April 15, 2026
The US dominates AI's IP-heavy layers. It's losing the ones that actually matter.
The conventional wisdom on the AI race goes something like this: the US leads on chips and models, China leads on manufacturing and energy, and export controls keep us ahead on the metrics that count. This framing is largely incorrect. It misunderstands what "winning" the AI race will actually require, and it dangerously misunderstands how we view AI superiority as a permanent advantage.
The AI Stack Is Wide
Talk about "AI leadership" in Washington and you'll hear about chips, models, and talent. These matter. But they're only the top layers of a much larger structure. The AI stack, defined properly, is everything that we need to unlock tangible security and economic impacts from AI. That includes the IP-heavy layers we're good at: semiconductor design, frontier model development, and high-quality data. But it also includes the foundational layers we're bad at: energy and infrastructure, critical mineral supply chains, and basic and advanced manufacturing.
On the IP-heavy layers, the US holds real advantages. Our highest-end AI model (GPT 5.4 Pro) averages 8% better on selected benchmarks than the highest-end Chinese model (Kimi K2.5). US data and benchmarks frequently appear in Chinese model development.
But on the foundational layers, the advantage inverts.
Energy: US total electrical generation capacity is roughly 1,190 GW. China's is almost triple that: 3,350 GW. In 2024, China added 500 GW of new generation capacity. We added 48.6 GW. They’ve already got more energy than we do, and they’re building more at an order of magnitude faster than us.
Supply chain: Only 12% of global rare earth production comes from the US, compared to 70% from China. Rare earths are a crucial input for AI chips.
Manufacturing: China exceeds the US by $2.5 trillion in total manufacturing value added. In advanced semiconductor fabrication specifically, China leads total output at the 7nm process node. We hold the ability to produce higher end 4nm wafers, which China has not demonstrated in high volumes, but our total expected 2026 production of leading-edge chips (<10nm) will make up only 10% of global advanced chip production. Not a decisive advantage.
This is the core problem. The US and China have models benchmarked within 6% of each other, despite our advantages in chips and talent. When AI begins to unlock real security and economic impact, the question will not be who has the highest benchmark scores. It will be which state can physicalize the impact of AI: manufacturing autonomous drone swarms, scaling autonomous systems at speed and low cost, and rapidly building and powering the data centers to deploy AI. That is a race we are not winning.
Why Tighter Export Controls Could Backfire
The instinct to choke China off from advanced chips is understandable, in a vacuum. America’s the only country that can design leading end-chips, and China’s the only country that can deploy them at scale. But the dynamics of Chinese industrial policy suggest it may be counterproductive.
For three decades, the CCP supported China's industrialization through a multi-level financing system: national policy banks, national funds, and budgetary support, augmented by aggressive investment at the provincial and city levels. The sheer volume of capital deployed drove enormous total gains.
China’s changed their tune. Now, they’ve discovered that debt from local governments makes up massive amounts of their GDP, and tighter rules on borrowing have limited how aggressively the government can direct capital into the industrial base. In semiconductors, while China has increased nominal investment through vehicles like Big Fund III (344 billion yuan), state and city-directed funding has fallen, and market deal flow has dropped. New vehicles like the government venture capital guidance fund are increasingly structured like private-sector VC funds: they have to make market-based decisions, must meet return expectations, and cannot simply subsidize development with cheap loans or grants the way previous efforts did.
This is a deliberate Chinese move prioritizing long-term economic stability over brute-force industrial subsidy. Chinese entities still access some US chips through gray-market channels, which appears to slow rather than halt Chinese AI progress.
Here's the risk: a full cutoff from American chip production would give China a stronger incentive to reverse that fiscal discipline and re-accelerate domestic semiconductor investment. AI development is a stated national priority. Chinese leadership has signaled that they expect AI to generate large economic impacts. And past volatility in US export controls has already pushed China to increase domestic production capacity. A full cutoff could be the catalyst that forces China back to the all-of-the-above model in the one sector where we can least afford it.
Build the Stack, Then Tighten the Vise
The most important thing the US can do right now is not cut off China from buying any chips. It's building stack sovereignty: an AI stack that is resilient to external disruption and firmly controlled by the United States.
Why sovereignty matters: it's easier to cheat on the IP-heavy layers of the stack. Hardware is smuggled. Chip designs are stolen and used to inform Chinese development. American models face cyberattacks targeting trade secrets; frontier labs are infiltrated by Chinese actors. Our data and benchmarks are freely used to develop Chinese models. But foundational layers are harder to replicate, even without regulatory barriers: they're limited by buildout time and capital expenditure. You can only build so quickly, and there's no cheating capital requirements like you can with IP theft.
Stack sovereignty requires three lines of effort.
First, remove regulatory barriers. The Administration has active efforts on permitting reform, energy buildout, and manufacturing support. The pace must accelerate. Congress should pass the SPEED Act and the Council on Environmental Quality should continue their work on supporting permitting reform, as directed by President Trump’s Executive Order, “Accelerating Federal Permitting of Data Center Infrastructure”. The Administration should explore streamlined federal pathways for critical mineral project permitting, including accelerated Bureau of Land Management claim processing, providing clearance for mining projects on DOW, DOE, and DOI lands, and expanded FAST-41 coverage for AI-stack-relevant projects.
Second, partner with allies on supply chain resilience. Pax Silica was a strong first step: leveraging the AI stack's strategic importance to strengthen alliances while building resilience. The State Department should continue to pursue bilateral critical minerals deals that develop local infrastructure and jobs for partner nations alongside structured US off-take, financed through the Export-Import Bank and DFC. Secretary Rubio’s leadership in the Critical Minerals Ministerial and the creation of the Forum on Resource Geostrategic Engagement is crucially important in advancing American supply chain resilience and AI stack sovereignty.
Third, deploy capital across the stack. Through the US Investment Accelerator, Pax Silica Fund, and the Department of War’s Office of Strategic Capital, the government can identify investment partners and deploy capital through direct investment, subsidization, and tax incentives in the areas that matter most. Government funding should be used sparingly; private capital is better suited for effective development. The American AI Exports Program and deregulatory efforts can attract private capital at scale.
Finally, prevent China from beating American chip companies. America needs to maintain our advantage in chip production: this means leveraging export controls on semiconductor manufacturing equipment (SME) to prevent China from building the capacity to compete with American champions. Selling chips to China isn’t the lose condition. Selling them the tools to ignore America is.
The US does not win the AI race by having the best model on a benchmark. It wins by having the stack to turn AI into power: manufactured, deployed, and sustained at scale. Right now, we're only optimizing the advanced layers. Time to build the foundation.
.png)
