The US Wants to Approve Every AI Chip Sale on Earth. Here's What That Actually Means.
The US Commerce Department is circulating a 129-page draft that would require government approval for exporting advanced AI chips to virtually any country — not just adversaries, but allies too. The framework creates tiered licensing, reciprocal investment requirements, and could reshape the global AI data center map. Here's what's actually in the document, who's fighting it, and why Nvidia should be nervous.
Close-up of a computer microchip on a circuit board
Key Points
•The US Commerce Department is circulating a 129-page draft that would require government approval for exporting advanced AI chips to virtually any country on earth — not just adversaries like China, but allies too
•The most striking provision is a buy-America-to-sell-America requirement: large foreign buyers would need to invest $1 in US AI infrastructure for every $1 they spend on their own domestic AI buildout
•The Trump White House is publicly distancing itself from the draft, yet industry sources describe it as Diffusion 2.0 — potentially even more intrusive than Biden-era rules
•Nvidia, AMD, and Broadcom are watching closely as these controls could reshape the global AI data center map
A 129-page document that nobody was supposed to see
Somewhere inside the US Commerce Department, a 129-page draft is making the rounds. It hasn't been formally announced. It hasn't been published in the Federal Register. And if you ask senior White House officials about it, they'll tell you it doesn't reflect the president's position.
But the document exists. Multiple outlets — TechCrunch, Tom's Hardware, The Register, Axios — have confirmed its contents. And what it describes is one of the most ambitious attempts to control the global flow of technology since the Cold War [1][2][3].
The core idea is straightforward: the US government would require a license for exporting advanced AI chips — think Nvidia's GB300, AMD's MI400 series, and similar accelerators — to essentially any country on earth. Not just China, Russia, and Iran, which are already under heavy restrictions. Every country. Canada. Germany. Japan. Australia. Everyone.
If that sounds like a dramatic expansion of federal authority over the semiconductor industry, that's because it is. And the details get more interesting from there.
The tiered licensing framework uses Nvidia's GB300 GPU as the benchmark for order-size thresholds.
The draft doesn't treat all chip exports the same. It creates a scaled licensing framework based on order size, using Nvidia's next-generation GB300 GPU as the benchmark [2].
Small orders (under ~1,000 GPUs): Expedited review. This covers most enterprise customers, startups, and research institutions. The idea is that a company buying a few hundred GPUs for a training cluster doesn't pose a national security risk worth agonizing over. You'd still need a license, but the process would be fast.
Medium-scale deployments: Pre-authorization required, plus transparency requirements. Buyers would need to disclose their business activities, allow potential on-site inspections by US authorities, and demonstrate that the chips won't be diverted to prohibited end users.
Massive deployments (200,000+ GPUs in one country by one operator): This is where it gets serious. Orders at this scale would require formal intergovernmental negotiations — meaning the buyer's government would need to provide national security assurances directly to Washington. And here's the kicker: the buyer would also need to make a matching investment in US AI infrastructure [2][3].
That last requirement — the matching investment — is the real story.
Buy America to sell America
The Commerce Department has already tested this model in practice. Export licenses granted to buyers in the UAE and Saudi Arabia reportedly include a condition: for every dollar spent on domestic AI infrastructure, the buyer must invest a dollar in AI infrastructure inside the United States [2].
Think about what this means. If a sovereign wealth fund in the Gulf wants to build a 100,000-GPU AI cluster to power a national AI program, they'd also need to fund a comparable facility in the US. The "price" of buying American chips isn't just the sticker price — it's a commitment to build on American soil.
This isn't traditional export control. Traditional export control says: "You can't buy this because we don't trust you." This says: "You can buy this, but you have to pay twice — once to us, once into our economy."
It's mercantilist tech policy. And if formalized in the 129-page draft, it would fundamentally reshape how the global AI data center map develops over the next decade. Countries and companies planning large-scale AI deployments would need to factor in the cost of co-investing in American infrastructure. Some would comply and bring investment into the US. Others might look for alternatives — Chinese chips, custom silicon, or simply scaling back their ambitions [2][3].
The question isn't whether this is clever industrial policy. It clearly is. The question is whether it's legal, enforceable, and whether the diplomatic blowback is worth it.
The White House says one thing. Commerce does another.
Here's where the politics get messy.
President Trump rescinded Biden's AI Diffusion Rule shortly after taking office, calling it "burdensome, overreaching, and disastrous." The Biden-era rule divided the world into three tiers: allies who could buy freely, a middle tier with volume caps, and adversaries who were cut off entirely. It was complicated, and the semiconductor industry hated it [1][3].
Trump's position, at least publicly, is that he wants to expand AI chip exports, not restrict them. A senior White House official told reporters that the Commerce draft "does not reflect what President Trump has said on export controls nor the direction of the Trump administration on encouraging export of the American AI stack" [1].
But the Commerce Department isn't freelancing. The draft exists because career officials and political appointees at Commerce believe the US needs a framework for managing AI chip exports that goes beyond case-by-case decisions. The current system — where each major export deal gets negotiated individually — doesn't scale as demand explodes globally.
