There’s a race underway right now to shape the future of AI. The leading American AI companies are trying to capture the market, spending money like it might rot. They’re buying chips, building data centers, securing the electrical connections necessary to power it all.
Bragging rights for the most powerful model keep changing hands, while these firms are lobbying lawmakers to help shut down competition with compliance costs. OpenAI and Anthropic are trying to get their models integrated into government and business workflows. OpenAI even offered Washington a five percent equity stake in the company.
Both are heading toward IPOs. Neither is profitable.
Meanwhile China is running a different kind of race. Its strategy is to weaken American dominance by giving the models away and serving inference for a fraction of the cost.
Chinese models typically cost 60 to 90 percent less than comparable systems from OpenAI and Anthropic. Kimi K3, a Chinese model that competes directly with the latest Claude Opus models, is available for a third of the price. DeepSeek is even cheaper than that. Testing by Chamath Palihapitiya’s firm has shown that the open-source options are 16.4 times cheaper than the American frontier models.
For most companies adopting AI, the calculus is becoming clear. They don’t need the most powerful model. They need reliable systems that can optimize a supply chain, predict a maintenance failure, and analyze data in real time. What matters is whether the model does the job, what it costs, and who controls it.
So while America has been racing to build the most expensive intelligence in history, China has been racing to make intelligence more useful.
This should have been a straightforward competitive story. Better product versus cheaper product. The free market can sort it out.
But it became something else.
How China Got Cheaper
In 2022, in an effort to stall China’s AI industry, the US enacted sweeping export controls on advanced AI chips.
The controls squeezed Chinese AI, but they didn’t cut it off entirely. To compete with Western AI, the Chinese firms had only one viable path: learn to do more with less. So they worked within the constraint they’d been given.
The optimization techniques they developed paid off. DeepSeek V4 Flash runs on about a tenth of the computing power required for a comparable Western AI model.
To hear Anthropic tell it, the Chinese got so competitive so fast because they copied Anthropic’s homework. That is, the Chinese models were trained on output from Anthropic’s models. And the Chinese didn’t have to pay for all the expensive training. Okay, maybe so, but that doesn’t explain the serving cost, the memory efficiency, the systems engineering. Those are all architectural decisions made under constraint.
The pricing is a bit more complicated. A dollar buys 1.15 million output tokens from DeepSeek’s V4 Pro. That same dollar buys 40,000 from Claude Opus 4.8. That’s a 28-to-1 gap. The advances in optimization helped make the price possible, but they probably don’t explain it by themselves. China has every reason to subsidize inference as part of a broader strategy to capture adoption and make its AI models impossible to refuse on cost. The efficiency is real. The price may also be a weapon.
Alibaba’s Qwen models have been downloaded over a billion times. Roughly 70 percent of all new open models that build on an existing architecture start from Qwen. Singapore’s national AI program abandoned Meta’s Llama and chose Qwen as its foundation. The ecosystem compounds. China is giving the models away to build a platform underneath them.
On July 17, 2026, Xi Jinping took the stage at the World AI Conference in Shanghai and called for international cooperation. He warned against “overstretching the national security concept in the field of AI” and announced a 29-nation open AI alliance.
While Beijing opened up, Washington started reaching for the off switch.
The Off Switch
Imagine being a company using closed American AI. You’re pouring your proprietary data into systems you rent from another company, under terms they can change at any time. The AI company owns the endpoint your bottom line depends on. It sets the price, and can retire the model or even be ordered by Washington to cut you off completely.
Alex Karp, the CEO of Palantir, sells software to some of the world’s most security-conscious institutions. He recently said at Davos, “I want something I own. I want to own the GPUs. I want to own my data. I want to own the model.”
Capable Chinese AI undercuts the American labs on price, while also weakening vendor lock-in. Some models like DeepSeek can integrate with existing OpenAI-compatible tools and workflows as a drop-in replacement. A small business can begin as an API customer, then bring the model in-house, running on their hardware. The company stops renting. It owns its AI.
There’s three options for American AI vendors now.
They can lower prices, but that would undermine their revenue story and tank their trillion-dollar valuations. They can release their weights, but that would give anyone with enough compute the ability to serve their models, damaging the one thing generating revenue for them. Or they can ask Washington to remove the competitor from the market…
On July 25, the New York Times reported, citing five people close to the discussions, that OpenAI and Anthropic had been privately lobbying Washington regulators to restrict Chinese open-weight models. The administration had spent weeks considering entity list designations, procurement bans, liability requirements, and security advisories. On July 22, Treasury Secretary Bessent went on television and threatened sanctions.
