Ramp counts 6.1% of AI-spending businesses paying for platforms that serve Chinese weights

Ramp's index counts businesses, not dollars, and that one design choice reconciles two findings sitting in the same dataset: the share of AI-spending companies buying model-serving platforms went from 4.5% in January to 6.1% in July, while Ramp's lead economist Ara Kharazian writes that the growth has not yet dented spending at OpenAI or Anthropic, where new buyers still land ^[3][11]. One buyer is one buyer whether it spends hundreds of dollars or millions. So the 1.6 points of movement, a 35.6% relative rise in six months ^[17], tells you adoption is broadening rather than that dollars are leaving; about one in sixteen AI-spending businesses now pays a serving platform ^[19], and running the same six-month pace once more gets you to roughly 8.3% by January ^[23], which is a minority habit and not yet a procurement default.

The named cases are where cash actually moves. Thomson Reuters put document-review work that previously ran on Claude onto Thomson-1, built from Snowdon, its own adaptation of Alibaba's open-source Qwen, with CTO Joel Hron arguing that a strong open foundation specialised deeply produces capable AI at lower cost than buying ever-larger models ^[5][6]. Read that as an allocation decision rather than a technology one: Thomson Reuters has pointed an engineering team at owning a model lineage instead of at whatever else that team would have shipped, and Anthropic no longer meters that volume. Harvey, whose backers include OpenAI, post-trained Harvey Tenet on Moonshot's Kimi K3 and says it outperformed both the base model and US frontier systems including Fable 5 and GPT-5.6 Sol on complex legal agentic tasks ^[7], having previously built its product by customising closed models from Anthropic, OpenAI and Google ^[8][9].

The price sheet is the mechanism. Z.AI's renamed GLM-5.3-Flash lists at 15 cents per million input tokens and 50 cents per million output ^[16], so output costs 3.33 times input ^[20], which is where an agentic loop's bill accumulates; a million tokens each way is 65 cents, and a billion each way is $650 ^[21]. Hugging Face's count puts the largest and most capable open model at a Chinese lab in almost every month of 2026, with the American challengers from Thinking Machines Lab and Meta not matching Kimi K3's scale or developer pull ^[10].

This is probably wrong, but the pressure I would expect first is in the high-volume, unglamorous tier, where a buyer with an ML team can post-train a cheap base and put two invoices side by side, rather than in the frontier price, which new and less sophisticated buyers keep paying ^[11]. The counter-thesis is Kharazian's own: what Ramp is capturing may be experimentation by the few firms staffed to do it. Three things would prove me wrong, and they are cheap to check: Ramp's next prints flattening near 6%, Thomson-1's document review quietly returning to Claude, or no third buyer of that size naming a production workload it moved. One more piece of arithmetic complicates the cheerful version, though, because Alex Brunicki of Backed VC describes these models as free to use ^[12] while DeepSeek is seeking $7.4 billion at a $74 billion valuation, or 10% of itself, to keep making them ^[13][18].