In January, Rep. Randy Fine of Florida introduced H.R. 6996, the Full AI Stack Export Promotion Act. The bill would make it U.S. policy to promote the export of the entire American AI technology stack—semiconductors, cloud infrastructure, AI models, data pipelines, and technical standards—as a bundled package to foreign governments and markets. It advanced through the House Foreign Affairs Committee last week on a bipartisan vote of 37-7. President Trump had already signed an executive order in July 2025 directing the Commerce Department to establish the American AI Exports Program, and this April, Commerce published a Federal Register call for proposals from industry-led consortia to deliver these packages. Selected consortia receive federal financing, expedited export licensing, and diplomatic support.
The circuit extends globally, with the state constructing new markets.
This is the export-side extension of a program already built with public money. The CHIPS Act subsidized the core infrastructure for AI. This legislation promotes selling it abroad. It’s the same companies and the same national security framing, applied to new markets. The question nobody in Congress seems to be asking is whether the financial architecture underneath justifies either.
What has emerged is a state-constructed market—one where public subsidies, national security framing, and self-referential financial architecture combine to create a closed circuit that benefits a concentrated cluster of firms while making itself politically impossible to question. That prevents serious scrutiny into whether this circuit is financially sustainable, or a house of cards waiting to come apart.
The Domestic Architecture
The CHIPS and Science Act committed $52.7 billion in public subsidies for semiconductor manufacturing. The national security case for that investment was constructed, in significant part, by the National Security Commission on Artificial Intelligence, which operated from 2018 to 2021. The NSCAI’s recommendations shaped key legislative provisions. The Semiconductor Industry Association anchored a coalition of more than 20 organizations that pushed the bill through Congress with overwhelming bipartisan support.
Former Google CEO and executive chairman Eric Schmidt chaired the NSCAI during this entire period while actively investing in AI startups—including Beacon, just five months after his appointment—and holding over $5.3 billion in Alphabet stock. Government ethics advisers told CNBC that the arrangement raised enormous conflict-of-interest concerns. It did not matter. The geopolitical framing—that semiconductor independence was a matter of national survival against China—made scrutiny politically costly. Questioning the subsidies sounded like questioning American security.
Pandemic supply chain disruptions, a broad industry coalition, a real and serious deficit in domestic semiconductor manufacturing—the United States had almost no leading-edge chip fabrication on its own soil—and the NSCAI’s recommendations converged to produce a policy architecture that no individual participant fully controlled. That convergence is precisely what makes the problem harder to address than simple corruption. The overlap between the people who built the national security case for AI subsidies and the people whose portfolios benefit from them didn’t need to be hidden. The framing made it irrelevant.
The Cloud Credit Circuit
The domestic subsidies were justified by projections of massive, growing demand for AI compute. But a January 2025 Federal Trade Commission report documented something the financial press has started calling “AI circular deals.” The structure is straightforward: A cloud provider—Microsoft, Google, Amazon—invests billions in an AI startup. A substantial portion flows back as cloud computing credits for the investor’s own platform. The startup spends the credits training models on the investor’s infrastructure. The investor books the resulting usage as revenue. That revenue growth supports the stock price that justifies the next round of investment.
Microsoft’s $13 billion commitment to OpenAI is the most visible example—much of that money never left Microsoft’s own balance sheet, circulating as Azure credits that Microsoft then recorded as cloud revenue. The FTC found similar patterns across Google-Anthropic and Amazon-Anthropic partnerships. Nvidia reinvests GPU profits into infrastructure startups that use the funds to buy more Nvidia chips. What looks like booming market demand for AI compute is, in significant part, financially manufactured.
This does not mean there is zero genuine demand. But the demand projections used to justify $52.7 billion in public subsidies—and now an export promotion program—may be substantially inflated by this self-referential architecture. Chinese AI model DeepSeek’s demonstration that competitive AI can be built with dramatically less compute makes the question unavoidable: How much of what we are subsidizing and exporting reflects real technological need, and how much reflects a financial loop?
The Export Turn
H.R. 6996 and the American AI Exports Program complete the circuit. The companies that received CHIPS Act subsidies—Intel, TSMC’s Arizona operations, Samsung—are the same companies that will anchor the consortia delivering “full-stack AI export packages” under the new program. The Commerce Department’s April 2026 Federal Register notice specifies that these packages must include AI-optimized hardware, cloud services, data pipelines, and networking—the entire vertical stack, bundled for export with federal support. The bill directs the Economic Diplomacy Action Group, chaired by the secretary of state, to mobilize every available federal tool to open foreign markets. The U.S. government becomes the export agent for infrastructure it already subsidized.
The structural problem is specific. Export promotion creates new demand for the same infrastructure whose domestic demand is already partly self-referential. If the cloud credit circuit inflates domestic demand through financially engineered consumption, and that inflated demand justifies both public subsidies and further build-out, then promoting the export of that infrastructure adds another layer of demand generation—this time backed by the full weight of U.S. trade diplomacy. The circuit extends globally, with the state constructing new markets.
Consider what this means in practice. The government subsidizes chip production. The financial architecture generates demand signals that justify the scale of production. Congress then authorizes the executive branch to promote export of the resulting infrastructure. Each stage reinforces the next, and national security provides the justifying discourse at every turn. The same framing that made it politically impossible to question the subsidies now makes it politically impossible to question the export promotion.
The Question Nobody Is Asking
Congress is simultaneously debating export controls and export promotion—restricting chip sales to China while promoting the full AI stack to everyone else. Both sides of this debate accept the same premise: The AI infrastructure build-out reflects genuine strategic necessity. Export hawks want to deny adversaries access. Export promoters want to ensure allies build on American systems. Nobody is asking whether the demand projections underlying both the domestic subsidies and the export program are partly artifacts of the financial architecture itself, and what happens if what has been described broadly as a bubble pops while we are in the midst of inflating it.
The United States has committed tens of billions in public funds to build AI infrastructure on the basis of demand signals generated, in significant part, by a self-referential financial loop. It is now preparing to use diplomatic and trade machinery to export that infrastructure globally—creating new markets for the same firms, justified by the same national security logic, without ever establishing that the underlying demand is real.
Subsidize, build, export, repeat. That is the circuit. And the bipartisan consensus that holds it together has made it all but impossible to examine.
