“CAPEX was the race. ROIC is the reckoning.”
The first phase of the AI trade is ending. Compute scarcity created the winners; capital efficiency will separate them.
The U.S.–China contest is accelerating that transition. America still controls much of the frontier. China is attacking the economics of that advantage through cheaper models, domestic substitution, offshore compute access and alternative financing. The market consequence is decisive: AI CAPEX is no longer the signal — the return on AI CAPEX is.
The investment equation has moved from GPUs → Capacity → Valuation to CAPEX → Utilization → Revenue → Free Cash Flow → ROIC. The next winners will not be those building the most compute. They will be those making compute earn the most.
The Great Split
Two systems, two equations
-
The American model is straightforward: Capital → Frontier Compute → Superior Intelligence → Premium Workloads → Cash Flow → More Capital. Its advantage remains formidable — advanced silicon, hyperscale infrastructure, frontier models, software ecosystems, enterprise distribution and deep capital markets. But that advantage is becoming extraordinarily expensive.
The financial question is no longer whether U.S. companies can keep spending. They can. It is whether AI revenue can grow faster than AI capital employed. If yes, the American frontier becomes one of history's great compounding machines. If not, today's infrastructure boom becomes tomorrow's ROIC compression. signal: America seeks higher value per unit of compute — ROIC through capability.
-
China is optimizing a different equation: Efficient Models → Lower System Cost → Cheaper Intelligence → More Deployment → Higher Utilization. DeepSeek compresses inference economics, Huawei attacks compute dependency, domestic memory and networking reduce system costs, and offshore capacity bridges remaining hardware gaps.
China does not need to maximize revenue per token. It needs to maximize economic output per dollar of AI capital. signal: China seeks more utilization per unit of capital — ROIC through volume.
-
GPUs are migrating from technology equipment toward financeable productive infrastructure. The relevant economics increasingly resemble Asset Cost → Financing Cost → Utilization → Contracted Revenue → Cash Flow → Residual Value — opening AI to private credit, infrastructure funds, sovereign capital, insurers and leasing vehicles.
The AI supercycle can now lever the balance sheet of the financial system, increasing the potential size and duration of the boom. It also introduces the next risks: leverage, refinancing, utilization and obsolescence. AI is becoming a credit cycle as well as an equity cycle. signal: AI is becoming a credit cycle as well as an equity cycle.
-
The first phase rewarded scarcity: GPU → HBM → Networking → Data Centers. The second rewards efficiency around scarcity: inference optimization, memory efficiency, networking utilization, cooling, power efficiency and workload orchestration.
The thesis changes from owning the bottleneck to owning who makes the bottleneck earn more — a much broader, and far more selective, trade. signal: A broader trade, but a more selective one.
-
China's biggest contribution to the global AI economy may ultimately be price compression. As intelligence becomes cheaper, Inference Price ↓ → Consumption ↑↑ → Model Margins ↓ → Infrastructure Utilization ↑ → Application Economics ↑.
Model scarcity falls, infrastructure volume rises, application value rises. The biggest winners may not be the companies producing the model. They may increasingly be the companies owning scarce infrastructure — or using commoditized intelligence to generate measurable productivity and cash flow. signal: Economic rents migrate from the model to infrastructure and applications.
The Great Split Becomes a Financial Contest
ROIC, two ways
🇺🇸 United States
Frontier CAPEX → superior intelligence → premium workloads → higher revenue per compute unit. ROIC through capability.
🇨🇳 China
Lower-cost stack → cheaper intelligence → mass deployment → higher utilization. ROIC through volume.
The decisive question
No longer who wins AI? but which system converts AI capital into cash flow faster? That is the metric financial markets will eventually price.
The Next Market Divide
Inside the AI trade itself
What Markets Should Watch
Five numbers that matter more than another benchmark
The valuation premium moves to capital efficiency
- The indiscriminate AI CAPEX trade is ending — the next phase rewards returns, not announcements.
- U.S. frontier leadership remains the strongest capability trade, but its capital intensity means monetization must now catch up with investment.
- China is the global AI deflation engine, pressuring model pricing while expanding inference volumes.
- Infrastructure remains structurally advantaged, but selection matters — favor assets with scarcity, utilization and pricing power across HBM, networking, optics, power and cooling.
- As models commoditize, economic rents migrate toward companies turning cheap intelligence into productivity.
- Credit becomes the next frontier and the next risk — financing can extend the cycle, but leverage will expose weak utilization quickly.
- The decisive metric becomes Free Cash Flow / AI Capital Employed.