Redistributed

General

The Hardware Bottleneck

Published March 8, 2026

infrastructureaccesssemiconductorsconcentration

The AI industry runs on software that can be copied infinitely at zero marginal cost. But it depends on hardware that cannot. The chips that train and run AI models flow through supply chains that are fragile, concentrated, and increasingly contested. Understanding these chokepoints is essential to understanding who controls AI.

The Physical Foundation

Training a large language model is not a software problem — it's an energy problem. GPT-4-class models require thousands of GPUs running for months, consuming megawatts of power continuously. The compute requirements have grown roughly 10x per year since 2010, far outpacing Moore's Law (Epoch AI, 2024).

This isn't abstract. It means physical chips, manufactured in physical factories, from physical materials. As Kate Crawford documents in Atlas of AI, the "cloud" is a misleading metaphor — there is no cloud, only other people's computers, sitting in data centers that consume electricity, water for cooling, and rare earth elements refined through extractive processes (Crawford, 2021).

The physical foundation of AI creates a hierarchy of access. At the top: companies that can afford to buy or rent thousands of the latest GPUs. In the middle: organizations that can access last-generation hardware or cloud credits. At the bottom: everyone else.

Nvidia's Dominance

One company controls the critical path: Nvidia.

In 2024, Nvidia held roughly 80% of the data center GPU market (SemiAnalysis, 2024). But market share understates their position. For training frontier models, Nvidia's share is closer to 95%. Their H100 and subsequent chips are not merely better than alternatives — for many AI workloads, they are effectively the only option.

This dominance rests on two foundations.

Hardware design. Nvidia's GPUs are architecturally suited to the matrix multiplications that neural networks require. They invested in this capability for years before AI became lucrative, initially for graphics rendering. That head start compounded.

CUDA lock-in. More significant than the hardware is the software ecosystem. CUDA, Nvidia's proprietary programming framework, has become the de facto standard for GPU computing. As Chris Miller documents in Chip War, this creates a moat: millions of lines of code written for CUDA don't run on competitors' chips (Miller, 2022).

Developers learn CUDA. Frameworks like PyTorch and TensorFlow optimize for CUDA. Research papers assume CUDA. Switching costs accumulate at every layer.

AMD offers competitive hardware (their MI300 series) but struggles against CUDA lock-in. Intel has spent billions on data center GPUs with limited success. Google's TPUs are powerful but only available on Google Cloud — they're not for sale. The competitive landscape shows that hardware alone is insufficient; the ecosystem matters as much as the chip.

TSMC: The Single Point of Failure

Every advanced AI chip — Nvidia's GPUs, AMD's competitors, Apple's M-series, Google's TPUs — is manufactured by one company: Taiwan Semiconductor Manufacturing Company (TSMC).

TSMC fabricates over 90% of the world's most advanced semiconductors, the chips built on process nodes below 7 nanometers (CSET, 2021). Their facilities in Taiwan represent the most sophisticated manufacturing capability on Earth. No other company can match their yield rates at these scales.

This concentration is remarkable. The AI industry's trillion-dollar valuations rest on factories in a single geography, operated by a single company, on an island that exists in a state of perpetual geopolitical tension with a nuclear power that claims sovereignty over it.

The implications extend beyond AI. Advanced chips go into phones, servers, medical devices, weapons systems. The global economy's most critical hardware flows through TSMC's fabs.

Why is it this concentrated? Semiconductor manufacturing exhibits extreme economies of scale. Building a leading-edge fab costs $20 billion or more and takes years. Staying at the frontier requires continuous investment. Intel, once dominant, fell behind and has struggled to catch up. Samsung competes but trails TSMC on yield and reliability.

ASML, a Dutch company, provides another bottleneck. Their extreme ultraviolet (EUV) lithography machines — which cost $150 million each and require a cargo plane to ship — are the only equipment capable of printing features at the scales required for cutting-edge chips. ASML has no competitors for EUV.

The Mineral Layer

Below the fabs sit the mines.

Advanced semiconductors require rare earth elements: neodymium for magnets, gallium for substrates, and dozens of others in trace amounts. China dominates this layer — controlling roughly 60% of rare earth mining and 90% of rare earth processing (IEA, 2024).

This isn't geology; it's policy. The minerals exist elsewhere. But China invested heavily in refining capacity while other countries didn't. Building alternative supply chains requires years and billions in investment.

