Major technology companies including Airbnb and DoorDash have integrated Chinese artificial intelligence models into their core operations, according to findings from House committees. Simultaneously, India's government-backed artificial intelligence initiative confronts backlash after its chosen startup began commercializing Chinese models rather than developing homegrown alternatives.

Cursor, the coding platform owned by SpaceX, is considering a transition away from Chinese open-weight artificial intelligence models due to what House investigators describe as "geopolitical risk." The platform's signature offering, Composer 2, was constructed using Moonshot AI's Kimi K2.5 as its foundation.

According to reporting by CNBC's Kai Nicol-Schwarz, details of this potential shift emerged from an advance version of a report jointly released Friday by the House Homeland Security Committee and the Select Committee on China. The investigation found that Anysphere, the company behind Cursor, communicated these intentions to the committees.

Anysphere "plans to explore phasing out PRC-developed open-weight models because of geopolitical risk,"

House Homeland Security Committee and Select Committee on China

The committees characterized this as an exploratory initiative without a concrete timeline or final decision. Requests for comment from Cursor, SpaceX, and DoorDash went unanswered, while Airbnb declined to provide a statement.

Three companies under scrutiny

The committees initiated their investigation in April by sending inquiries to Anysphere and Airbnb, expanding to include DoorDash in July. The three organizations employ Chinese models in distinct operational contexts, the report determined.

Airbnb operates modified versions of Alibaba's Qwen infrastructure on American servers. One deployment manages an assistant responding to roughly 80,000 daily exchanges between guests and hosts. A separate system processes approximately 250,000 support inquiries and calls each day.

DoorDash disclosed deployment of 25 separate Chinese model families in active production, with each requiring a minimum investment of $5,000 to operate. Replacing a single model embedded within a complex system could demand six months of work. The company indicated that most other implementations could transition almost immediately.

In Cursor's case, the Chinese foundation only became widely known after an independent developer made the discovery public. Cursor's official technical documentation identifies Kimi K2.5 as the basis for its development. The company subsequently enhanced this foundation through additional training focused on code. SpaceX finalized its acquisition of Cursor in August at an implied valuation of $60 billion in equity value, per CNBC reporting.

Affordability, performance and opacity

Adoption of Chinese models stems from three primary factors: reduced expenses, capacity for customization, and performance approaching cutting-edge capabilities, according to the committees' findings. The United States currently lacks an open-weight alternative offering comparable pricing. The report further notes that much of China's advancement derives from techniques and knowledge extracted from American-developed models. OpenAI disclosed in September that individuals connected to Moonshot had attempted to access its reasoning capabilities. Maintaining model infrastructure on American servers prevents data transfer to China, the committees observe, though this does not alter the model's original training methodology.

The committees advocate for broadening the existing federal prohibition on DeepSeek to encompass additional Chinese developers. Their models would face exclusion from federal procurement and critical infrastructure applications involving elevated risk. The committees additionally propose establishing standards for tracking model origins and launching a federally-supported American open-weight program.

Cautionary language appears throughout the report regarding rapid implementation. Mandating corporate abandonment of Chinese models before dependable substitutes become available would elevate operational expenses, the report warns. This scenario would also concentrate market power among a limited number of proprietary systems. The committees identify emerging open models from Reflection AI and Mistral as potentially positive developments.

India's sovereign AI initiative embraces Chinese models

A parallel situation is unfolding within India. Sarvam, a startup receiving backing from the government's IndiaAI Mission to construct a domestically-developed model, has commenced offering access to DeepSeek and other Chinese models, according to reporting by Bloomberg's Saritha Rai.

The company's newly launched Token Factory service hosts models from DeepSeek, Moonshot, and Z.ai within India's borders, alongside Sarvam's proprietary model and Google's Gemma. The service already processes more than 450 billion tokens daily using DeepSeek V4.1 Flash, Sarvam reported. Online commentators questioned the rationale behind a lab established with subsidized computing resources reselling Chinese models.

"We haven't changed our motto. We continue to build AI models for India,"

Pratyush Kumar, Sarvam co-founder

Vivek Raghavan, the other co-founder, stated that the computing infrastructure supporting the service derives entirely from private sector investment. Retaining inference operations within India additionally ensures that user information remains within national boundaries, he explained.

Sarvam maintains supporters of its approach. Nikhil Pahwa, founder of MediaNama, contended to the publication that Sarvam must generate revenue within a competitive marketplace. The controversy emerged shortly after Sarvam secured a contract with India's defence ministry to develop artificial intelligence applications for military use, MediaNama reported. The startup achieved India's newest AI unicorn status in June.

Source: The Next Web