Major tech firms push back on proposed open-weight AI restrictions as startups face cost pressures

A philosophical divide over open-weight artificial intelligence has crystallized within the technology industry, with major infrastructure providers standing apart from closed-model champions on the question of whether policymakers should restrict access to powerful, freely available models.

The debate erupted this week when the White House alleged that China's Moonshot AI engaged in intellectual property theft through model distillation—a technique where developers extract knowledge from one AI system to train another. The accusation triggered competing pressure campaigns: nearly 200 startups urged the Trump administration to preserve access to affordable Chinese alternatives, while a coalition of 25 major organizations including Microsoft, Nvidia, Meta, and Hugging Face released a statement cautioning Washington against "premature restrictions" on open weights. Anthropic and OpenAI pointedly declined to sign.

Why cost has become the deciding factor for builders#

The underlying tension stems from a stark pricing gap that has already shifted developer behavior. Moonshot's Kimi K3, a capable open-weight competitor, delivers comparable performance to Anthropic's premium Claude Fable 5 at roughly one-third the cost—though with slower inference speeds. For resource-conscious startups managing token expenses, that price differential has proven decisive. Across OpenRouter, a model routing platform, Chinese alternatives have claimed over 30% of weekly token usage from American customers since February, peaking at 46%.

Even Anthropic's recent Opus 5 launch, which halved the company's frontier-model pricing to $5 per million input tokens, still costs more per token than Kimi K3. Anthropic argues its edge lies in cost-per-completed-task rather than raw token pricing, emphasizing that more capable models may solve problems in fewer steps. But the math remains challenging for bootstrapped teams building on constrained budgets.

The industry's conflicting interests#

The statement defending open weights—titled "Open Weights and American AI Leadership"—characterizes distillation as a "widely used technique" fundamental to model development, not unlawful extraction. Signatories span infrastructure providers (Microsoft, Nvidia), cloud platforms, chip makers, and developer-focused organizations.

Nvidia's Jensen Huang told Axios that learning from other systems constitutes how intelligence itself develops, and that American companies "absolutely" should use excellent Chinese models. He acknowledged that privacy and contractual breaches warrant consequences, but rejected framing distillation as theft in principle.