In Scaling the Wrong Layer, I argued that the market is conflating two types of efficiency: hardware efficiency, which comes from building bigger infrastructure, and structural efficiency, which comes from better orchestration, routing, and architecture. I argued that if structural efficiency arrives first, significant portions of the current infrastructure buildout become mispriced.
There is a country-scale experiment running right now that tests that thesis in real time.
The United States spent three years designing export controls to restrict China’s access to advanced AI chips. The logic was straightforward… if intelligence is a function of compute, and compute requires GPUs, then restricting GPU access restricts AI capability. Control the silicon, control the outcome.
The policy assumed the hardware-efficiency frame. It assumed the only path to competitive AI ran through scale.
It may have trained the wrong model.
The Constraint
The numbers are stark. The best U.S. AI chips are currently roughly five times more powerful than Huawei’s best offerings1. By 2027, that gap is projected to widen to seventeen times.2
Huawei’s next-generation chip in 2026 may actually be less powerful than its current one, SMIC appears to be hitting a ceiling at 7nm.3 By raw compute, the policy is working exactly as designed.
But capability is not the same as compute. And competitive AI turns out to require less of the latter than anyone in Washington assumed.
What Actually Happened
Constrained on hardware, Chinese labs did what constrained competitors always do. They optimized on a different axis.
DeepSeek’s V3 has 671 billion parameters but activates only 37 billion per query through a Mixture-of-Experts architecture. StepFun’s Step 3.5 Flash activates 11 billion of its 196 billion parameters per token. Zhipu AI trained GLM-5 entirely on 100,000 domestic Huawei Ascend chips, hardware that would have seemed inadequate for frontier training a year ago.
These are not workarounds. They are architectural decisions that produce competitive output at a fraction of the compute cost. The constraint didn’t slow them down. It pushed them toward the very approach that Scaling the Wrong Layer argues is the higher-leverage path… structural efficiency over hardware scale.
DeepSeek is the name most people know. But DeepSeek isn’t an outlier, it’s a signal within a pattern. Alibaba’s Qwen family has overtaken Meta’s Llama in cumulative downloads on Hugging Face. Chinese open models on OpenRouter have gone from nearly zero in late 2024 to roughly 30% of usage.4 Martin Casado at Andreessen Horowitz estimates that among startups pitching with open-source stacks, there’s about an 80% chance they’re running on Chinese open models.5
The constrained ecosystem isn’t falling behind. It’s setting the efficiency frontier.
The Christensen Frame
This pattern has a name. Clayton Christensen called it disruption, and the version that matters here isn’t the overused Silicon Valley kind. It’s the original mechanism.
The incumbent competes on the performance axis it already leads. It invests in sustaining that lead, more compute, bigger clusters, faster chips. This is rational. It’s where the margins are. It’s where the customers are. It’s where the capital is flowing.
The constrained entrant can’t compete on that axis. It doesn’t have the hardware. So it competes on a different one, efficiency, architecture, distribution. The models are smaller. The inference is cheaper. The weights are open. None of this looks threatening from the incumbent’s position, because the incumbent isn’t measuring along that axis.
The disruption doesn’t happen when the entrant matches the incumbent on performance. It happens when the market discovers that performance was overweighted and efficiency was underpriced all along.
That repricing is already underway.
The Policy Paradox
Here is the irony that Washington hasn’t processed yet.
Export controls were designed to prevent China from building competitive AI by restricting access to the hardware the U.S. believed was prerequisite. The policy assumed the Build → Efficiency → Affordability loop, that you need scale before you get capability.
Instead, the restriction forced Chinese labs into the other loop…
Structural Efficiency → Reduced Demand → Selective Buildout
They couldn’t brute-force their way to capability, so they engineered around the constraint. They built Mixture-of-Experts architectures that activate a fraction of their parameters. They trained on inferior chips and published the techniques. They open-sourced the weights and let the ecosystem compound the gains.
The policy designed to maintain an American advantage in hardware scaling may have accelerated China’s advantage in the one area that makes hardware scaling less necessary.
This is not a failure of execution. The controls are working exactly as intended on the axis they were designed to constrain. The problem is that the axis they were designed to constrain may not be the one that determines the outcome.
The Wall
There is a version of this where architectural efficiency hits a ceiling. Where the next real leap in capability, not incremental improvement but a phase change, genuinely requires trillion-parameter models trained on hundred-billion-dollar clusters. Where the scaling laws hold indefinitely and hardware primacy reasserts itself.
This is possible. Maybe likely, on a long enough timeline.
But capital cycles don’t operate on long enough timelines. They operate on depreciation schedules, earnings calls, and financing assumptions. If structural efficiency buys the constrained ecosystem three to five years of competitive parity, three to five years of developer adoption, open-weight distribution, and architectural compounding, then by the time the wall arrives, the market has already repriced. The hardware advantage doesn’t disappear. But the window in which it was decisive may have already closed.
The bet isn’t that architectural efficiency wins forever. It’s that it wins long enough to change the board.
What This Means for the Fork
In Scaling the Wrong Layer, the fork was a market question, where should capital flow? The China experiment adds a geopolitical dimension to the same fork.
If hardware efficiency is the dominant path, export controls are sound strategy. Restricting access to the best chips restricts access to the best AI. The buildout is justified. The capital allocation is correct.
If structural efficiency is the dominant path, or even a co-dominant one, then export controls are an accelerant for the competition they were supposed to suppress. Every restriction that forces a lab to find an architectural workaround is a restriction that trains the ecosystem to need less hardware. And an ecosystem that needs less hardware is an ecosystem that your hardware advantage can’t control.
The $650 billion dollars in Western AI infrastructure spending is a bet on hardware primacy. Export controls are a bet on hardware primacy. The IPO trajectories of OpenAI and Anthropic assume hardware primacy.
China’s AI ecosystem is producing competitive models at a fraction of the compute cost, publishing the architectures openly, and gaining global distribution through open weights.
One of these trajectories is mispriced.
The Uncomfortable Question
The debate in Washington is still about whether to tighten or loosen chip restrictions. The AI OVERWATCH Act6 would give Congress veto power over export licenses. The Trump administration has oscillated between allowing H200 sales and restricting them. The assumption on both sides is that chips are the lever.
But if the lever is architecture, if the thing that determines AI competitiveness is not how many GPUs you can rack but how efficiently you can route a query, then the entire export control framework is optimizing for the wrong variable.
The question isn’t whether to sell China chips.
It’s whether chips are still what matters.