US Pushes Full-Stack AI Exports as China Captures Market with Open-Source Models
White House program aims to lock allies into American AI infrastructure while Beijing's low-cost alternatives gain ground in developing markets.
The Trump administration launched a coordinated push to export integrated American AI technology packages to allied nations, opening a proposal window in April 2026 for industry consortia to deliver hardware, software, and standards as unified systems—even as Chinese open-source models captured seven of the top ten download spots on developer platforms by late 2025.
The Department of Commerce began accepting applications for the American AI Exports Program on April 1, with a June 30 deadline for proposals. The initiative, mandated by a July 2025 executive order, requires participating companies to bundle compute infrastructure, foundation models, deployment software, and governance frameworks into complete Technology stacks rather than selling components separately. Secretary of Commerce Howard Lutnick framed the program as delivering on “President Trump’s directive to ensure that American AI systems are deployed at scale around the world.”
The program stems from Executive Order 14320, signed July 23, 2025, which established a three-pillar AI strategy encompassing domestic innovation, infrastructure buildout, and international technology diplomacy. The third pillar explicitly tasks federal agencies with mobilising financing tools and diplomatic channels to promote adoption of US AI platforms by allied governments.
China’s open-source counter-strategy
While Washington structures its exports around proprietary full-stack integration, Chinese developers have pursued the opposite approach. Models from DeepSeek and Alibaba’s Qwen family dominated download rankings on Hugging Face between November and December 2025, per analysis by the US-China Economic and Security Review Commission. The commission’s March 2026 report found that Chinese models accounted for seven of the platform’s ten most-downloaded offerings during that period, reflecting a strategy prioritising rapid diffusion over margin protection.
DeepSeek released a preview of its V4 model on April 24, 2026, pricing output tokens at $3.48 per million—less than 12% of OpenAI’s $30 rate and 14% of Anthropic’s $25 rate, according to Fortune. The pricing gap reflects fundamental architectural differences: most leading US developers restrict access through paid APIs and subscription tiers, while Chinese competitors distribute model weights freely and monetise through cloud hosting or enterprise support.
Export control contradictions
The full-stack export initiative runs parallel to tightened controls on advanced semiconductors, creating strategic tensions. White House Office of Science and Technology Policy Director Michael Kratsios accused Chinese entities of conducting “industrial-scale” campaigns to extract knowledge from frontier US models through systematic querying, as reported by CNN Business in April 2026. Yet the administration simultaneously approved exports of Nvidia’s H200 chips to China in December 2025, a policy reversal that Commerce Secretary Lutnick defended by noting Huawei’s limited domestic production capacity of roughly 200,000 AI chips annually.
Smuggling networks exposed in March 2026 revealed systematic diversion of restricted chips through Southeast Asian shell companies, prompting Congress to pass the Chip Security Act on March 26. The legislation mandates embedded location verification in advanced semiconductors, though enforcement mechanisms remain under development, per the Bloomsbury Intelligence and Security Institute.
“We believe that independent partners are critical to unlocking the prosperity AI adoption can open to all of us. That is why the President launched the American AI Exports Program.”
— Michael Kratsios, White House Office of Science and Technology Policy Director
Allied nations face ecosystem choice
The American AI Exports Program offers federal financing tools, government-to-government advocacy, and data centre operational support to countries adopting US technology stacks. Microsoft’s announced $10 billion four-year investment in Japan, beginning in 2026, exemplifies the integrated infrastructure deployments the program aims to replicate across allied markets. Japan has explicitly aligned the investment with its “Sovereign AI” strategy, creating a template for how other nations might structure partnerships.
But the sovereignty framing contains inherent contradictions. Analysis by the New Lines Institute notes that full-stack integration deepens technical dependencies on US platforms, standards, and update cycles—limiting recipient nations’ ability to modify systems or switch providers without prohibitive switching costs. China’s open-source approach, by contrast, allows governments to fork codebases and maintain local control, though at the cost of ongoing compatibility with upstream development.
| Dimension | US Full-Stack | China Open-Source |
|---|---|---|
| Access Model | Proprietary APIs, licenses | Permissive open weights |
| Pricing | $15-30/M tokens | $3-5/M tokens |
| Recipient Control | Limited customisation | Full fork rights |
| Standards Lock-In | High | Low |
| Update Dependencies | Vendor-controlled | Community-driven |
Compute capacity projections
The strategic stakes centre on raw computational advantage. Modelling by the Institute for Progress, cited by the Council on Foreign Relations, projected that a total embargo on advanced chip exports to China would maintain US compute capacity at roughly ten times China’s level in 2026. Aggressive exports of H200-class chips, however, could compress that advantage to single-digit multiples within two years, narrowing the performance gap between frontier models developed in each country.
The December 2025 policy shift permitting H200 exports suggests the administration prioritises semiconductor industry revenue and diplomatic flexibility over maximising the compute gap. China legally imported approximately one million downgraded Nvidia chips in 2024, Secretary Lutnick testified, providing substantial capacity even under restricted specifications.
What to watch
Proposal submissions to the American AI Exports Program close June 30, with initial awards expected in Q3 2026. The composition of winning consortia will reveal whether hyperscalers like Microsoft and Google dominate, or whether chip manufacturers and cloud infrastructure providers secure independent pathways. Watch for allied government responses to US versus Chinese offers—particularly in Southeast Asia and Latin America, where both powers are competing for first-mover adoption. Open-source model performance benchmarks will indicate whether China’s cost advantage translates into capability parity, potentially eroding the technical rationale for paying premium prices for closed US systems. Finally, enforcement actions under the Chip Security Act will test whether location verification closes smuggling channels or merely adds friction to an adaptable gray market.