Thomson Reuters' Pivot to China AI

Programming note:

• We've invited a small group of former leaders from frontier AI labs and enterprise software platforms for a limited Q&A series on whether enterprise AI adoption and spending are gaining real traction. We'll introduce the first Q&A topics and experts this Friday, along with details on how readers can take part.
• ARPU returns next Tuesday with a look at Stripe's acquisition of OpenRouter.

Downsizing the AI Landlord

Thomson Reuters is a 175-year-old information giant whose core business is selling indispensable data to corporate lawyers billing by the hour. If you run a business like that, you generally do not want to get creative with your supply chain. The standard playbook is simple: you license the frontier AI everyone else is using, bolt it onto your proprietary data platform, and package the markup as a lucrative upsell to White-Shoe law firms.

But Thomson Reuters seems to have decided that paying Silicon Valley top-dollar for every single AI query is a bad business model.

This week, the company rolled out Thomson-1, a proprietary in-house AI model designed to take over high-volume workloads for its legal assistant tool, CoCounsel. Previously, that heavy lifting was handled by Anthropic's Claude.

And what is striking isn't just that they are weaning themselves off Silicon Valley's darling, but how they did it.

The company went on the internet and downloaded Qwen—an open-weight model built by the Chinese tech powerhouse Alibaba. They handed it to some researchers at Imperial College London to scrub out any political quirks, fine tuned it on their own vault of legal data, and put it to work reading contracts.

The company's CTO Joel Hron explained it like this:

Renting a house, you still have a roof over your head, and somebody's taking care of it, and it's great. But you're not building any equity that compounds into something valuable for you long term.

That is a polite way of saying the frontier labs had become his landlord—and that paying luxury rent to summarize boring contracts was a terrible way to build long-term enterprise value.

You Do Not Need God to Read a Contract

Why is a Western data giant suddenly deploying code from Hangzhou? Arithmetic—but arithmetic that comes in two flavors. You can rent cheaper, or you can peel off your high-volume workloads and build something in-house. Most companies fleeing frontier pricing have taken the first exit. Thomson Reuters took the second.

Start with the first, because it is where the money is most obvious. If you want to run every routine workflow through the absolute frontier of Western AI, it will cost you. Claude Fable 5 costs $50 per million output tokens.

Meanwhile, the Chinese AI ecosystem is running the classic manufacturing playbook known locally as involution—a state of hyper-competitive, cannibalistic price wars. It is a destructive domestic race to the bottom where virtually no one makes a profit, but it has the convenient side effect of exporting radical deflation to the rest of the tech world. DeepSeek's V4-Pro model costs $3.96 per million tokens during peak hours, and literally half that at night. Moonshot's Kimi K3 delivers near-frontier performance for $15 per million tokens.

For an enterprise customer running complex workflows that constantly read, parse, and verify multi-page legal filings, that pricing delta can be the difference between a viable product and an accounting disaster.

As Bloomberg recently documented, the margin math is already forcing American founders into pragmatic compromises.

Consider Polsia, a San Francisco startup whose platform went viral and saw its monthly AI bill go vertical—hitting $1 million in a single month. To stop incinerating his runway, the founder quietly shifted the bulk of his backend to Chinese models, instantly cutting his monthly spend to $100,000. Or take Flo Crivello, the founder of AI assistant Lindy, who slashed his company's AI operating costs by 90% after ditching Anthropic for DeepSeek—using the surplus capital to fund an offsite event in Napa Valley.

Crivello put it bluntly: "You do not need God to write your emails."

Apparently, you don't need God to read a non-disclosure agreement either.

Brute Force Capex vs. Constraint Engineering

The price gap is the direct result of a structural divergence in how both countries build technology.

The American strategy for AI is essentially brute-force capital. US hyperscalers are on track to spend $760 billion in capex in 2026 alone. The strategy is to pour oceans of concrete around scarce Nvidia GPUs, build massive 50-gigawatt data centers, and keep the resulting models locked behind proprietary, closed-weight APIs. If you spend $50 billion building a 1-gigawatt facility, you have to charge $50 per million tokens just to cover the depreciation on the hardware.

Which is the part Hron was really talking about. Every enterprise relying entirely on a frontier API is helping service somebody else's multi-billion-dollar mortgage on a data center they will never own.

China, by contrast, was forced into constraint engineering.

Deprived of the latest American chips by Washington's export controls, Chinese labs couldn't simply throw more silicon at the problem. Operating with roughly one-seventh the high-end compute of their Silicon Valley rivals, they had to optimize the software. DeepSeek and Moonshot pioneered hyper-efficient architectures where only 3% to 4% of a model’s neural parameters activate for any given prompt, and then they gave the weights away for free.

The irony of Washington's export controls is that they functioned as an accidental industrial policy for efficiency. By starving China of brute-force hardware, the US forced Chinese developers to commoditize the software layer. In doing so, they created the exact weapon that undermines Silicon Valley's entire economic model: open, portable intelligence.

The American tech sector is already addicted to the resulting margin relief, with nearly 200 firms—including Proton and Y Combinator—actively lobbying against restrictions on Chinese models.

Thomson Reuters represents the next, inevitable phase of this shift. They didn’t just switch to a cheaper API; they took China's open-weight commodity, fine-tuned it on proprietary domain expertise, and began chipping away at their dependency on closed American models.

Silicon Valley's frontier labs bet hundreds of billions of dollars that intelligence would be an oligopoly where they could collect premium rents forever. They are about to discover how hard it is to maintain a luxury real estate empire when their biggest enterprise tenants only rent the penthouse for special occasions—and build the rest of the house themselves for free.

Signal Stack

The operating reality beneath the headlines.

• OpenAI's Head of Data Centers Chris Malone Is Out in Latest Exec Exit (CNBC) — Malone joined from Meta and Google in March 2025 to help oversee a roughly $600 billion compute buildout through 2030 and is now the fourth senior departure in about six weeks, following four more in April—meaning the management layer responsible for the largest private infrastructure commitment in tech is turning over faster than the buildout it was hired to execute.
• US to Tell Partners They Must Pick Sides in AI Race With China (Reuters) — A draft State Department letter tells 35 countries that joining Beijing's new AI organization will disqualify them from the US-led one, meaning Washington is now treating access to American chips and models as an alliance requiring exclusivity rather than a commercial relationship countries can hold alongside others.

📺 On Our Channel

What Could Break the AI Spending Boom?

The dot-com boom left behind enormous amounts of unused fiber. Many investors lost money, but the infrastructure eventually powered broadband, streaming and cloud computing. AI may follow the same pattern—with one crucial difference. Fiber could wait decades for demand. AI hardware cannot. In our latest video, we examine the three timing risks that could prevent the AI buildout from earning back its cost before the next spending cycle begins.

Watch ARPU's deep dive on YouTube (28 Mins)

You received this message because you are subscribed to ARPU newsletter. You can unsubscribe using the link below.

If a friend forwarded you this message, sign up here to get it in your inbox.