Industry lobbyists are already calling the new draft Diffusion 2.0 — and in some ways it could be even more intrusive than what Biden proposed.
Industry lobbyists are already calling the new draft "Diffusion 2.0." And they have a point: in some ways, the new framework could be even more intrusive than what Biden proposed. Biden's rule at least gave close allies a free pass. This draft would require licensing for everyone — allies included. Biden's rule didn't condition exports on reciprocal investment. This one does [2][4].
The tension between the White House's pro-export rhetoric and Commerce's pro-control instincts is real, and it won't resolve easily. Export controls live in a bureaucratic gray zone where career officials have significant latitude. Even if Trump publicly opposes broad restrictions, the machinery of government can move in its own direction — at least until someone in the Oval Office picks up the phone.
What this means for Nvidia, AMD, and the chip industry
Nvidia is the most exposed company in this story, and it's not even close.
Nvidia controls an estimated 80%+ of the AI training chip market. Its revenue from data center GPUs hit $62.3 billion last quarter alone. A significant and growing portion of that revenue comes from international customers — hyperscalers building out non-US data centers, sovereign AI programs in the Middle East and Asia, and a global ecosystem of cloud providers, research labs, and enterprises [1][4].
If every international sale above a certain threshold requires a Commerce Department license — and large sales require reciprocal US investment — the friction in Nvidia's sales process increases dramatically. Deals that currently close in weeks could take months. Customers who planned to deploy in Dubai or Singapore might scale back or delay. And competitors — particularly Chinese chipmakers who are racing to close the gap — would get more time to catch up.
AMD faces similar exposure, though its AI chip market share is smaller. Broadcom's custom ASIC business could actually benefit: if hyperscalers face regulatory friction buying off-the-shelf Nvidia GPUs for international deployments, they might accelerate their shift toward designing their own chips — which is exactly what Broadcom helps them do.
Nvidia's stock has already lost roughly $200 billion in market capitalization since late February, driven partly by broader AI sentiment concerns and partly by the sell-the-news dynamic after blowout Q4 earnings. These export control headlines add another layer of uncertainty [4].
The global AI data center map, redrawn
Step back from the policy details and look at the big picture. What the Commerce draft is really trying to do is use America's chip dominance as a tool for reshaping global AI infrastructure development.
The logic goes like this: Every major country wants to build AI capabilities. To do that, they need the best chips. The best chips are American. Therefore, America can set the terms.
Those terms, if the draft becomes policy, would push the world toward a hub-and-spoke model where the US sits at the center. Countries that want large-scale AI deployments would need to co-invest in American infrastructure, accept US oversight of their facilities, and provide national security assurances. In exchange, they get access to the best hardware.
This is not how the global technology market has worked for the past 30 years. The post-Cold War tech ecosystem was built on relatively free trade in hardware and software, with export controls reserved for weapons, nuclear technology, and a narrow list of dual-use items. AI chips are now being treated like dual-use technology — civilian products with military and intelligence applications that justify strategic control.
The race to build alternatives
China has been under heavy US chip restrictions since 2022. The result? A massive domestic effort to build competitive AI chips. Huawei's Ascend 910C is now being deployed at scale for AI training in Chinese data centers. It's not as good as Nvidia's H100 or GB200 — but it's improving rapidly, and every year of US restrictions gives Chinese chipmakers more time and more motivation to close the gap.
The Commerce draft could accelerate this dynamic globally. If buying American chips comes with investment mandates, inspections, and intergovernmental negotiations, some countries might decide the hassle isn't worth it. The Middle East has the money to comply with reciprocal investment requirements. India might not. Southeast Asian nations building AI capabilities on tight budgets could find themselves priced out entirely.
The semiconductor industry's nightmare scenario is a fragmented global chip market where different regions develop incompatible AI hardware ecosystems — an AI iron curtain where American chips dominate one hemisphere and Chinese chips dominate another, with everyone in between forced to choose sides.
That scenario might sound extreme. But every new layer of export control makes it slightly more likely.
What happens next
The draft is just a draft. It hasn't been formally proposed, let alone finalized. The White House is publicly distancing itself from the most aggressive provisions. Industry lobbyists are pushing back hard. And the internal clash between Commerce's control instincts and the White House's pro-export messaging has to resolve one way or another [1][3].
But the direction of travel is clear. Both the Biden and Trump administrations — for all their differences — agree that AI chips are strategic assets that require government oversight. They disagree on how much oversight and what form it should take, but neither is arguing for a free-market, sell-to-anyone approach.
For the semiconductor industry, the message is: plan for a world where international AI chip sales require government approval, where large deals come with strings attached, and where the regulatory landscape can shift with every election cycle.
For the rest of the world, the message is different but equally clear: if you want the best AI hardware, be prepared to invest in America's AI future too. That's the deal on the table. Whether it's a fair deal depends on who's sitting across from you.
