There are now too many related articles to read as isolated episodes. The clearer frame is to ask through which channels Chinese AI models are entering or reshaping the U.S. market: enterprise adoption, price pressure, open-weight distribution strategy, and capability transfer through distillation and defense reuse.
This hub reorganizes the existing coverage into four linked themes. DoorDash and Kimi show the adoption path. DeepSeek V4-Flash and OpenAI's price cuts show the pricing and workflow pressure. Kimi K3 and the U.S. open-weight split show the distribution fight. The distillation explainer and the PLA article show why the argument is now moving from commercial competition into national security.
Adoption
Chinese models first enter through practical workloads, not necessarily full-stack replacement.
Pricing
Low-cost APIs and response-format compatibility shift competition toward real workflow economics.
Distribution
Open-weight strategy changes who can deploy, adapt, and spread strong models.
Capability transfer
Distillation and defense reuse move valuable model behavior across political and institutional boundaries.
1. Adoption: DoorDash and the practical entry point
AdoptionThe DoorDash story matters because it shows how Chinese models can enter through lower-level or auxiliary workloads before anything like a full replacement. CNBC reported that DoorDash highlighted Moonshot AI's Kimi K2.6 as strong on cost and quality for some internal tasks, and that two House committees then asked for details on evaluation, safeguards, and exposure. That tension is the real point: what is rational for a company on price-performance grounds is increasingly treated by Washington as a strategic dependency question.
2. Pricing and workflow pressure: DeepSeek's practical push
Agent / PricingDeepSeek V4-Flash matters less as a benchmark event than as a workflow event. According to DeepSeek's own updates, the official public-beta release keeps the model name as deepseek-v4-flash, changes only the V4-Flash API, and leaves V4-Pro and the app/web models untouched. The Responses API documentation goes further by explicitly framing support around Codex-style and OpenAI-compatible workflows. That is how a model stops being just a cheaper alternative and becomes a drop-in operational option. In that frame, OpenAI's price cuts look like a direct response to mounting cost pressure.
3. Distribution strategy: Kimi K3 and the U.S. open-weight split
Open WeightKimi K3 should not be read only as another large-model launch. Reuters described Moonshot's move as a major open-model release closing in on U.S. rivals, while Microsoft and NVIDIA argued that open weights are a foundation for American AI diffusion and competitiveness. The important refinement here is conceptual: open weight is not the same as open source. OSI stresses that weights alone do not provide full reproducibility, code access, or data transparency. Yet they are still useful enough to shape enterprise deployment, which is exactly why the policy split is becoming more consequential.
4. Capability transfer and security: from distillation to military reuse
SecurityAnthropic said DeepSeek, Moonshot, and MiniMax used roughly 24,000 fraudulent accounts and more than 16 million Claude exchanges to extract reasoning, tool use, and coding capability. The key distinction is that knowledge distillation itself is a standard ML technique; what makes a “distillation attack” different is unauthorized, industrial-scale, evasive extraction. Reuters then pushed the story further by reporting that Chinese military researchers used outputs from OpenAI and Anthropic models in domestic defense-related research. Combined with Xi's call to accelerate military AI and unmanned intelligent technology, the issue becomes much bigger than contract violations. It becomes a question of how high-value AI capability crosses borders and gets repurposed.