“America leads the creation of frontier intelligence. China is challenging the cost of reproducing it.”
America finances the frontier. China compresses its economics. The next battle is over who captures the value.
New reporting indicates that two leading Chinese developers, DeepSeek and Moonshot AI, have relied substantially on outputs from OpenAI's most advanced model to improve their own systems. At the same time, independent testing shows Chinese open-weight models acquiring sophisticated agentic capabilities once associated primarily with America's proprietary frontier.
These developments strengthen our thesis but reveal a deeper mechanism: China's advantage may arise not only from producing intelligence more efficiently, but from accelerating the transfer of capabilities developed at far greater expense elsewhere. This introduces the economics of imitation, diffusion and value capture to the AI Supercycle.
The DeepSeek Inflection
Who pays for intelligence?
On October 28, The Information reported that DeepSeek and Moonshot AI relied heavily on outputs from OpenAI's GPT-6 Astra when developing DeepSeek's V4.1-Flash 1017 and Moonshot's Kimi K3.1 Pro.
The technique, model distillation, can reproduce aspects of a more capable model's performance while cutting training and deployment costs. Distillation itself is established; the controversy concerns unauthorized use of proprietary systems. The claims have not been independently verified, and neither company offered a substantive response.
The economics could be substantial. American developers incur enormous costs building frontier intelligence; competitors may acquire portions of it at far lower incremental cost: frontier investment → advanced capabilities → knowledge extraction → lower-cost models → global price competition. The company financing the breakthrough may not capture most of its subsequent value.
The Agentic Inflection
China's open models cross a threshold
Research published October 28 by security firm Irregular, reported by Reuters on October 29, provides a second signal. Z.AI's GLM-5.3 completed three of thirteen complex cybersecurity scenarios in controlled testing — the strongest performance Irregular has measured from an openly available model. Leading American proprietary models performed better in separate, non-comparable evaluations.
The important development: a downloadable Chinese model can now complete multistage tasks requiring sustained planning and coordinated execution. As open-weight models undertake complex work, their addressable market expands across software engineering, research, industrial operations and enterprise automation.
But greater autonomy adds security and compliance costs. The next test is therefore risk-adjusted productivity per completed task, not token price alone. Chinese models may offer attractive deployment economics while requiring more investment in oversight and reliability.
Wall Street Confirms The Next Test
Infrastructure vs. capacity to sustain it
Microsoft reported $18.6 billion in quarterly free cash flow despite $45.2 billion in computing infrastructure. Alphabet reported negative $9.07 billion even as Google Cloud revenue more than doubled to $30.71 billion. After the October 28 releases, Microsoft rose more than 3% in after-hours trading while Alphabet fell nearly 10%.
This does not establish superior technology. It shows capital markets differentiating between expanding AI infrastructure and the financial capacity to sustain it — a distinction that sharpens as competitively priced Chinese models broaden the supply of deployable intelligence. The potential result: expanding consumption, declining model-level pricing power, and increasingly selective returns on infrastructure.
The Great Division Of Intelligence
How the surplus gets split
America's advantage
Frontier innovation, advanced computing infrastructure, proprietary technology and access to capital.
China's challenge
Rapid capability absorption, software optimization, open-weight distribution and industrial deployment.
Commoditization
Capable open-weight systems may accelerate the commoditization of standardized inference workloads.
The deploying business
Firms deploying AI could capture a rising share of the surplus as capabilities become more accessible.
Five Signals For Investors
What to watch next
-
Whether American developers can limit unauthorized distillation without sacrificing legitimate adoption. signal: Can the frontier be defended without closing the market?
-
Whether DeepSeek, Moonshot and Z.AI can sustain performance and margins without reliance on proprietary American outputs. signal: Independence, not imitation, is the real test.
-
Whether open-weight models show comparable risk-adjusted performance across valuable workflows. signal: Autonomy must prove profitable, not just capable.
-
Whether hyperscalers can sustain free-cash-flow generation as model prices and architectures evolve. signal: Cash flow, not capex, is the scoreboard.
-
Whether AI adoption delivers measurable gains in margins, asset utilization and returns on invested capital. signal: Deployment must become measurable return.
The Investment Verdict
Ownership, not rankings
The U.S.–China AI competition is becoming a contest over the economic ownership of intelligence. The United States retains significant frontier advantages; China is narrowing capability gaps and expanding access. But the next market rotation will depend less on model rankings than on the distribution of profits across frontier development, infrastructure, software and industrial deployment.
Model capability, infrastructure ownership and shareholder returns must now be evaluated separately. The decisive question is whether leaders can preserve enough economic return to finance continued innovation while rival models grow progressively more capable and accessible.
America leads the creation of frontier intelligence; China is challenging the cost of reproducing it. The next AI Supercycle will reward those who control the economic value — not merely the technology — of intelligence.