The AI export program that only five companies can enter

The American AI Exports Program launched with the aim of pushing US AI into friendly and swing markets. Nine months in, the executives it was meant to help are telling Politico it is "not where they hoped it'd be." The structural reason is buried in the eligibility rules, and it is more interesting than the political complaint.

The program, run out of Commerce's International Trade Administration, offers three things to qualifying US vendors: expedited export licensing, access to federal financing, and diplomatic backing on foreign deals. To qualify, applicants must present a complete "stack package", chips, cloud infrastructure, models, cybersecurity, and applications, assembled as a single offering to the foreign buyer.

Read that eligibility list against the US AI market. The set of companies that can assemble a qualifying stack without going hat-in-hand to a competitor is very small. Nvidia supplies the chips. AWS, Azure, and Google Cloud supply the infrastructure. OpenAI, Anthropic, and Google supply the frontier models. Cybersecurity and applications are the only layers with real competition, and they are the layers doing the least work in the package.

A mid-sized inference-optimisation vendor cannot qualify alone. A vertical AI firm in healthcare or logistics cannot qualify alone. They can join a consortium led by a hyperscaler, on the hyperscaler's terms — which is the same distribution arrangement they already had, now with a government stamp on it. The stated pitch of the AAEP was to widen the field of US AI exporters. The mechanism narrows it.

What the primary release actually says. Commerce's implementation notice frames the stack requirement as an interoperability and security control: buyers get an integrated US system rather than cherry-picked components that might not work together or might carry supply-chain risk. That logic is defensible. It is also the logic that, applied honestly, guarantees the beneficiaries are the four or five companies that already own the interoperable stack. Both things are true at once, which is why the complaint is structural rather than fixable by paperwork.

Then there is the Anthropic episode. Earlier this quarter the administration ordered Anthropic to cut foreign access to its latest models on security grounds. Within days the order was reversed. Commerce is now trying to sell the AAEP to foreign buyers as a stable channel for US AI supply — while the same administration has demonstrated it will restrict a frontier lab's foreign access by executive order and then un-restrict it under industry pressure, on a timescale of days.

If you are a sovereign AI procurement officer in Riyadh, Jakarta, or Warsaw, you have just been shown two things. First, that US AI supply carries a regulatory optionality, the possibility that access is cut at short notice, that was not priced into contracts signed before July 2026. Second, that the reversal mechanism runs on lobbying speed, not on any published process you could plan against. Both facts belong in the risk column of any deployment decision.

This is the structural tension Axios flagged in June, and it has not been resolved. The same executive branch is running export promotion (AAEP) and export restriction (chip controls, entity-list actions, ad-hoc model cutoffs) in parallel, with no disclosed arbitration body when they conflict. "Dominate globally" and "deny adversary access" are compatible strategies only if you specify a priority ordering. The administration has not specified one. Each individual decision therefore reads to foreign buyers as ad-hoc, because it is.

The Zhipu piece in Bloomberg on the same day as the Politico report is worth noticing here. Zhang Peng, arguing frontier AI should stay open to all countries, is running the standard Chinese AI diplomacy line — open source, predictable access, no export-control surprises. Every US promotion stumble is a receipt for that pitch. Swing-market buyers who were told a year ago that Chinese AI came with political risk are now being told that US AI comes with a different political risk, delivered on a faster clock.

The contrarian read is worth stating. Commerce Secretary Lutnick's line, "ambitious by design", is not empty. Large government export programs routinely draw industry complaints in their first six to twelve months, and some of them (Ex-Im Bank, USTDA) grew into functional channels once the initial friction was worked out. If the stack-package rule is a configuration choice, it can be revised — Just Security's July analysis argues it is correctable through specific structural changes rather than abandonment, which is probably right on the merits.

The question is whether it is a configuration choice or a captured feature. Once Nvidia, the three hyperscalers, and two frontier labs are the effective gatekeepers of AAEP participation, the incentive to lobby for the stack requirement's preservation is substantial. The default trajectory of programs that concentrate benefits on a small number of incumbents is that the incumbents keep them concentrated. Q4 will tell us whether the design brief or the lobbying brief is winning.

For a strategist trying to price this. If you are a mid-sized US AI vendor with international ambitions, the AAEP is not a distribution channel you can plan around; it is a channel you can access only through a hyperscaler partnership, which was your position before the program existed. If you are a hyperscaler, the program is a modest tailwind on deals you were already going to win. If you are a foreign buyer, the effective price of US AI supply just went up by the size of the regulatory-optionality discount you now have to apply.

The intelligence-explosion frame reads this as friction in the US ability to project AI capability abroad at the pace the frame's own logic would require. A country that believes it is in a race, and simultaneously restricts its own runners, is telling the market something about how seriously it takes the race, or about how many races it is actually running.

The AAEP can still work. It requires Commerce to decide whether the program's job is to help the incumbents export what they were exporting anyway, or to widen the set of US AI companies that reach foreign buyers. Those are different programs. The stack-package rule chose the first one. The pitch abroad promised the second one. Something has to give.

Footnotes

¶