The Great AI Financial Divergence

“America finances the frontier. China challenges its economics. The next battle is over who captures the profits.”

The U.S.–China AI competition is entering a new phase of financial divergence. Semiconductor earnings are accelerating, hyperscaler investment is expanding, and software efficiency is beginning to redistribute the economics of intelligence.

Three market signals expose the emerging fault line.

China — The Memory Earnings Boom. Chinese memory manufacturer CXMT projects revenues of RMB 265–270 billion for the first nine months of 2026, compared with RMB 32.08 billion a year earlier. That represents approximately 726–742% revenue growth. Projected net profit reaches RMB 131–132.5 billion, compared with a RMB 5.3 billion loss in the corresponding period of 2025. The implied nine-month net margin is approximately 48.5–50.0%.

These figures remain preliminary company projections rather than audited results. CXMT manufactures DRAM. Its earnings reflect memory pricing, product mix, manufacturing economics, utilization and demand across multiple end markets, including AI infrastructure. They cannot be attributed entirely to AI. Nor do they demonstrate technological parity with American frontier accelerators. The distinction matters because the earnings boom may contain both structural growth and cyclical pricing effects.

United States — The Infrastructure Investment Cycle. Amazon's reported $220 billion AI infrastructure investment plan illustrates the extraordinary capital requirements of American frontier computing. The buildout extends beyond GPUs into networking, memory, power, cooling and data centers. America's technological advantage remains formidable, supported by leading models, advanced accelerators, hyperscale infrastructure and deep capital markets.

But the financial burden is rising. The market must increasingly assess how rapidly installed capacity becomes productive and whether recurring revenues can absorb depreciation, operating costs and financing requirements.

Software — The Efficiency Counterforce. Nvidia-backed doubleAI has introduced Maxwell, a code-execution model designed to improve computing efficiency. The company reports a 67% reduction in computing costs for one customer and a doubling of inference speed for another. These are company-reported results, not independently established industry benchmarks. Nevertheless, the direction is consequential. Software optimization can potentially increase the economic output generated by existing infrastructure without proportional additional capital expenditure.

The three signals reveal a fundamental change in how the AI Supercycle should be valued. The first phase rewarded exposure to infrastructure expansion. The next phase will increasingly differentiate companies according to their ability to retain the economic benefits of that expansion.

Three mechanisms are now operating simultaneously. Technological scarcity supports premium pricing for advanced computing. Industrial substitution creates new markets for alternative hardware and domestic suppliers. Software efficiency reduces the cost of completing commercially valuable AI workloads.

The resulting financial transmission is not uniform. Infrastructure suppliers may benefit from shortages. Hyperscalers may experience growing revenues alongside rising capital intensity. Model providers may face pricing pressure. Industrial enterprises may capture productivity gains from increasingly affordable intelligence.

The next AI investment cycle will be defined by profit migration, not computing capacity alone. That is the central thesis.

The United States retains significant advantages in frontier models, accelerated computing, software ecosystems and capital formation. But maintaining those advantages requires sustained investment in increasingly complex physical infrastructure.

Four financial pressures are becoming more important. Capital expenditure: larger computing clusters require substantial investment in chips, networking, electricity and cooling. Depreciation: rapid technological progress can shorten the economic life of installed equipment. Execution: construction delays and grid constraints can postpone productive deployment. Financing: greater external funding requirements increase sensitivity to interest rates and credit conditions.

The result is growing differentiation within the American AI ecosystem. Highly profitable hyperscalers can finance expansion through operating cash flows and strong balance sheets. Leveraged infrastructure providers face greater exposure to refinancing, utilization and technology-obsolescence risks.

For investors, the relevant distinction is therefore not simply between companies investing heavily and those investing cautiously. It is between companies whose infrastructure spending is supported by durable demand and those whose financial models depend on future utilization assumptions that remain unproven.

