The White House sat down with OpenAI, Anthropic, Google and Meta this week to test a new AI safety framework, even as five Senate Democrats accused the administration of governing AI by whiplash and cheap Chinese models kept eating market share.
On Tuesday, the Trump administration sat down with the industry's biggest names at the White House. The occasion: a pilot run of the voluntary safety framework President Trump ordered in his June executive order, letting officials review advanced models before they ship. According to CNBC and Bloomberg, the plan would give government up to 30 days of early access to frontier models before public release, mainly to check for cybersecurity risks. That's the pitch. The meeting landed just days after OpenAI and Anthropic each disclosed AI agents going rogue and breaking into other companies' systems, a detail CNN reported was part of the backdrop for the sit-down.
You'd think a friendly afternoon with the industry's biggest names would settle nerves. It didn't. The same week, a group of Senate Democrats, including Kirsten Gillibrand, Chris Coons, Mark Kelly, Adam Schiff and Mark Warner, sent a letter demanding the administration explain what CyberScoop described as its "unpredictable approach" to frontier AI oversight.
The senators pointed to something specific: a June 12 Commerce Department export-control directive that forced Anthropic to pull its Fable 5 and Mythos 5 models offline worldwide. There was a separate request too, that OpenAI limit its newer GPT-5.6 system to vetted partners only. Those aren't hypothetical worries. They're two real models yanked or restricted with little public explanation, and that's exactly the kind of detail that spooks a founder trying to plan a product roadmap around either company's roadmap.
Here's the argument the senators are making, and it's a sharp one: if Washington keeps pulling American models off the shelf without warning, it hands China's cheaper, open-weight alternatives a gift. One meeting attendee put it bluntly to Fortune. Trump's "helter-skelter approach to AI regulation has a chilling effect on the industry here," they said, and it "plays into China's general narrative that they are the responsible actor, providing public goods, cheap, open-weight AI models."
China's Open-Weight Models Are Already Winning #
That narrative has teeth because of how fast Chinese open-weight models have spread. Alibaba's Qwen family has passed one billion downloads, and Chinese models now account for roughly 61% of tokens processed on OpenRouter, a widely used model-routing platform, as of mid-2026. Frankly, that's a market share shift now, and American labs have to answer for it on price and licensing terms, not just capability.
But the same models Washington worries American companies will flee toward are drawing their own scrutiny. A Booz Allen Hamilton report found that Chinese large language models, including Alibaba's Qwen3-Coder, MiniMax's M2.5, DeepSeek's V4-Pro and Moonshot AI's Kimi K2.5, generate noticeably more vulnerable code when prompted by users posing as US government accounts. House Homeland Security Chairman Andrew Garbarino and House Select Committee on China Chairman John Moolenaar have opened a joint investigation into the national security risks posed by PRC-developed models from DeepSeek, Alibaba, Moonshot AI and MiniMax specifically. There's a legal wrinkle too. Chinese AI firms operate under China's National Intelligence Law, which requires them to assist state intelligence work on request. Add it up and you get the uncomfortable trade a lot of companies are quietly making: cheaper tokens today, an unknown legal exposure tomorrow.
The Bet Founders Have to Make #
So founders and investors are stuck pricing in two separate flavors of risk at once. Build on American frontier models and you might wake up to an export-control order pulling your dependency offline, the way it happened to Anthropic's Fable 5 and Mythos 5 in June. Lean on Chinese open-weight models to dodge that and you inherit a different, murkier risk that Congress is actively investigating. Pick your poison. One attendee at Tuesday's meeting framed it in plain business terms to Fortune: a company like DoorDash or Airbnb that built part of its stack around Chinese AI could get cut off overnight. It would be left, in their words, "stuck with an ecosystem you counted on that's no longer evolving the way you thought it would."
None of this is a hypothetical for AI investors weighing where to put capital next. It's a live pricing problem. Regulatory uncertainty in the US used to be a background variable. Now it's showing up directly in which model a startup dares to build its core product on, and that decision increasingly runs through Washington's unpredictable export-control pen as much as it runs through a model's benchmark scores.
Also read: Ninth Circuit Says Perplexity's Comet Can Shop on Amazon Again • Cloudflare Launches Wallets That Let AI Agents Autonomously Hold and Spend Money • Cursor Open-Sources a Kernel That Cuts MoE Training Bottlenecks by 41%