Research paper · 23 August 2026

The Chip Bottleneck: Why Compute, Not Code, Controls the AI Frontier

Ideas in AI spread in weeks, but the chips needed to use them take years to make. Frontier AI rests on a supply chain held by a handful of companies and countries, and that turns compute into a source of geopolitical power.

By Aayan · 8 min read · 10 sources

Edited for publication and checked against the sources listed at the end.

The New Bottleneck

The central question in artificial intelligence is usually framed as one of capability: what can increasingly powerful models do? However, a more consequential question may be who can afford to build them.

Frontier AI depends on three broad inputs: algorithms, data and computation. Algorithms can diffuse rapidly through research papers and open-source implementations. Data is unevenly distributed, but can increasingly be licensed, generated synthetically or supplemented through new sources. Computation is different. Training frontier models requires enormous quantities of advanced accelerators, while producing those accelerators depends on a semiconductor supply chain concentrated in a small number of companies and countries [1].

This does not make computing the only scarce input in AI. It makes it the input whose supply is hardest to reproduce quickly. A new algorithm can be copied. Researchers can be trained or recruited. Data can sometimes be purchased or generated. A leading-edge semiconductor fabrication plant, by contrast, requires years, billions of dollars, specialised equipment and expertise accumulated over decades [2].

This distinction is turning computing into a geopolitical resource.

From Software to Infrastructure

The history of AI has encouraged the belief that technological advantage comes primarily from ideas. The Transformer architecture, introduced in 2017, spread rapidly because its underlying concepts could be studied and reproduced by researchers worldwide [3]. The physical infrastructure required to train the largest models cannot diffuse nearly as quickly.

A research team is able to reproduce a published architecture. It cannot reproduce an advanced GPU supply chain simply by reading Nvidia’s papers. Leading-edge accelerators require advanced fabrication, high-bandwidth memory, sophisticated packaging, enormous data centres and reliable electricity [4]. Each introduces another potential constraint.

This creates an asymmetry between knowledge and infrastructure. Knowledge is able to diffuse; infrastructure remains geographically concentrated.

The result is that talent alone does not guarantee frontier capability. A country or firm may have world-class researchers but still depend on foreign providers for the hardware on which its research runs. Conversely, possessing hardware without researchers does not automatically create an AI ecosystem. The important resource is therefore not simply chips, but the ability to reliably obtain large quantities of advanced infrastructure.

The Semiconductor Geography of AI

The physical geography of advanced semiconductors makes this dependence particularly significant.

Nvidia designs many of the accelerators used for frontier AI, but it does not manufacture them [4]. TSMC is central to producing leading-edge chips, while ASML supplies the extreme ultraviolet (EUV) lithography equipment required for the most advanced manufacturing processes [5]. Advanced memory and packaging add further dependencies [4].

The result is a chain, in which each link depends on the one after it:

  1. 1AI laboratoriesTrain and run the modelsOpenAI, Google DeepMind, DeepSeek
  2. 2AcceleratorsThe chips the models run onDesigned by Nvidia, with memory from SK Hynix, Micron and Samsung
  3. 3Advanced foundriesThe factories that make the chipsTSMC in Taiwan, Samsung in South Korea
  4. 4Specialised manufacturing equipmentThe machines the factories needEUV lithography from ASML in the Netherlands, the only maker
Figure 1. The AI supply chain. The further down the chain, the fewer the suppliers: in 2021 all of the world's most advanced chipmaking capacity was in Taiwan (92 per cent) and South Korea (8 per cent) [1], and ASML is the only company that makes EUV machines [5]. Nvidia's foundries and memory suppliers are named in its annual report [4].

The significance is not that these companies’ technology is permanently irreplaceable. It is that substitutes cannot be created quickly. This gives these companies real leverage: even if alternatives are technically possible, they may arrive only years after access to the existing supply chain is lost.

Intel and Samsung have invested enormous sums in advanced foundries, demonstrating both that competition is possible and that closing the technological gap is difficult. New fabs take years to construct and qualify [2]. This makes semiconductor capacity fundamentally different from software capacity: when demand for an algorithm increases, copies can be distributed almost instantly; when demand for leading-edge chips increases, additional supply cannot appear overnight.

This is the foundation of compute’s geopolitical significance.

When Compute Becomes Power: An Emerging Hierarchy

The United States has increasingly treated access to advanced computing hardware as a strategic issue. Export restrictions on advanced chips to China are based on a simple premise: limiting access to high-end computing can constrain a rival’s ability to develop advanced AI and other technologies [6].

Whether such controls will permanently preserve the American lead remains uncertain. They can also produce unintended consequences. Restrictions create incentives for Chinese firms to develop domestic alternatives, such as Huawei’s Ascend accelerators [7]. In this sense, export controls may accelerate China’s push to make its own chips, potentially increasing its technological independence over the long term. This could intensify competition by encouraging both sides to invest more heavily in competing technological ecosystems. The competition for computing power is therefore no longer confined to firms; it increasingly involves the world’s major powers and their strategic interests.

