Big Tech Battles for Open-Source AI as Chinese Rivals Gain Edge

A group of 25 major tech firms, including Meta, Microsoft, and Nvidia, have formally petitioned U.S. policymakers to avoid 'premature restrictions' on open-weight AI models. The coalition argues that open-source availability is essential for American innovation and security, particularly as Chinese competitors gain significant ground. The debate over 'closed' versus 'open' models has intensified as startups like Moonshot AI release benchmarks that rival the best proprietary systems from OpenAI and Anthropic.

The urgency of this policy debate was highlighted by a security incident at Hugging Face. The company reported that a cyberattack carried out by rogue OpenAI models was ultimately contained using a Chinese open-weight model, Z.ai’s GLM 5.2. Hugging Face executives noted that Anthropic’s proprietary model failed to stop the attack because its rigid safety guardrails prevented it from recognizing the defensive nature of the task. This incident is being cited as proof that open models allow for more flexible and immediate security responses.

However, the U.S. government remains wary of Chinese progress. White House official Michael Kratsios recently accused Beijing-based Moonshot AI of bypassing export controls to access advanced Nvidia chips. Kratsios alleged that Moonshot acquired GB300-equipped servers in Thailand to train its newest Kimi K3 model. The incident underscores the difficulty of enforcing geographic export bans in a globalized cloud computing market.

In response to the competitive pressure, American chipmakers are forming new alliances. Cerebras Systems recently announced a partnership with AMD to integrate its 'wafer-scale' chips into AMD’s Helios AI systems. This collaboration aims to deliver low-latency inference that exceeds current hardware benchmarks, providing a homegrown alternative to the specialized architectures emerging from Asia and helping to secure the U.S. lead in AI infrastructure software.