Innovation & Tech

Anysphere Weighs Exit From Chinese Open-Weight AI Models

House investigators say Anysphere may phase out Chinese open-weight models, while warning that tighter federal rules could raise costs before trusted U.S. substitutes emerge.

By Daniel Okafor

4 min read

Anysphere Weighs Moving Cursor Away From Chinese Open-Weight Models
Anysphere Weighs Moving Cursor Away From Chinese Open-Weight ModelsQiqiGuo / pixabay

What's News

  • House report released Nov. 20, 2026: Anysphere will explore phasing out Chinese-developed open-weight models, with no timetable.
  • Cursor’s technical report identifies the Moonshot AI base behind Composer 2 as Kimi K2.5.
  • Airbnb reported roughly 80,000 daily guest-host interactions and 250,000 daily support cases and calls handled by customized Qwen models.
  • DoorDash deployed 25 Chinese model families; non-trivial use meant at least $5,000 in computing and serving costs.
  • The report makes 12 recommendations, including NIST-led provenance records and removal rules for high-risk critical infrastructure.

A joint House report released November 20, 2026, says Anysphere, developer of the Cursor AI coding platform, “plans to explore phasing out Chinese-developed open-weight AI models because of geopolitical risk.” The disclosure includes no final decision or timetable.

The House Committee on Homeland Security and House Select Committee on China published the findings after an investigation that began in April 2026. The inquiry expanded from Airbnb and Anysphere to DoorDash on July 31, 2026.

What did the investigation find about Composer 2?

The committees define open-weight models as systems whose trained parameters can be downloaded, run and modified without the original developer. That independence lets derivatives spread without a continuing commercial relationship with the model maker.

Cursor’s technical report identifies the base model behind Composer 2 as Moonshot AI’s Kimi K2.5. The committees said an independent developer first brought the base model’s identity into public view.

Composer 2 underwent continued pretraining on a code-heavy data mix. The developers then applied large-scale reinforcement learning in realistic Cursor sessions, using the same tools and harness as the deployed model.

The committees described Composer 2 as Cursor’s flagship proprietary model. The report treats its Kimi lineage as a case study in geopolitical supply risk.

Why did companies adopt Chinese open-weight models?

The committees identified cost, customization and “near-frontier performance” as adoption drivers. They found the United States lacks a competitive open-weight alternative at the price and performance the market demands.

Airbnb reported that customized versions of Alibaba’s Qwen models support an AI assistant handling roughly 80,000 guest and host interactions daily. A second fine-tuned model summarizes roughly 250,000 support cases and calls each day, Airbnb reported. Both run on U.S.-based service-provider infrastructure.

DoorDash reported that it had authorized and deployed 25 Chinese model families in a non-trivial manner. The company defined that threshold as “incurring at least $5,000 in computing and serving costs.”

DoorDash estimated that replacing one embedded Chinese-developed model could take up to six months. Most other uses could transition almost immediately, the company told investigators.

The committees compared published API prices for one million input tokens and one million output tokens. Their table listed these approximate costs:

  • Qwen3-235B-A22B-Instruct, Alibaba: $1.15
  • DeepSeek-R1, DeepSeek: $2.74
  • Kimi K2.6, Moonshot AI: $4.95
  • Claude Opus 4.8, Anthropic: $30
  • GPT-5.5, OpenAI: $35

The table put Qwen at $1.15, against $30 for Anthropic’s Claude Opus 4.8 and $35 for OpenAI’s GPT-5.5. These figures do not adjust for capability differences, caching, batch discounts, promotions or self-hosting costs, the committees cautioned.

They illustrate price pressure, not a like-for-like capability comparison. Open weights also let companies fine-tune models on their own data and deploy them on infrastructure they control, according to the report.

What would federal restrictions require?

The report’s 12 recommendations do not themselves enact restrictions. They would build on a Department of Homeland Security order excluding DeepSeek products and services from federal civilian procurement and requiring their removal from agency systems, including third-party products built on DeepSeek.

The committee chairmen recommend extending those restrictions to other Chinese-developed models and their derivatives across federal information systems and federal contract work. They would allow narrow waivers for research, testing and defensive security.

They also recommend barring the models and derivatives from narrowly defined high-risk critical-infrastructure functions. Removal schedules would depend on trusted alternatives, with temporary exceptions where no trusted substitute exists, according to the report.

To make the rules enforceable, the committees recommend a model-provenance standard led by the National Institute of Standards and Technology. A machine-readable AI bill of materials would document base models, significant fine-tuning and distillation sources.

The standard would cover federal contractors and AI used in high-risk critical-infrastructure functions. It would remain voluntary elsewhere, according to the recommendations.

The report also proposes a national American open-weight AI initiative. It would combine public computing resources, shared evaluation infrastructure and voluntary licensing of American frontier-model outputs to qualified open-weight developers for training.

Federal anchor demand would complete the proposed package, the committees said. The initiative seeks to give domestic developers computing scale, shared testing and distribution.

What could make the phase-out costly?

The committees warn that removing Chinese models before trusted open-weight alternatives exist could raise costs. It could also concentrate the market among closed proprietary providers.

Anysphere’s exploratory plan therefore illustrates the implementation test identified in the report. Restrictions need reliable lineage records to locate embedded models and credible substitutes to avoid trading Chinese open weights for higher costs or closed-model dependence.

The security concern remains separate from hosting location, the report says. U.S. infrastructure can reduce direct data exposure, but it does not change how the underlying model was trained or how it may behave under conditions ordinary evaluations miss.

Under the report’s plan, agencies and contractors would need traceable model histories, while developers would need cheaper trusted alternatives. The committees warn that restrictions without those substitutes would exchange Chinese open-weight dependence for higher costs and concentration among closed-model providers.

Source: unite.ai

Share this article:

More from Daniel Okafor

Daniel Okafor

Show full bio

Correspondent covering money, funding and market shifts around entrepreneurship, startups and venture funding at Business Bearings.

2,538 articles

Related articles

« Previous article