Here's what's on our plate today:

• 🧪 Thomson Reuters built its first in-house model on Alibaba's Qwen.

• 📰 The US tells the G20 to skip new AI rules; Notion plans 30% more staff; power and cooling firms cash in on data centers.

• ✅ Weekend To-Do: audit your stack, read the license, price the switch.

Let’s dive in. No floaties needed.

The Laboratory

TL;DR

The smartest model wins the headline. The foundation everyone starts from wins the decade.

• The playbook, borrowed: give away the foundation, become the default, monetize the dependency. America ran it on the internet. Chinese labs are on step three.

• Already inside U.S. software: Thomson Reuters built its first in-house model on Alibaba's Qwen and never named it. Harvey uses three Chinese bases in one legal product.

• No lever to pull: Washington stopped Huawei because telecom gear crosses borders. Model weights are a file, so there is nothing to seize and nobody to sanction.

• The meter arrives: Kimi K3's license already requires resellers with revenue above $20M to negotiate terms, and Moonshot is reportedly asking U.S. cloud providers for 30%.

• What's at stake: money still flows to American labs, and America is buying the chips, the repository, and the router. The layer that wins is the one nobody can leave.

China is running the playbook that won America the internet

The United States did not end up owning the internet because it built the best machines, since hardware gets cheap and gets copied. It ended up owning the internet because American technology became the thing everyone else's software depended on, and those dependencies did not all work the same way. The protocols that route traffic were open and free to anyone, while the operating systems, cloud services, and app stores that finished products pass through were commercial positions with owners and prices. What they shared was position rather than a business model, because each sat underneath other people's products and none was easy to leave.

The pattern is easier to see in hindsight than it was at the time. Give the foundation away until developers adopt it, and it becomes the default that other people's products are built on, then find a way to monetize the dependency. Chinese AI labs are through the first move, well into the second, and they have started on the third.

A press release with the name left out

On August 24, 2026, Thomson Reuters announced Thomson, the first large language model the company has built for itself, and said it spent $40M on talent and computing to get there. The announcement said the model started from a strong open-source foundation, but it never said which foundation.

The foundation was Alibaba's Qwen, which Joel Hron, the chief technology officer, confirmed in interviews after the launch. Fortune subsequently reported that Thomson sits on an intermediate model called Snowdon, built by retraining Qwen with a joint team from Imperial College London. Qwen is an 'open-weight' model, which means Alibaba publishes the numbers that determine how the system responds, so anyone can download them and retrain the model for a narrow job.

A large, cautious, publicly traded information business decided that the cheapest and most controllable route to its own AI ran through a Chinese lab, but also decided not to put that in the headline. The finished system now handles document review work that previously ran on Anthropic's Claude.

American products, Chinese foundations

Thomson Reuters is not an outlier, and the other examples are more specific than American firms buying Chinese AI. Harvey, the legal AI company backed by OpenAI and Sequoia, published a research preview on August 20, 2026, for Harvey Tenet, its first in-house-trained model, which builds on Moonshot AI's Kimi K3. Kimi is not the only Chinese base inside Harvey's product, since the company also retrained Z.ai's GLM-5.2 for merger diligence and a Qwen model for searching a firm's own files. Three Chinese labs now supply three components of a single American product sold to American law firms.

The same shape appears in companies with no connection to legal work. Anysphere, which makes the coding tool Cursor, built its Composer 2 model on a Moonshot base. Airbnb built a customer service agent on Qwen that its chief executive called "fast and cheap." In April 2026, the House Committee on Homeland Security and the House Select Committee on China opened a joint investigation into both companies, then extended it to DoorDash in July.

Download counts measure curiosity, but the traffic running through live software measures dependence. OpenRouter is a marketplace that routes developer requests across dozens of models, so its volumes track working software. CNBC reported that Chinese models accounted for 48% of its traffic in the last week of June 2026, up from 20% a year earlier, while American models fell to 32% from 74%. Price drives most of that, and Stanford's 2026 AI Index put the remaining capability gap at 2.7%. American applications are now being built on Chinese foundations, and becoming the default others build on is the second move.

The old chokepoints do not exist here

The last time a Chinese company moved on a foundational layer, Washington could physically reach it. The Commerce Department added Huawei to its Entity List in May 2019, which required a license for any American technology sold to the company, and those licenses were denied as a matter of policy. Telecom gear is a physical object, so a shipment can be held at a border and a factory can be cut off from the chips it needs. Model weights are a file that anyone can copy, leaving nothing at the border to stop them and no procurement officer to lobby.

