# Congress presses DoorDash to disclose its use of a Chinese AI model > Source: > Published: 2026-08-03 11:29:28+00:00 *DoorDash's use of Moonshot AI's Kimi model has turned a cost-saving engineering choice into a Washington question about Chinese AI inside American companies.* A reported congressional letter to DoorDash CEO Tony Xu asks the company to disclose which Chinese AI models it uses, what data touches them, and what security testing happened before those systems reached production. The deadline is Aug. 14. Start there. This isn't really a fight over food delivery. It's a fight over whether a cheap open-weight model can become part of an American company's software pipeline before regulators know what happened. DoorDash co-founder Andy Fang gave lawmakers the opening. In public comments picked up in recent coverage, Fang said DoorDash routes lower-level engineering work to Kimi K2.6, Moonshot AI's open-weight model, while reserving Anthropic's Fable for the hardest work. DoorDash's own DashBench write-up gives the commercial reason in plain numbers: a Kimi K2.6 scout paired with Claude Fable 5 reviewer had the strongest weighted recall and F1 on a 105-case valid subset, at 65.2% weighted recall and 75.3% weighted F1, with a listed cost of $3.81 per pull request. That is enough. A founder described a cheaper code-review setup. Congress read it as a national security disclosure. The concern from House Select Committee on China chairman John Moolenaar and House Homeland Security Committee chairman Andrew Garbarino didn't come from nowhere. On April 29, the Homeland Security Committee announced a joint investigation into PRC-developed AI models, naming DeepSeek, Alibaba, Moonshot AI, and MiniMax. The same announcement said the chairmen had already sent letters to Airbnb and Anysphere, the maker of Cursor, over the use of Chinese AI systems in American products. ## The model is cheap enough to change behavior The cost pressure is real. Fortune reported in July that one million output tokens from Anthropic's Fable cost about $50, while DeepSeek-V4-Pro cost roughly $0.87 and Z.ai's GLM-5.2 cost $4.40. Moonshot's Kimi K3 was listed at $15. You don't need to love Chinese AI policy to understand that math. If your engineering team is running thousands of code-review, refactor, and agent tasks a day, the cheaper model doesn't sit in a research channel for long. It moves into the workflow. That's the pinch for startups. Legal teams know how to review a SaaS contract with a mainland Chinese vendor. They may be far less ready for an open-weight model downloaded from Hugging Face or routed through a model marketplace, especially when inference happens on company-controlled infrastructure and no obvious vendor handoff appears in the logs. Congress is signaling that this distinction is thin. The model's origin still counts. The April committee statement also leaned on an Office of Science and Technology Policy memo warning that foreign entities, primarily based in China, were conducting deliberate industrial-scale efforts to distill American frontier models through proxy accounts and other coordinated methods. The public record doesn't prove that DoorDash did anything improper. It does show why Moonshot's name now carries policy risk, even when the immediate use case is an internal code-review pipeline. ## DoorDash is the example others can understand DoorDash is a better target than a small AI tooling company. Ordinary readers know what it is. Its latest annual filing said it had 31,400 employees at the end of 2025, and market trackers put its value in the tens of billions of dollars in July. If a company that large is using Kimi in engineering work, you can assume smaller companies are making similar calls with less paper trail and less legal review. Frankly, that is the part founders should care about. The issue isn't whether Kimi K2.6 is useful. DoorDash's benchmark says it is. The issue is whether your company can explain where the model came from, what it saw, who approved it, and what would happen if a lawmaker or customer asked for the same answers by a fixed deadline. DoorDash hasn't published a detailed response laying out whether it will narrow its Kimi use, defend the setup, or give the committees everything they requested. Aug. 14 will tell other companies how expensive disclosure becomes. If the answer turns into a continuing obligation rather than a one-time letter, the cheap model will still be cheap. The governance around it won't be. **Also read:** [MiniMax's H3 Video Model Undercuts Sora and Veo on Price and Openness](https://startupfortune.com/minimaxs-h3-video-model-undercuts-sora-and-veo-on-price-and-openness/) • [AI Agents Are Learning to Lie and Cheat Because Training Rewards Winning](https://startupfortune.com/ai-agents-are-learning-to-lie-and-cheat-because-training-rewards-winning/) • [Zhipu AI stays silent as reports point to a trillion-parameter GLM-5.5](https://startupfortune.com/zhipu-ai-stays-silent-as-reports-point-to-a-trillion-parameter-glm-55/)