← Back to feed

The US State Department sent a classified diplomatic cable to every embassy and consular post worldwide on April 25, instructing diplomatic staff to raise “concerns over adversaries’ extraction and distillation of US AI models” with their foreign government counterparts.

The cable, dated Friday and seen by Reuters, names DeepSeek, Moonshot AI, and MiniMax by name. It states that “a separate demarche request and message has been sent to Beijing for raising with China” — making it a two-track pressure campaign, both multilateral and bilateral.

The timing is precise and intentional. DeepSeek launched its V4 model the same day, explicitly engineered for Huawei Ascend 950 chips. That’s not a minor detail: by building on Huawei silicon, DeepSeek sidesteps the export controls that blocked Nvidia hardware sales to Chinese entities. The V4-Pro runs 1.6 trillion parameters across 49 billion active at inference time. The V4-Flash runs 284 billion total, 13 billion active. Both default to 1 million token context. Pricing: $3.48/M output for Pro, $0.28/M for Flash, against $30/M for GPT-5.4 and $25/M for Claude Opus 4.6.

What the Cable Actually Says

The State Department cable has three stated purposes:

1. Warn allies of risks from using AI models distilled from US-origin proprietary models
2. Lay groundwork for potential follow-up US government action
3. Counter the ability of foreign actors to release competitive products at “a fraction” of development cost

Distillation is the mechanism at issue — training smaller models on outputs from larger frontier systems to dramatically cut training cost. Anthropic alleged in February that DeepSeek used 24,000 fraudulent accounts to run 16 million Claude API exchanges as training data. OpenAI made parallel allegations to Congress.

The White House Office of Science and Technology Policy director posted publicly that the US has “obtained evidence that foreign firms, primarily in China, are conducting large-scale distillation campaigns” and that “action will be taken.”

Legislative Acceleration

The House Foreign Affairs Committee acted in parallel, unanimously passing two bills:

• Prevent Theft of U.S. AI Models Act — criminal penalties for extracting AI model weights or architecture without authorization
• Hardware Technology Controls Multilateral Alignment (MATCH) Act — coordination mechanism for allied export controls on advanced semiconductor equipment

Both passed unanimously, signaling cross-party consensus unusual for tech legislation. They still require Senate action.

China’s Response

The Chinese Embassy dismissed the allegations as “groundless” and “deliberate attacks on China’s development and progress in the AI industry.” DeepSeek maintains its V3 model was trained on naturally occurring web-crawled data, without intentional use of OpenAI synthetic outputs.

DeepSeek V4-Pro claims open-source SOTA on agentic coding benchmarks at 80.6% SWE-bench Verified — placing it level with Gemini 3.1 Pro Preview, below Claude Opus 4.7 (87.6%) and GPT-5.5. The model is MIT licensed and fully available on Hugging Face.

The Geometry of the Conflict

The State Department cable and DeepSeek V4’s Huawei launch on the same day create the sharpest illustration yet of the structural conflict: US export controls are designed to slow China’s AI hardware access; China is routing around them at the model level via distillation and at the hardware level via Ascend silicon.

V4-Flash at $0.28/M output is cheaper than most US models on a per-token basis. If it performs within reach of frontier US models — and independent benchmarks will run in the coming weeks — the economic argument for using US-origin proprietary models in non-classified settings weakens.

A Trump-Xi summit is expected next month, with semiconductor controls and AI IP disputes explicitly on the agenda. The cable and the Huawei launch together mean AI is arriving at that summit with sharper edges than either side may have planned.