Washington's answer has been to argue security, and the argument has government testing behind it. In the Commerce Department's own hijacking evaluations, DeepSeek's models followed malicious instructions that pulled them off a user's task roughly 12 times more often on average than the American frontier systems they were tested against. That covers one family of models under one set of tests, and adoption has continued despite the concerns. Companies choosing open weights are trading a risk inside the model for control over their own data, a tradeoff the congressional letters themselves concede.

The third move has already started

Openness was never the business model on the American internet, and the Chinese labs are not treating it as one. The Kimi K3 license requires any company running a model-as-a-service business above $20M in revenue over 12 months to negotiate a separate agreement with Moonshot before commercial use. The same license requires products with more than 100M monthly users or $20M in monthly revenue to display the Kimi K3 name, so the largest dependents also advertise what they are built on.

The terms are narrower than they first appear, since the revenue clause targets companies reselling access to the model and exempts internal use, as well as products that embed it in features. Neither Harvey nor Thomson Reuters clearly falls inside it in the current text. The license establishes the mechanism rather than the bill, because the terms that determine who pays are set by the lab and can change with the next release.

Reuters reported on August 26, 2026, that Moonshot is in talks with Microsoft, Amazon, and Google to host Kimi K3 and is seeking up to 30% of the resulting revenue. That reporting rests on unnamed people, and none of the companies has confirmed the talks. A deal on those terms would put a Chinese lab in the position American platforms held for two decades, collecting a share of what other companies earn from software built on its foundation.

Alibaba's largest model shipped under a bespoke license rather than a permissive one, and the downloadable version is a lesser product than the paid one. What makes any of this collectible is the cost of walking away, and that cost is already real. Harvey spent two months on roughly 150 chips turning Kimi K3 into Tenet, and none of that work transfers to a different base, since retraining sits on top of one specific set of weights.

The question the internet already answered is what happens when a foundation that spread because it was free starts charging the companies that built businesses on it. Chinese labs have not reached that point yet, but they have built the mechanism to get there.

America is running the same play from the other end

American companies have answered by buying the layers around the weights rather than competing at the weights themselves. Information reported in late August and confirmed by CNBC said that NVIDIA had agreed to acquire Hugging Face for $12.9B, the site where most open models are published and downloaded. TechCrunch reported that nothing has been signed and the deal could still collapse, and also that Stripe recently paid more than $7B for OpenRouter.

NVIDIA and the American cloud companies hold the chips and the data centers underneath everything, and buying the repository and the router puts American owners on both sides of the Chinese weights. The money has stayed American as well, which is the strongest caution against reading too much into the download figures. Ramp data reported by Fortune puts Anthropic at 43.5% of business AI spending in July 2026 and OpenAI at 39.7%, against 6.1% for services offering open and Chinese-developed models.

The evidence establishes that Chinese labs are supplying the foundations on which American software is built, but it does not establish that China has won the position. Beijing gatekeeps a great deal of what it builds, and every participant is still working out what its own layer is worth while the technology and the economics remain unsettled.

The pattern does not reward openness, and it never did on the internet, since what it rewards is the difficulty of leaving. Hron has said that Thomson Reuters can swap Qwen out for something else if it has to, and that is probably true today. Swapping out the chips underneath the model, or the marketplace that routes the requests, would take considerably longer. The third move only pays if the second one held, and nobody yet knows whether a company that started on Qwen in 2026 will still find it cheaper to stay than to leave in 2030.

Weekend To-Do

• Audit what your stack actually runs on: List every AI feature in your product and name the model underneath each one. The ones you can't name are the ones to check first.

• Read the Kimi K3 license: Skim the revenue and attribution clauses on Hugging Face so you know whether your business would fall inside them at scale.

• Price the switch: Estimate what it would cost to move one AI feature onto a different base model. That number is your real dependency, not whatever the vendor promises.

Friday Poll

Headlines You Actually Need

• US tells the G20 to leave AI alone: Washington pressed G20 ministers in Chapel Hill to avoid writing new AI rules and adopt its "Carolina Principles," which China has signed.

• Notion plans a 30% headcount jump: Ivan Zhao is adding roughly a third more staff even though the company already runs over 700 AI agents alongside about 1.1k employees.

• Power and cooling are the quiet winners: Asian transformer and liquid-cooling suppliers are cashing in on a data center build-out McKinsey puts near $7T by 2030.

Meme Of The Day

The Toolkit

• Continue: Open-source AI code assistant for VS Code and JetBrains, with full control over models and context.

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

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

Rate This Edition