Middle Powers Face AI Access Test As U.S. And China Dominate Compute

Compute Control Becomes A Geopolitical Gate

AI access is becoming a diplomatic and infrastructure problem for countries outside the U.S. and China.

At a New York Tech Week discussion titled “The Great AI Divide,” Sam Winter-Levy of the Carnegie Endowment for International Peace described AI as mainly a two-country race, with the U.S. and China controlling 90% of global computing power and attracting between 70% and 80% of global AI investment.

That concentration changes the policy question for middle powers.

The issue is not only whether they can build local AI products, but whether they can keep reliable access to models, chips, data centers and research talent when frontier systems are rationed by companies or shaped by government export and access decisions.

Managed Access Narrows The Field

Winter-Levy pointed to three pressures that make dependence more visible: managed access to frontier models, compute constraints, and a more assertive role for both Washington and Beijing.

In his framing, U.S. companies can choose the first customers for advanced systems, while limited compute capacity forces providers to decide who receives access first.

That creates a practical risk for countries that want AI capacity but do not control the full stack.

If a state depends on a small group of foreign model providers, its access can be delayed, narrowed or tied to strategic conditions.

The same concern applies to companies in markets that need AI for finance, scientific research, cybersecurity or public-sector services.

Bias And Representation Stay In The System

Aditya Vashistha of Cornell University shifted the debate from access to representation.

He said many AI technologies are built around WEIRD societies — Western, educated, industrialized, rich and democratic — which he estimated at 14% or 15% of the world’s population.

That leaves the remaining 85% at risk of being treated as an afterthought in data, benchmarks and product design.

The concern is not limited to language coverage.

Vashistha described religious, linguistic and identity biases that have improved since early generative AI systems but remain embedded in models.

He also noted that many safety benchmarks do not account for ableism, even as the world has 1 billion people with disabilities.

Sovereign AI Still Depends On External Inputs

The discussion treated sovereign AI as a partial answer, not a complete escape.

Winter-Levy said the UAE and India talk often about building their own models, but the effort remains difficult because local AI programs can still rely on U.S.-designed Nvidia chips, foreign-serviced data centers and supply chains controlled elsewhere.

Another path is a coalition of middle powers that pools models, data centers and technical resources.

A third option is to align closely with either the U.S. or China to secure access.

Each path carries a trade-off: independence can be expensive, pooling requires coordination, and alignment gives great powers leverage if policy disagreements arise.

Leverage Comes From Supply Chains, Data And Energy

The most concrete bargaining tools named in the discussion were researchers, energy, chips and data.

Winter-Levy said the Netherlands, Taiwan, Japan and South Korea hold leverage through semiconductor supply chains.

India was described as a major source of data, while Ukraine’s battlefield data and energy resources in other countries were also framed as assets that can matter in AI negotiations.

For Gulf and Asian policymakers, that makes AI strategy less about slogans and more about what can be exchanged for durable access.

Countries with energy, capital, data, chip roles or deployment markets can use those assets to seek stronger guarantees from model providers and governments.

The Watchpoint Is Durable Access

The next test is whether middle powers can turn local advantages into enforceable AI access rather than one-off partnerships.

The panel did not prove that sovereign models or open-source systems will be enough for national-security, finance, research or cybersecurity use cases.

It did define the constraint clearly: access to advanced AI is not guaranteed.

Countries outside the two dominant AI powers will need to decide which assets they can bargain with, which systems they can build locally, and where dependence on U.S. or Chinese technology is an acceptable risk.