U.S. Leadership on AI Requires Strategic Interdependence

F or years, Washington has approached the U.S.–China artificial-intelligence competition largely through the lens of the technology stack. The logic is straightforward: Advanced AI requires advanced chips, enormous computing power, sophisticated semiconductor equipment, frontier models, and vast data centers. If the United States can control access to critical layers of that stack, it can preserve its technological lead while slowing China’s progress.

But a recent Forbes investigation by Anna Tong exposes the limits of that framework. She reports that American companies specializing in AI training data are doing substantial business with Chinese technology companies. These firms do much more than label photographs or clean datasets. They organize programmers, scientists, mathematicians, financial professionals, and other experts to construct difficult problems, devise answers, establish evaluation criteria, and create sophisticated datasets used to improve AI models after their initial training. Tong reports that China’s leading AI labs collectively spend roughly $500 million annually with American training-data companies. Some reportedly seek the same kinds of datasets that American frontier laboratories purchase.

This produces an extraordinary paradox. Washington may prevent a Chinese AI company from buying the most advanced Nvidia chip, while an American company can sell the knowledge that helps make its existing chips and models substantially more capable.

The problem is not simply a loophole in export controls. It reveals something more fundamental: The United States has been thinking too narrowly about what constitutes an AI strategic asset.

We need a concept broader than technology transfer. I propose calling it the AI knowledge pipeline: the organized process through which human expertise, scientific and technical knowledge, task design, evaluation methods, training data, and post-training know-how are converted into transferable inputs that increase the capabilities of artificial-intelligence systems.

Imagine American programmers solving difficult coding problems, mathematicians constructing reasoning exercises, scientists evaluating model responses, physicians distinguishing good answers from dangerous ones, and engineers designing sophisticated benchmarks. Their work can eventually become machine-readable training material. What has been transferred is therefore not merely “data.” Human knowledge has been distilled, structured, evaluated, and converted into an input for machine intelligence.

That distinction becomes increasingly important as AI advances. The first phase of the current AI revolution emphasized scale: more data, more GPUs, more parameters, more computing power. But frontier competition increasingly depends on what happens after massive pretraining — reasoning, reinforcement learning, synthetic data, expert evaluation, tool use, agentic behavior, and specialized domain knowledge. The scarce strategic resource is increasingly not simply computation. It is organized intelligence used to train artificial intelligence.

This leads to a larger conclusion about the U.S.–China AI competition. It is commonly described as a competition between two AI “stacks.” But the real competition is between two AI ecosystems.

The stack consists of identifiable technological layers: energy, semiconductor fabrication, chips, cloud infrastructure, foundation models, and applications. But an ecosystem encompasses much more: universities, researchers, engineers, entrepreneurs, venture capital, immigration, open-source communities, energy and manufacturing capacity, data markets, intellectual-property regimes, regulatory systems, alliances, international standards, and knowledge pipelines.

No single component determines the winner. China may trail America in frontier GPUs while possessing advantages in electricity generation, manufacturing scale, industrial deployment, engineering manpower, and its enormous domestic market. America may lead in frontier models while depending heavily on Taiwan for advanced semiconductor manufacturing. China may be denied the newest American chips but can compensate partially through efficiency improvements, domestic hardware, open-weight models, or American-generated training data. The final outcome will be the aggregate result of advantages and disadvantages across the entire ecosystem.

This perspective exposes the weakness of treating technology export controls as the centerpiece of an AI strategy. Export controls are necessary, particularly for technologies with direct military or intelligence applications. But controlling physical technologies while ignoring the knowledge, talent, capital, services, expertise, and networks that make those technologies useful is strategically incomplete.

At the same time, trying to control everything creates a different danger. America’s greatest technological advantage has historically arisen precisely because it is open. The United States attracts the world’s scientists, educates foreign students, operates globally, collaborates across borders, and spreads American technological standards through open-source software and widely used platforms.

