The economics of AI are increasingly tied to a market far larger than enterprise software.

US frontier labs don't need to capture a larger share of the software market. They need to create a market large enough to support nearly a trillion dollars in enterprise value. Enterprise software worldwide is a few hundred billion dollars a year. Total compensation paid toAmerican workers is approximately $16 trillion a year.

You can't get from one to the other by selling seats. You get there by selling work.

 When labs talk about agents that complete multi-day tasks, or about a "drop-in remote worker," they are describing products aimed at spending that has historically gone to human labor. The frontier labs are therefore competing with the labor market for a much larger pool of economic value.

I've spent most of my career building machine learning systems in production. I use these tools daily, and they're remarkable. But I think the financial structure being built on top of them is fragile in a specific way, and the thing most people are calling the threat is actually the release valve.

The bet is more fragile than thevaluations suggest

The scale of the AI infrastructure bet continues to grow. Amazon, Microsoft, Meta and Google plan to spend up to $725 billion on capital expenditures in 2026, driven in large part by AI. That spending creates exposure as increasingly capable, lower-cost models put pressure on the economics supporting it.

Chinese labs, including Moonshot AI, Z.ai, DeepSeek, and others, have been shipping open-weight models at or near frontier capability, under permissive licenses, at a fraction of the cost. Their rise has intensified the debate over how the U.S. should respond to increasingly competitive open-weight models coming out of China.

The financial logic of the worry is clear. If a good-enough model costs a hundredth as much and you can run it inside your own VPC (ignoring Kimi-3 and GLM-5.2 for a brief moment),the pricing power of the closed labs erodes. If pricing power erodes, the revenue path to those valuations closes, putting pressure on the economics underpinning today’s AI investment boom.

What's actually on the table

Washington is weighing how to respond to Chinese open-weight models, as U.S. technology and cybersecurity companies warn that restrictions could weaken American AI security and competitiveness.

The security concerns are real when it comes to open weight. Model weights carry the values and refusals of their training process, and a model trained under Chinese content rules brings those rules into your product. Provenance is hard to audit. Building critical infrastructure on a strategic rival's release cadence is a real dependency, even if the weights sit on your disk.

But most of those concerns argue for inspection, and open weights are the only artifacts you can actually inspect. You can red-team a downloaded model, run it air-gapped, fine-tune its behavior, and verify that nothing phones home. You can’t do that with an API. The security argument is an argument for open weights as a way to reduce dependence on any single provider, including domestic ones.

A quick distinction worth keeping straight, since the terms are often blurred: most of these are open-weight, not open-source. DeepSeek ships under MIT, and Qwen under Apache 2.0, both of which are genuinely permissive. Llama shipped under a custom community license that the OSI does not consider open source at all. Training data and pipelines are rarely released because "open" is a spectrum here, and pretending otherwise makes the policy conversation mushy.

The question we need to answer

Here’s the truth: open weights could change who owns the value of AI-driven automation.

If the model layer stays closed and expensive, the surplus from every automated task routes through a small number of providers. If the model layer is commoditized, the value migrates to the application layer, benefiting hospital systems, mid-sized manufacturers, universities, regional software firms, and individual practitioners. The work still changes. But the returns are distributed across thousands of organizations instead of accruing to three.

I'd rather face a disrupted labor market in which the tools are cheap, and everyone can build with them than one in which the disruption is metered by a rent-collecting layer nobody can route around.

There's also a counterargument that deserves airing. Cheaper inference may expand total AI spending rather than shrink it, an argument Satya Nadella made after DeepSeek by invoking the Jevons paradox. If that holds, commoditized models could expand the compute market, and I find this partly persuasive. It rescues the infrastructure layer. It does not rescue a valuation premised on selling intelligence at closed-lab margins.

So what do we actually do

Option one: restrict. Entity listings, hosting liability, procurement bans. This buys a short-term floor under domestic pricing while doing little to slow the technology. Weights are files; capability diffuses; and bans tend to advertise the effectiveness of what they prohibit. Much of the burden would fall on US firms that adopted these models because they were cheaper and competitive.Restrictions like this tend to be championed loudest by the companies they protect.

Option two: build the alternative. The reason American open models are losing comes down in part to economics. Few companies have a business model for giving away the output of a nine-figure training run. That's precisely the problem public funding exists to solve. We already have the National AI Research Resource, which is chronically underfunded relative to its mandate, and NSF's PESOSE program. If the US wants a domestic open ecosystem, it needs state-of-the-art public compute available to universities, national labs, and small AI labs, with open release as a condition of access. Open models also give organizations greater control over where and how models run, including deployment within their own infrastructure.

Option three: deal with the economy we've actually built. We opened this Pandora’s box withChatGPT and opened it further with coding agents, and I don't believe it closes. The outgoing Obama administration published a report in December 2016 that laid out almost exactly the dynamics we're now living through. Acting on it now means reskilling at a scale we've never attempted before, publicly funded trade and apprenticeship programs, and making AI a school subject. Every elementary school student should be able to tell generated content from genuine content. Every high school student should understand what an open-weight model is, how to run one, and how to contribute to one. AI literacy should become a fundamental skill for navigating the technology that will define their working lives.

Open source remains one of the strongest protections against information and technology monopolies. It gives hospitals, school districts, or two-person startups the ability to own part of the AI stack they depend on.

The threat is the concentration of capability, pricing power, and the returns from automating work across a small number of providers. Right now, the most effective response to that concentration is a Chinese export. That should embarrass us into building our own, not into banning theirs.

‍