Meanwhile Anthropic sent a letter to the Senate accusing a Chinese competitor of distillation attacks and updated its terms of service to block companies with majority Chinese ownership. Every move pointed in the same direction: make Chinese models too risky for American companies to touch.
On the day Moonshot released the Kimi K3 weights, Anthropic’s CEO published a blog post conceding that ordinary open-weight models are a public good, that banning American companies from using Chinese models wouldn’t address the real threats, and that such restrictions could amount to protecting American companies from competition. Then he proposed mandatory testing for all “sufficiently capable” models. The category argument was lost. The fallback was a licensing regime where the companies already at the frontier help define what counts as sufficiently capable.
The rest of the industry pushed back. Twenty-five major tech companies signed an open letter opposing the restrictions. Anthropic was the only major American AI company that refused. Trump’s AI adviser David Sacks called it what it was: “The weaponization of regulatory uncertainty as a competitive tool.”
Two weeks later, the White House delivered its answer in a closed-door meeting with OpenAI, Anthropic, and Google. The framework covers closed-source frontier models only. Open weights are explicitly exempt.
So now Kimi K3, which scores above Claude Opus 4.8 in benchmarks, is exempt from Washington’s controls. If you can download it, it’s exempt.
And that framework is “voluntary.” The White House isn’t publishing it. The thresholds are classified.
The Irony
So all the lobbying the big AI labs have been doing appears to have hurt them. At least for now. They helped create this regulatory ambiguity and the burden now looks like it will fall on their customers.
A small business will see this development and not bat an eye. They’ll use whatever AI is best for their budget.
But a large Fortune 500 company with compliance departments, government contracts, and boards that answer to shareholders will see this very differently. They’ll look at “voluntary framework, classified thresholds, subject to revision without public notice” and see the safe move. They’ll stick with the American vendors. The ambiguity does the work.
The result operates like a tariff. It raises the cost of AI for American enterprises while collecting nothing from foreign competitors.
A manufacturer in Monterrey using DeepSeek is paying a fraction of the price the Americans are obliged to pay. And that difference ends up being a production-cost advantage.
Someone who wants to open a new factory in Detroit now does the math and sees the advantage of building it somewhere else.
And the supply of those open models is untouchable. The weights are on Hugging Face, on ModelScope, on mirrors and torrents and direct downloads. Washington can make it risky for Americans to touch them. Enforcement thins out past the G7.
And if Washington tightens further, it pushes the countries outside the American ecosystem toward the only alternative left. Huawei hardware running Chinese software. The restriction drives customers toward the Chinese stack, supplying it with exactly the users, workloads, revenue, and developer attention it needs to mature.
Protecting OpenAI means taxing Detroit.
The Lever
Washington restricted China’s access to chips. The restrictions were supposed to slow Chinese AI down. Instead, they helped create an industry organized around producing more intelligence from less hardware.
Washington considered restricting American access to these cheaper models and backed off. The framework covers closed models only. Open weights are exempt.
The lever is still installed. The framework is voluntary, unpublished, with classified thresholds and undefined terms. The exemption can be revised without public debate on any given Tuesday.
The weights are already in the world. The things Americans would have built with them aren’t.
Enjoyed this piece?
I do all this writing for free. If you found it helpful, thought-provoking, or just want to toss a coin to your internet philosopher, consider clicking the button below and donating $1 to support my work.
The buyers you describe, choosing on reliability, price and control rather than raw capability, are hitting a fourth variable: who's liable when it breaks. RAND's review of US insurance filings this year found most carriers silent on AI losses, exclusions spreading into directors-and-officers coverage, and no standalone AI policies filed in the regulated market. Whether the stack runs on Opus or DeepSeek, the liability for a bad output lands on the deployer, and no one upstream wants to underwrite it.
That's a separate race from the one you're tracing, cost per token against compute spent on frontier capability. It's the race to get indemnified, and nobody's winning it yet. Does the price gap you describe make US buyers more willing to self-insure, or just slower to deploy at all?
China wins this round. Capitalism shot itself in the foot.