Here's what's on our plate today:

• 🧪 The $500M nobody licensed: how Chinese labs legally buy American expert judgment.

• 📰 Beijing rejects picking AI sides, Stripe buys OpenRouter for $7.5B, Apple's camera AirPods stay in 2027.

• 🎯 Weekend To-Do: audit your vendors, read the Forbes report, benchmark Kimi K3.

• 📊 Poll: what should Washington do about expert training data?

Let’s dive in. No floaties needed.

The Laboratory

TL;DR

Washington guarded the chips for four years. Nobody thought to guard the judgment.

• The other scarce input: frontier models need compute and expert human judgment. Export controls cover the first and say nothing about the second.

• A $500M open door: the six largest Chinese labs buy roughly $500M a year in training data from the same vendors serving OpenAI and Anthropic.

• Not a few bad actors: thin margins and a short customer list make resold datasets the most profitable thing these firms sell. Shaming will not stop that; only a rule will.

• Beijing moved first: Chinese regulators are already drafting limits on training data leaving the country, a conclusion Washington has not reached.

• The stakes: closing the gap means deciding which professional expertise a citizen may sell abroad, a line no democracy has drawn. Until someone draws it, the American lead narrows at the speed of a purchase order.

Why Silicon Valley is selling to both sides of the AI race

Stopping a rival country from building something advanced usually comes down to one question: which part of the process is hardest for them to replace on their own. Every export control is a bet on the answer, and the bet only pays off if the thing stopped at the border is the thing nobody can build around. When it is, a single restriction can slow an entire industry for years. When it is not, the control still gets enforced beautifully, while the capability it was meant to contain arrives through a door nobody thought to watch.

For four years, Washington has placed that bet on a single component. Advanced chips, the specialized processors that perform the arithmetic behind large AI models, were added to a restricted list, and the rules around them have grown longer and more detailed every year since. Officials at the Commerce Department spend their days drawing and redrawing performance thresholds, deciding at what point a processor becomes too fast to sell abroad, and the licensing rules now reach the overseas subsidiaries of Chinese firms. And even as those restrictions tightened, Congress spent this year arguing the line still sits in the wrong place, with the House Foreign Affairs Committee advancing a slate of bills to tighten it further.

Four years of that argument have made it easy to lose sight of the assumption the bet depends on. Both sides of the chip fight treat compute as the scarce ingredient in a frontier model, so controlling advanced chips is taken to mean controlling the ability to build advanced AI, which accounts for only half of what a model actually needs.

The other AI bottleneck

The other half is expert human judgment, which is every bit as scarce and much harder to see. Someone has to teach a model how a lawyer analyzes a contract or how an engineer finds a bug, turning that expertise into examples and grading rules the model can learn from. Unlike chips, this knowledge is not sitting in a factory that Washington can put behind an export license, because it has to be created by people, and those people can be paid for their work by anyone.

That is where the four-year bet begins to come apart, because the companies producing this expert training data are largely American, while their customers do not have to be. A Chinese AI lab with enough money can buy much of the same expertise that American labs are buying. The chip controls may limit how much computing power that lab can use, without limiting how much knowledge it can acquire. And if the same suppliers are selling their expertise to both sides, every improvement an American lab pays to develop can become something the Chinese side can buy. Which means that the question Washington has not fully confronted is whether chips are still the part of the AI supply chain that matters most.

The clearest answer to this conundrum came in August, when sales teams from America's biggest data-labeling companies traveled to Seoul for a machine learning conference. They expected to meet their usual customers: demanding American AI labs. What they found instead were Chinese labs waiting with their own shopping lists.

Tencent, for example, had prepared a request for training data covering finance, cybersecurity, and self-improving AI systems. Forbes reported on August 5 that the six largest Chinese AI labs spend roughly $500M a year with American data-labeling companies, a figure two data-labeling entrepreneurs gave Forbes based on market-size estimates from Tencent and ByteDance executives.

The individual deals look less like a scandal and more like ordinary business expansion. Mercor gets about 2% of its second-quarter revenue from Chinese labs and has hired an engagement manager for the region. AfterQuery gets at least $50M in recurring revenue from Chinese customers, while Surge AI CEO Edwin Chen has traveled to China to meet lab executives. Mercor declined to comment, while AfterQuery and Surge said they do not disclose customers. But none denied doing business with Chinese labs, because nothing in the current rules prevents them from doing so.

What the buyers are really getting

The sum itself is small by the standards of an industry that counts in tens of billions, and what matters is what it contains. Teaching a model to handle complex accounting is not mostly about collecting finished spreadsheets. The valuable part is the instruction manual, which outlines which problems to focus on and the grading rules, written by working professionals, that tell the model when it has the answer right. Those rules come out of long and expensive arguments inside American labs about what good work looks like. Vendors build the machinery to produce them on custom projects for OpenAI or Anthropic, then point the same machinery at a standard package sold to several buyers at once. A Chinese lab that buys one skips months of trial and error, inheriting reasoning somebody else paid for.

None of this is what an export control was ever built to stop. A restricted chip is a physical object that crosses a border, can be counted, and can be switched off remotely. A dataset is a file, and the expertise inside it belongs to the people who wrote it rather than to any country. These vendors hold what TechCrunch has called some of the model makers' biggest trade secrets, and they owe Washington nothing.

