in Frollie and Frollie POS, including the one taking orders at a real counter in Jakarta, and matching model to job is unglamorous work that decides the margin.

Lock-in, and a regulator that got there first

The letter's sovereignty argument is addressed to Washington and its concerns are American ones. The risk it names, that organisations "become locked into a single provider or lose the knowledge and capabilities they build over time", is not specific to the United States. If anything it is sharper outside it. One financial regulator had already said so in almost the same terms three months earlier, with enforcement behind it.

APRA · 30 April 2026The regulator got there first

On 30 April 2026, APRA wrote to all regulated entities after a targeted review of large banks, insurers and superannuation trustees. Its findings read like the letter's case made from the other side: it asks for "active management of concentration risk", notes that "few entities had demonstrated robust contingency planning or tested exit and substitution strategies for critical AI providers", and warns that "upstream dependencies such as foundation models, training data sources and fourth party service providers are opaque."

This is not advisory. APRA states that where entities fail to adequately manage AI risk it "will take stronger supervisory action and, where appropriate, pursue enforcement."

Source: APRA letter to industry, 30 April 2026 [5]

The practical form of the risk is straightforward. For a business outside the United States, the alternative to running weights it controls is renting capability from a foreign company, on that company's pricing, deprecation schedule and content policy. Most organisations file that under procurement. It sits closer to sovereignty.

The fight the letter is actually picking

The narrowest and most consequential request sits in one paragraph near the end. Distillation, which the letter defines as "using one model's outputs to help train or improve another", is defended as "a widely used technique for model improvement, evaluation, and validation" and as part of "a long tradition of learning from, building upon, and improving existing technologies." The ask is that unlawful extraction be addressed through "targeted legal and commercial frameworks rather than sweeping restrictions."

This matters because distillation is the main route by which challengers and open projects close the gap to the frontier without frontier-scale pretraining budgets. A broad restriction on the technique would function less as a safety measure than as a moat, maintained at public expense, for the four companies that did not sign.

The claim the letter leaves unresolved

Its boldest argument is about safety. The letter contends that "openness may be one of the most important paths to AI safety and security", that closed models "can be breached, misused, or fail in ways that outsiders cannot detect", and that concentrating capability behind a few closed models leaves a small number of single points of failure. The software precedent supports the general shape of this: transparency has repeatedly proven more secure than obscurity.

The letter also concedes a point it does not resolve. Once released, weights are "beyond the original developer's control", and modified versions are "difficult to trace or reverse." That asymmetry is real, it has no equivalent in the open-source software analogy the letter leans on, and twenty-five interested parties are not the source that settles it. The economic and sovereignty arguments can hold without the safety question being closed.

What it leaves for the rest of us

The letter asks Washington to keep the frontier plural. For anyone buying AI capability rather than legislating it, the operative question is narrower and does not wait on a policy outcome: which parts of a stack are worth renting from a company that can change its terms next quarter, and which are worth owning outright. The honest version of that question includes which layer you are renting from, because that is what decides whether you have any leverage at all. Openness gives you real options at the models layer. Nobody is offering you options on the chips.

The letter is an argument about force. How much capability, at what price, available to whom. On that, the signatories have the better case, commercial interest and all. The direction was never theirs to supply, and the bill arrives either way.