The Split Just Changed Shape

“America is building the most powerful AI system. China is trying to make AI cheaper, more portable and harder to contain.”

Washington still controls the frontier. Beijing is learning to route around it. There is a consequential inflection.

The core thesis is strengthening, but in a more important way than a simple benchmark race. The United States still holds the superior frontier stack. China is increasingly attacking the economics, software dependency and geographic enforceability of that advantage.

The Big Take

Three developments now fit together

THE U.S. LEADS IN FRONTIER CAPABILITY. CHINA IS ATTACKING THE COST OF ACCESS, DEPLOYMENT AND DIFFUSION.

The Tencent structure matters because it suggests export controls may constrain ownership of advanced chips inside China more effectively than they constrain access to frontier compute abroad. If that model scales, Chinese firms can rent U.S.-class compute offshore for training while deploying domestically on increasingly capable Huawei infrastructure. The DeepSeek–Huawei collaboration attacks the second moat: CUDA lock-in. The hardware gap remains meaningful, but China is trying to reduce the software penalty that makes non-Nvidia accelerators difficult to use at scale. Meanwhile, Google's Gemini 4 "Argon," unveiled September 30, reinforces that the American frontier remains very strong: Google described it as comparable with OpenAI's Astra and Anthropic's Opus on key coding and cyber benchmarks, even though it still trails them on some measures.

The Hidden Pattern

China does not need to beat the U.S. chip-for-chip

What This Means for the AI Supercycle

Capacity gives way to efficiency

Five Signals to Watch Next

What moves the balance

Whether Washington closes the offshore-cloud route. A regulatory move against remote access to advanced accelerators would directly attack the Tencent model. signal: Watch for rules targeting remote access to advanced accelerators.
If other Chinese labs move onto Ascend through the new open software stack, CUDA dependence begins to weaken structurally. signal: Broad Ascend adoption would erode CUDA lock-in.
If Hunyuan improves materially after access to offshore frontier compute, this becomes evidence that cloud geography can partially bypass hardware geography. signal: Hunyuan gains would prove cloud geography beats hardware geography.
If model prices continue falling faster than U.S. prices, China keeps pressure on industry-wide AI rents. signal: Falling Chinese inference prices squeeze global AI rents.
Today's Reuters numbers make this the decisive financial metric of the next phase. signal: Revenue-to-capital is now the phase's defining ratio.

Key takeaways

The Velocity Insight

The U.S.–China AI race is no longer simply about denying China the best chips.
China is building a three-part workaround: rent the frontier, rebuild the stack, undercut the economics.
The United States still leads the frontier — but China increasingly does not need to own it to challenge its economics.
The decisive metric of the next phase is U.S. AI revenue versus AI capital employed.

Francesco De Leo Kaufmann · The Velocity Edge