Serving both sides at once

The vendors are by no means neutral, since they sell to the American government too. Surge AI has listed the U.S. Army and the U.S. Air Force among its customers, and Mercor told its staff in late July that it had won a U.S. federal contract, days before the same reporting placed it inside Tencent's supply chain. A company can hold a defense contract and a Beijing purchase order in the same quarter and break no law doing it, because procurement rules and export rules were both written to inspect physical things.

However, not every company operates on this assumption. One company drew the line anyway, when Scale AI walked away from a ByteDance deal in 2024 on national security grounds and watched its rivals keep selling, which is what restraint costs when nobody else has to match it. The people defending the trade still have a real argument that goes beyond self-interest. Data consultant Sean Cai told Forbes that restricting these exports would damage American open-source work, since the only labs able to survive such a rule are the two or three rich enough to build their own expert networks. Others in the industry call the trade shameful, and neither side has persuaded the other. While they argue, the selling continues for reasons unrelated to either of them.

Why the selling continues

The simplest explanation for why these companies continue to sell to China is financial, and it shows up plainly in the accounts. Those reasons are commercial, and they matter because they decide whether the trade can correct itself. A company with room to turn down awkward revenue will sometimes do so, and a company without that room will not, however loudly its competitors object. The accounts of the firms doing the selling suggest that there is very little room for that.

Mercor made $614M in gross revenue in the first half of 2026, with roughly 91% of it coming from a handful of foundation model companies. Gross is the word that matters, since most of that money goes straight back out to the contractors who do the work, leaving a thin margin on a very short customer list. What a short list costs became clear in March, when a single security breach led Meta to pause its contracts entirely.

A business shaped like that needs a product it can sell more than once, and the off-the-shelf dataset is exactly that, built for one lab and then resold at the best margin these companies make. The most eager buyers are Chinese labs, which cannot get this material anywhere else. The pressure is not going to ease either, since the money raised assumes it will not, and Mercor is reportedly raising at around $20B while Surge has been in talks at $25B or more.

So selling into China is not a lapse of judgment by a few firms that could be shamed into stopping. It is holding up part of the business, which means the trade will outlast every argument the industry has with itself, and the only thing that ends it is a rule.

Beijing reaches the same conclusion

That puts the question back on governments, and only one of them appears to be asking it. Washington has not yet behaved as though it knows the trade exists. Beijing, coming at the same problem from the other side, has already decided that this kind of data is worth guarding.

Chinese regulators have been meeting with the country's largest AI companies about restricting foreign access to their best models, and the proposals reportedly go as far as limiting training data leaving China. The part worth noticing is not the restriction itself but who is imposing it. A government that spent years objecting to American export controls has reviewed its own supply chain and concluded that expert training data should be placed on the controlled list.

That is the clearest outside sign that this layer carries real strategic weight, since the government that changed its mind about which part of the supply chain matters is not the one that built the modern export control.

Wanting to act and knowing how to act are different problems, though, and the second is harder than it looks. The material in dispute is made by lawyers, doctors, and engineers working as contractors in dozens of countries, and its value comes from how their reasoning was organized rather than from where a file sits. A rule covering it would have to specify which kinds of professional expertise a person is allowed to sell abroad, a line no democracy has ever had to draw.

That is a question about citizens rather than cargo, and it is part of why the gap stayed open while everyone argued about chips. Whatever answer emerges will have to arrive soon, since Moonshot's Kimi K3 already beats models just below the best from Anthropic and OpenAI on coding tests.

The chip controls are real, expensive, and seriously enforced, and they may still turn out to be the right thing to build. They rest on a bet about which ingredient was hardest to replace, and for the past year, the market has quietly tested that bet by selling the other one to anyone who asked. Washington can widen the controls to cover expertise itself, accept that the lead narrows at the speed of a purchase order, or keep finding out after the fact which ingredient mattered. What nobody has settled is whether a country can control the movement of judgment without also controlling the people who produce it.

Weekend To-Do

• Audit who your vendors also sell to: List every AI data, eval, or labeling vendor in your stack and check their customer disclosures; the ones that name nobody are telling you something.

• Read the Forbes report behind this story: Anna Tong's August 5 piece has the $500M figure, the Seoul conference scene, and the vendor-by-vendor numbers in full.

• Benchmark Kimi K3 against your own task: Run a coding problem you already know the answer to; the gap you find is the gap the chip controls were built to hold.

Friday Poll

Headlines You Actually Need

• China pushes back on picking sides: Beijing urged respect for each country's digital sovereignty after a draft reviewed by Reuters showed Washington preparing to tell dozens of countries to choose an AI camp.

• Stripe's $7.5B bet on the router: Stripe bought OpenRouter for a reported $7.5B, moving from collecting payments into managing what developers spend on AI tokens.

• Camera AirPods stay parked in 2027: Apple left a Visual Intelligence demo video inside a macOS release candidate, but supply chain and software work keeps the earbuds off this year's schedule.

Meme Of The Day

The Toolkit

• Dust: No-code platform for building AI agents that connect to your company's tools and automate real workflows.

• Descript: Audio and video editor that lets you cut recordings by editing the transcript like a doc.

• Chroma: Open-source vector database for AI apps, quick to set up and easy to scale for RAG.

Rate This Edition