The case for building more U.S. data centers is consistently framed around one competitor: China. Executives, investors, and lawmakers have spent 2026 describing hyperscale buildout as a strategic necessity to preserve American leadership over Chinese AI labs. At the same time, the models those Chinese labs publish are being downloaded, hosted, and shipped into U.S. consumer products by the very hyperscalers running that buildout.
The framing
Investor Kevin O’Leary told Fox News in June 2026 that slowed state-level data-center approvals were allowing China to pull ahead, arguing the country now benefits because it is “getting more power sooner” and using it to train its own AI systems.
Testifying before Congress, OpenAI CEO Sam Altman told senators that continued AI leadership depended on rapid infrastructure investment, framing the buildout as foundational to sustaining a U.S. edge.
Research from the Brookings Institution and testimony before the House Select Committee on the Chinese Communist Party have made a similar case: the constraint on U.S. AI leadership is not model quality but the physical capacity to build and power data centers fast enough to keep pace with Chinese infrastructure expansion.
Combined 2026 capital spending by Microsoft, Amazon, Meta, and Alphabet on data centers and related AI infrastructure is projected near $785 billion, according to Moody’s Ratings, May 18, 2026.
What actually runs on that infrastructure
Inside the facilities built under that framing, Chinese open-weight models are now standard, hosted components. GitHub made Moonshot AI’s Kimi K2.7 Code generally available in the Copilot model picker on July 1, 2026, the first open-weight model added to a roster otherwise built from OpenAI, Anthropic, Google, and Microsoft systems. The model, a trillion-parameter system from a Beijing-based lab, runs on Microsoft Azure infrastructure rather than Moonshot’s own servers, placing a Chinese-developed model inside the same U.S. cloud capacity built partly to compete with Chinese AI development.
Microsoft Research’s own Fara1.5 browser-agent models, released in May 2026 and hosted on Azure AI Foundry, are built directly on Alibaba’s Qwen3.5 base architecture rather than a Microsoft-originated foundation model. Both cases follow the same pattern: once a Chinese lab publishes model weights, any U.S. hyperscaler can download the checkpoint, run it on domestic GPU clusters, and integrate it into products marketed to U.S. consumers without routing any traffic through Chinese-operated servers.
The hyperscalers are the hosts, not just bystanders
Chinese open-weight models are not reaching U.S. infrastructure through side channels — the same companies building out data centers under the China-competition banner are the ones adding these models to their managed catalogs. Amazon Web Services lists Qwen3, Kimi K2, MiniMax, DeepSeek V3.2, and GLM 4.7 as fully managed models in Amazon Bedrock and Amazon SageMaker JumpStart, served through Project Mantle, AWS’s distributed inference engine, on a catalog that expanded from 18 to roughly 24 open-weight models between re:Invent 2025 and February 2026.
Google Cloud’s Vertex AI Model Garden lists Qwen and DeepSeek among more than 200 validated open-weight models available as one-click managed endpoints, with DeepSeek-V3.2 also offered as a fully managed API. Microsoft’s position runs through both Azure and GitHub: Kimi K2.7 Code became a selectable model in GitHub Copilot on July 1, 2026, hosted on Azure infrastructure, and Microsoft Research’s own Fara1.5 agent models are built directly on Alibaba’s Qwen3.5 base and distributed through Azure AI Foundry.
GPU-specialized cloud tenants such as CoreWeave and Lambda, which lease capacity inside hyperscale and colocation facilities to AI startups, host the same model families for customers who need raw GPU access rather than a managed endpoint. In each case, the hosting company controls the infrastructure layer — the GPUs, the region, the billing meter — while the model weights and underlying architecture originate with a Chinese developer.
Local stakes
Wyoming illustrates the pattern at the state level. Gov. Mark Gordon signed Executive Order 2026-03 on June 3, 2026, directing agencies to accelerate data-center policy tied to the state’s AI-infrastructure push, and Laramie County approved Project Jade in January 2026, an AI-oriented campus initially sized at 1.8 gigawatts. Meta already operates a hyperscale facility in Cheyenne. Any hyperscaler with an existing practice of hosting Chinese open-weight models — Microsoft, Amazon, or Google among them — could run those same models on Wyoming-sited compute without that use being distinguishable, at the facility level, from any other inference workload marketed as part of the domestic AI buildout.
References
- Fox News, “Kevin O’Leary warns China is winning the AI race because U.S. states are slowing data center production,” June 19, 2026
- Seeking Alpha/AP, “OpenAI CEO Sam Altman and other US tech leaders testify to Congress on AI competition with China”
- China’s secret weapon in AI race with US? Lots of cheap energy, May 28, 2026, citing Morgan Stanley data
- Brookings Institution, “Competing AI strategies for the US and China,” April 2026 congressional testimony
- Windows Forum, Best-AI.org, AlphaSignal, TechTimes, Let’s Data Science — coverage of Kimi K2.7 Code general availability in GitHub Copilot, July 1, 2026
- Microsoft Research; KuCoin, Decrypt, MarkTechPost — coverage of Fara1.5 release built on Qwen3.5, May 2026
- SoftwareSeni, “AWS Bedrock Open-Weight Models — Running Qwen, Kimi K2, and MiniMax Without Infrastructure Overhead,” April 2026
- Google Cloud Blog, “Take an open model from discovery to endpoint on Vertex AI”; Google Cloud documentation, Vertex AI release notes and DeepSeek/Qwen model pages
- Office of Wyoming Governor Mark Gordon, Executive Order 2026-03, June 3, 2026, as reported by CryptoBriefing
- Hyperscaler Capex Forecast 2026 Raised to $785 Billion – Moody’s Ratings, May 18, 2026 – computeforecast.com