CXMT's preliminary results demonstrate that Chinese semiconductor manufacturers can capture substantial earnings during periods of exceptional technology demand. But the underlying investment thesis must distinguish three mechanisms.

Cyclical memory profits. Global shortages and higher selling prices can produce extraordinary earnings. These profits may reverse when additional manufacturing capacity becomes available or pricing conditions normalize.

Structural semiconductor substitution. China's domestic technology strategy supports alternative semiconductor suppliers and encourages investment in local computing infrastructure. That can create a larger addressable market for Chinese manufacturers. However, domestic market-share gains do not automatically establish superior technology, lower total costs or sustainable margins.

Industrial AI diffusion. The broader opportunity lies in applying affordable intelligence across manufacturing, logistics, robotics, energy and mobility. Here, the economic surplus may accrue primarily to the companies deploying AI rather than the semiconductor or model suppliers.

These are distinct sources of economic value. China's memory earnings boom demonstrates its growing participation in the AI infrastructure cycle — not superior profitability across the Chinese AI ecosystem. The sustainability of margins, manufacturing efficiency and end-user productivity will determine the structural value of that expansion.

The two AI economies are developing different mechanisms of competitive advantage.

United States — Frontier Economics. Superior capability × premium pricing × global demand − infrastructure costs. America's advantage is technological leadership, global distribution and access to advanced computing. Its vulnerability is that inference commoditization can reduce pricing power while the cost of maintaining frontier infrastructure remains high.

China — Diffusion Economics. Affordable intelligence × industrial adoption × productivity gains − integration costs. China's opportunity is to expand the deployment of sufficiently capable intelligence throughout the physical economy. Its vulnerability is that adoption can increase without producing sustainable supplier margins or measurable productivity gains.

The distinction is financially consequential. The United States is seeking to preserve the premium attached to frontier capability. China is attempting to expand the economic value generated by more accessible intelligence. Neither strategy guarantees attractive shareholder returns. The competitive advantage will ultimately depend on combining sufficient capability, sustainable deployment costs and effective monetization.

The AI value chain is experiencing two opposing forces. Infrastructure scarcity supports pricing power across advanced semiconductors, memory, networking and power. Software efficiency reduces the resources required to deliver commercially useful intelligence. Lower inference costs can expand demand, but the resulting profits will be distributed unevenly.

This creates three potential areas of investment differentiation. Scarce infrastructure: advanced semiconductors, high-performance memory, networking and power assets where technological constraints sustain pricing power. Efficient computing platforms: operators capable of increasing productive utilization and recurring revenues faster than infrastructure costs. Industrial AI value capturers: enterprises converting lower-cost intelligence into reduced operating expenses, higher throughput, better asset utilization and stronger margins.

The greatest risk lies with business models where infrastructure expansion consistently outpaces monetization. The next market rotation may therefore favor earnings quality and demonstrable productivity over headline AI investment growth.

The next phase of the AI Supercycle should be evaluated through a capital-to-cash-flow sequence: Capital expenditure → Operational capacity → Productive utilization → Economic value → Free cash flow.

The relevant metric is not simply the cost per token. It is economic value per successfully completed workload. This incorporates computing cost, reliability, utilization and the commercial value of the task performed. It provides a more meaningful basis for comparing infrastructure investment, software efficiency and industrial deployment.

The financial test is whether the value generated by intelligence exceeds its full economic cost, including depreciation, financing and operating requirements.

The analytical framework remains: Technological Velocity × Capital Formation × Diffusion Velocity × Economic Capture.

The critical insight is that these forces operate at different speeds. Computing investment can accelerate before demand becomes profitable. Inference costs can decline before infrastructure is depreciated. Industrial adoption can expand before productivity gains become visible in corporate earnings.

These timing differences create the next valuation opportunities — and risks. For The Velocity Edge, the most important forward indicators are therefore the relationship between capital expenditure and recurring revenues, productive computing utilization, cost per completed workload and measurable industrial productivity.

Key Takeaways for