“The next AI leader will not simply produce more intelligence. It will create more economic value from every dollar invested.”
Two developments materially refine the U.S.–China AI thesis. The United States is demonstrating that frontier leadership and low-cost diffusion can coexist. China, meanwhile, is revealing that its strategy of abundant intelligence requires substantial capital investment.
The distinction between the two AI economies remains valid. But the competitive boundary is shifting.
The Evidence
two shifts
-
On October 7, Anthropic introduced Claude Haiku 5.5, reporting average running costs approximately 75% below Haiku 4.5. For prompts below 100,000 tokens, published input and output prices fell by 90%. Anthropic also reduced Sonnet 5.5's cache-read pricing, lowering the cost of many agentic workloads.
American AI companies are no longer relying exclusively on frontier capability and premium pricing. They are competing directly on the economics of mass deployment. The U.S. is beginning to combine frontier superiority with lower-cost intelligence. signal: Frontier leadership and low-cost diffusion can now coexist.
-
An October 8 Financial Times report highlights research from SemiAnalysis, originally published September 25, estimating China's data-center fleet at more than 24 GW, with approximately 50 GW of additional projects at various stages of development. The U.S. comparison is approximately 56 GW projected for year-end 2026. Crucially, China's pipeline is not installed capacity, and much of its existing stock is not suitable for advanced AI workloads.
SemiAnalysis estimates that Alibaba, Tencent and Baidu together invested approximately $20 billion in the second quarter of 2026, more than doubling year over year. All three reported negative free cash flow. China is therefore confronting a familiar tension: accelerating AI infrastructure investment ahead of demonstrable returns. signal: Abundant intelligence now demands enormous, unproven capital.
From Evidence to Thesis
the next equilibrium
Implications for the Supercycle
what to measure
Cost Per Workload
Cost per successfully completed workload becomes a decisive gauge of whether cheaper inference actually converts into value.
Productive Utilization
Productive compute utilization separates installed capacity from genuine, revenue-bearing deployment.
Incremental ROIC
Incremental return on invested capital tests whether rising capital intensity is justified by returns.
Cash-Flow Conversion
Favor demonstrated cash-flow conversion over AI capacity announcements alone.
the verdict
- The immediate beneficiaries remain significant: semiconductors, networking, energy infrastructure and platforms able to increase utilization.
- The second-order effect is more selective as the line between technological and financial advantage sharpens.
- America's frontier advantage is expanding into the economics of diffusion.
- China's diffusion advantage is increasingly dependent on its ability to finance, equip and utilize substantial computing infrastructure.
- The next phase will reward neither superior models nor cheaper intelligence automatically — only sustainable returns on capital.