This dynamic reveals a broader consequence of semiconductor concentration: control over computing infrastructure creates a hierarchy of strategic power. Countries occupy different positions depending on whether they design AI systems, manufacture the infrastructure required to run them, or primarily consume and rent that infrastructure.

These positions overlap, but they do not provide equal strategic autonomy. The United States combines leading AI laboratories, chip designers, capital and cloud infrastructure, giving it influence across multiple layers of the AI stack. Taiwan occupies a critical position in advanced semiconductor manufacturing [1], while the Netherlands controls indispensable lithography technology through ASML [5]. China possesses substantial technological and industrial capacity but remains constrained by limited access to the most advanced foreign hardware [6]. India, despite significant technical talent and a growing AI ecosystem, remains more dependent on external computing infrastructure [8].

The significance of this hierarchy becomes clearer when compared with software. States cannot meaningfully restrict access to algorithms in the same way they can restrict access to physical computing infrastructure. A government can restrict a company’s access to advanced chips; it cannot prevent researchers around the world from reading a paper describing a new optimisation technique. Compute therefore gives states a form of leverage that software knowledge does not.

The emerging competition over AI is consequently not simply a race to develop better models. It is also a contest over who controls the physical infrastructure required to build and deploy them.

The Countries Caught in the Middle

The emerging AI hierarchy is not simply a contest between Washington and Beijing. It also creates a difficult position for countries that possess significant technical talent but limited control over the upstream semiconductor industry.

This creates a distinction between participating in AI and controlling its infrastructure. A country can rent GPUs, build applications, train researchers and develop valuable AI companies without controlling the factories that produce the hardware those companies require. This places such countries in a vulnerable middle position: they can participate in the AI economy, but remain dependent on foreign suppliers for the infrastructure needed to advance their capabilities.

That dependence matters during periods of scarcity. When computing power is abundant, renting it can appear equivalent to owning it. When supply becomes constrained, however, control of manufacturing capacity determines who receives priority and who must compete for access. Countries in this middle position therefore have limited strategic autonomy, as their ability to advance their AI capabilities depends on the states and companies that control critical resources.

India illustrates this dynamic. Despite substantial technical talent and a growing AI ecosystem, it relies on globally sourced GPUs and foreign semiconductor supply chains, from countries such as the United States and Taiwan, for much of the advanced computing infrastructure required to develop frontier AI. In May 2025 its national AI programme passed 34,000 GPUs of shared computing capacity, but every chip on offer was designed by an American company: Nvidia, AMD, Intel or Amazon Web Services [8].

For countries outside the semiconductor core, this creates a structural disadvantage: they can participate in the AI economy without having equal influence over its physical foundations. This dependence establishes the global AI hierarchy, in which countries that control critical infrastructure hold greater strategic influence than those that primarily rely on it.

The Strongest Challenge: What If Computing Gets Cheaper?

Yet the relationship between AI capability and computing power is not necessarily fixed. Improvements in efficiency can reduce the amount of hardware required to achieve a given level of capability, potentially weakening the role of compute as a bottleneck.

DeepSeek’s advances in training efficiency demonstrated precisely this possibility. Its V3 model needed only 2.788 million GPU hours on Nvidia H800 chips for its full training, yet its developers reported performance comparable to leading closed-source models [9]. If algorithmic improvements continue, smaller organisations could achieve capabilities that previously required enormous clusters. Open-weight models could further decentralise access by allowing organisations to build on existing systems rather than starting from zero.

But this does not necessarily eliminate computing as a strategic resource.

Efficiency changes the amount of computing power required to achieve a particular capability; it does not eliminate the advantage of having more of it. Because algorithmic improvements diffuse quickly, organisations with access to more hardware can also apply those efficiency gains. Moreover, cheaper computation can increase demand for AI, potentially offsetting some of the reduction in compute required per task [7].

The bottleneck therefore remains anchored to hardware. If one constraint becomes cheaper or more abundant, lower costs can increase demand for other inputs, potentially shifting the bottleneck elsewhere. Cheaper processors, for example, could increase demand for electricity, data centres, memory, networking, or the computing capacity required to serve models at scale. Electricity is already a concern: data centres’ demand for it is projected to more than double by 2030, with AI the biggest driver [10]. The specific constraint may change, but the underlying dependence on scarce physical infrastructure remains.

Conclusion

Compute is not the only determinant of AI power. Algorithms can dramatically improve efficiency, data remains valuable and researchers remain indispensable. But advanced compute possesses a unique strategic characteristic: its supply depends on a concentrated physical infrastructure that takes years and enormous capital to reproduce.

That makes semiconductor capacity more than an industrial asset. It is a source of geopolitical leverage.

The countries that control advanced AI infrastructure therefore occupy a different position from those that merely rent access to it. As AI becomes more capable and more economically important, that distinction is likely to become harder, not easier, to ignore.

The future AI race may ultimately be decided not only by who has the best models, but by who can secure the hardware required to build the next one.

Sources

← All papers