This openness is not merely generosity. It is a source of American power. A world in which engineers in India, Europe, Southeast Asia — and, to some degree, China — build upon American technologies can become a world dependent upon an American-centered technological ecosystem. America gains not simply customers but gravitational pull.

The United States therefore faces two coherent but dangerous extremes. One is unrestricted openness. If China can systematically absorb American innovation while combining it with its enormous engineering workforce, manufacturing capacity, state resources, data, energy infrastructure, and domestic market, openness could supply precisely the missing ingredients China needs to close the gap and eventually surpass the United States. America would bear much of the cost of creating knowledge while its principal competitor rapidly internalized it.

The opposite extreme — comprehensive containment — carries a different but equally serious risk. If Washington attempts to sever China from American technology, research, capital, talent, models, and knowledge altogether, Beijing will have overwhelming incentives to eliminate its remaining dependencies and construct a parallel ecosystem from chips to models to standards. Rather than preserving an American-centered global system, Washington could accelerate the creation of a Chinese-centered alternative competing for the rest of the world. Either extreme, badly executed, can therefore end in the same place: an American defeat in the AI competition.

This is why the strategic choice cannot simply be reduced to engagement versus containment, or openness versus closure. What America needs is managed precision openness.

Think of the relationship not as a single switch marked OPEN and CLOSED, but as a control board containing dozens of graduated switches. Advanced GPUs might be nearly closed. Certain scientific exchanges might remain largely open. Some frontier-model capabilities could be restricted while open-source research remains accessible. Investment, cloud computing, talent flows, datasets, research collaboration, semiconductor equipment, AI models, and AI knowledge pipelines would each occupy different positions on the board — and those positions would change as technology, Chinese capabilities, and strategic circumstances evolve.

The governing question for each switch should be neither “Can we stop China from obtaining this?” nor “Will restricting this hurt American business?” It should be: What degree of openness in this particular area maximizes the long-term strength and gravitational pull of the American AI ecosystem relative to China’s?

Sometimes the answer will require closing a gate. Sometimes it will require opening one wider.

American openness itself can be strategically deployed. The objective should not simply be to keep China technologically behind. A more ambitious goal is to keep the American ecosystem so innovative, attractive, indispensable, and globally connected that China — and, even more importantly, the rest of the world — continues to have strong incentives to participate in it rather than migrate toward a competing Chinese system.

In that sense, interdependence need not always be a vulnerability. Properly structured, it can be leverage. If critical parts of global AI development continue to depend upon American technologies, research institutions, standards, platforms, talent networks, and knowledge pipelines, the United States possesses advantages that cannot be measured simply by comparing the number of GPUs on either side of the Pacific.

Managed precision openness will be extraordinarily difficult to execute. Policymakers must identify genuine chokepoints, distinguish them from commodities, anticipate substitution effects, coordinate with allies and private companies, and repeatedly recalibrate policy as technology changes. Yesterday’s critical technology can become tomorrow’s commodity; restricting one input can unexpectedly accelerate Chinese innovation in another.

But difficulty is not an argument against necessity. The stakes are too high for a strategy built primarily around a few easily identifiable technologies. Winning may require a level of state capacity, technological understanding, private-sector cooperation, and strategic sophistication that democratic governments have rarely had to sustain in peacetime.

The Forbes investigation should therefore prompt a debate larger than whether American data companies should sell particular datasets to Chinese laboratories. It should force Washington to ask a more fundamental question: What exactly is America trying to protect, and what exactly is it trying to lead?

If the answer is the world’s leading AI ecosystem, America must protect the technologies and knowledge pipelines that constitute genuine strategic advantages while remaining open enough that scientists, entrepreneurs, companies, allies — and, where strategically advantageous, even competitors — continue to orbit the American system.

The AI contest will not be won by building the highest wall around the best technology. It will be won by building the strongest ecosystem, and knowing, with extraordinary precision, where to build walls, where to build gates, and when, and how widely, to open them.