Washington is packing its bags for Silicon Valley, and no, it’s not for a hoodie-and-kombucha tour.
Members of the House Foreign Affairs Committee are heading west to grill the people building the world’s most powerful AI systems, Google, OpenAI, Meta, and the rest, about a question that’s suddenly become a national-security obsession: when American AI gets “exported,” who exactly ends up holding the keys?
This trip lands in the middle of a widening tech cold war, where the hottest weapons aren’t missiles, they’re models, chips, and the know-how to train them.
AI exports: the new front line in the tech cold war
The U.S. has already been tightening the screws since 2022, especially on advanced chip exports to China. Those restrictions hit the high-end graphics processors, the workhorses used to train large AI models. If you can’t get the chips, training frontier AI gets a lot harder. That’s the theory, anyway.
Now lawmakers are looking beyond hardware. Because even if you block the chips, AI can still “travel” in other ways: cloud access, licensing deals, overseas subsidiaries, model weights shared with partners, or just American companies selling AI services abroad. The committee’s visit signals that Congress wants a clearer line between what’s normal business and what’s shipping strategic capability overseas.
Silicon Valley built the tools. Now it gets the scrutiny.
Silicon Valley likes to think of itself as the world’s innovation engine. Congress is increasingly treating it like a strategic asset that needs guarding.
The most advanced systems, think GPT-4 on the OpenAI side, or Anthropic’s Claude, already face limits in certain countries and contexts. That’s not just corporate caution; it’s the early shape of policy pressure. When a model can help write code, design chemicals, or supercharge intelligence analysis, “export” stops sounding like a trade term and starts sounding like a security briefing.
And the U.S. isn’t acting alone. Europe is finalizing its AI Act, Brussels’ big attempt to regulate AI by risk category. China, meanwhile, is pushing its own national champions and building an AI stack that doesn’t depend on American permission slips.
Money talks. Security yells.
AI is where the money is: hundreds of billions of dollars in investment sloshing around the sector, with companies racing to lock in customers, data, and compute. The problem for policymakers is that AI doesn’t stay politely “civilian.” The same model that drafts marketing copy can also help automate cyber operations or accelerate military R&D.
That’s why the likely core fight in these meetings will be classification, figuring out what can be sold abroad and what should be restricted. Sounds simple until you try to define it. Is the red line the model’s capability? The size of the training run? Access to certain tools? The ability to fine-tune? The availability of the model weights? In AI, the boundary between “harmless” and “dual-use” is a smudge, not a fence.
Big Tech’s usual argument is predictable: clamp down too hard and you’ll kneecap U.S. competitiveness while foreign rivals keep sprinting. And they’re not wrong that regulation can backfire. But they also have a habit of treating “national interest” as a synonym for “our quarterly numbers.”
This isn’t the first time America tried to bottle up code
If this all feels familiar, it should. The U.S. fought similar battles over cryptography in the 1990s, when Washington tried to restrict strong encryption exports like it was a weapons system. That didn’t age well. Encryption spread anyway, and the rules eventually loosened.
AI is trickier. A modern model can learn, improve, and get repurposed fast. Risk assessments go stale quickly. And because so much AI is delivered through cloud services, “export control” can turn into a game of whack-a-mole, less about shipping boxes and more about controlling access.
That’s the real reason this congressional trip matters: it’s a sign lawmakers are gearing up for a new era of tech regulation where the product isn’t a gadget, it’s capability. And Silicon Valley, for once, doesn’t get to pretend it’s just building apps.
