The Quiet Exile of US AI: Harvey, Thomson Reuters and Cursor Are Running on Chinese Models

Here’s your Friday digest. The big story today isn’t a launch — it’s a quiet migration that’s been hiding in API logs.

AI Agents

The Quiet Exile of US AI: Harvey, Thomson Reuters and Cursor Are Running on Chinese Models

What happened. A 36Kr investigation has stitched together the API logs and disclosures showing that some of the most hyped Western AI products are, quietly, running on Chinese foundation models. Legal-AI darling Harvey (an OpenAI investee, ~$11B valuation) launched a self-developed model in August that is based on Moonshot’s Kimi K3 — trained on 150 Nvidia B300 GPUs for two months, lifting its legal-task pass rate from 10.8% to 19.7%. Thomson Reuters confirmed its “Snowdon” model is built on Alibaba’s Qwen3.5, recalibrated with Imperial College London. And Cursor’s Composer 2 was exposed via API logs as running on Kimi K2.5 — which Elon Musk confirmed.

Why it matters. The “AI race” narrative assumed US labs would dominate the model layer and everyone else would build on top. The opposite is happening at the application layer: when a $11B startup and a $40M-investing media giant both quietly swap in Chinese weights, it signals that open Chinese models are now cost-competitive enough to beat out Western incumbents on price and capability for specific tasks. OpenRouter data backs this up — US traffic to Chinese models has exceeded 30% weekly since February and peaked at 46%. Senator Tom Cotton’s July letter (claiming ~80% of US startups use Chinese models) is no longer hyperbole.

What’s next. The pressure is now on two fronts. Washington is weighing expanding export bans beyond DeepSeek, while Beijing is negotiating export controls on model weights with Alibaba, ByteDance and Zhipu AI. Both sides moving toward the same wall means the window for quietly swapping in Chinese models may be closing — and the companies that already did it face a compliance reckoning. Watch for disclosures and licensing changes at any Western AI product whose unit economics suddenly improved.

Source: 36Kr

WhatsApp is about to let you plug in third-party AI agents. Spotted in Android beta 2.35.3, WhatsApp is building an “Agents” section that lets users connect up to five third-party bots per account, each with its own dedicated chat and API key. The catch: agent conversations won’t be end-to-end encrypted, and the agent developer can read messages sent to it. Source: Digit

Alibaba’s Accio is deploying an “Exhibition Agent” across six European trade shows. Alibaba.com’s agentic platform (used by 10M+ SMEs) is now the “Official Agentic AI Partner” for events like Autumn Fair and ISPO, offering natural-language booth finding, supplier comparison, and RFQ drafting. It’s a concrete, real-world test of whether agents can genuinely help buyers at scale. Source: PR Newswire

AI News

Broadcom’s AI chip revenue surged 221% to $16.7B — and it’s projecting $230B by 2028. CEO Hock Tan says AI chip revenue will double to ~$115B in fiscal 2027, with Anthropic set to become its biggest custom-chip customer and OpenAI its second. The signal: hyperscalers are diversifying hard away from Nvidia-only supply. Source: The Star / Bloomberg

Nvidia finalized its $12.93B acquisition of Hugging Face. The deal — one of Nvidia’s biggest ever — moves the chipmaker beyond hardware into the developer-tools layer. Leadership says the open-access philosophy stays, positioning Nvidia to be “essential” to AI workflows regardless of whose chips run underneath. Source: Blockonomi

Snowflake jumped 20%+ on AI-driven growth. Q2 product revenue rose 37% YoY to $1.49B, and CEO Sridhar Ramaswamy attributed roughly half the momentum to AI products — evidence that the data layer, not just the model layer, is where AI dollars are landing. Source: Blockonomi