The batteries powering AI data centers and the chips inside them both depend on materials extracted from a handful of countries, often under conditions Crawford documents as extractive: environmental damage, labor exploitation, and minimal benefit to local communities (Crawford, 2021).

The Chip War

The semiconductor supply chain has become a theater of great power competition.

In October 2022, the US imposed export controls blocking China's access to advanced chips, chip-making equipment, and the expertise to build them (Bureau of Industry and Security, 2022). The rules extend to Nvidia's latest GPUs, ASML's EUV machines, and any company using US technology to serve Chinese customers.

The logic: AI is dual-use technology. Advanced chips train AI systems that have military applications — surveillance, autonomous weapons, signals intelligence. The US aims to maintain a generational lead in AI capabilities by controlling the hardware layer.

China has responded with massive investment in domestic semiconductor capacity. Success is uncertain. As Miller argues, chipmaking is the most complicated manufacturing process humans have ever devised. It cannot be easily replicated through industrial policy alone (Miller, 2022).

The CHIPS and Science Act (2022) provides $52 billion for US semiconductor manufacturing, attempting to bring fabrication back onshore. TSMC is building fabs in Arizona. Samsung is building in Texas. Intel is expanding in Ohio. Whether these facilities can match Asian fabs' efficiency remains to be seen.

For AI, this creates uncertainty. A Taiwan contingency would disrupt global chip supply for years. Export controls limit where AI capabilities can be deployed. The "open" in open-weight models means nothing if you can't get the chips to run them.

What This Means for AI Access

The hardware bottleneck shapes AI access in several ways.

Training frontier models requires frontier hardware. The compute requirements for GPT-4-class models effectively limit training to organizations that can access tens of thousands of GPUs. That's a handful of well-funded labs and the cloud providers they partner with. Open-weight releases let others use frontier models, but they don't let others create them.

Inference costs determine deployment economics. Running AI models requires hardware too — less than training, but still substantial. The cost per query shapes which applications are viable. Healthcare diagnostics in wealthy countries: economically viable. Agricultural advice in low-income regions: harder to justify commercially.

Hardware access is geographically uneven. Export controls, data center locations, and cloud provider coverage create a map of AI access that mirrors existing global inequalities. As Birhane documents, data often flows from the Global South while compute concentrates in the North (Birhane, 2020).

The stack is vertically integrated. Cloud providers are investing in AI labs (Microsoft in OpenAI, Google in DeepMind, Amazon in Anthropic). The same companies sell cloud compute and develop models. Competition policy designed for horizontal markets struggles with this integration.

Alternative Futures

This concentration is not inevitable. Different policy choices could distribute hardware access more broadly.

Public compute infrastructure. The EU's EuroHPC initiative provides supercomputing resources to European researchers. Similar facilities for AI could provide compute to academic labs, civil society organizations, and small companies that can't afford cloud rates. The question is whether public investment can keep pace with exponential growth in AI compute requirements.

Alternative hardware ecosystems. Companies like Groq, Cerebras, and Graphcore are building specialized AI chips with different architectures than Nvidia's. Apple Silicon has made local inference viable on consumer laptops. If these alternatives mature, CUDA lock-in could weaken. The precedent is encouraging: Intel's x86 monopoly in PCs seemed unassailable until ARM-based chips (via Apple Silicon and Qualcomm) carved out a significant share within a few years. Ecosystem moats are real, but they are not permanent — especially when new use cases (like on-device AI) favour different architectural trade-offs.

Algorithmic efficiency. Research into more compute-efficient training and inference could reduce hardware requirements. Techniques like quantization, distillation, and sparse architectures have already enabled models to run on smaller chips (Dettmers et al., 2022). If efficiency improves faster than model size grows, the hardware bottleneck loosens.

Supply chain diversification. New fabs in the US, Europe, and Japan could reduce dependence on Taiwan. New mineral processing capacity outside China could diversify critical inputs. These investments are underway but will take a decade to mature.

The Redistribution Question

Hardware is the layer where software "openness" meets physical reality.

You can release model weights under a permissive license. You can open-source training code. You can publish papers with every implementation detail. None of that matters if the chips required to run the model are controlled by a handful of companies, manufactured in a single country, and allocated by whoever can pay the most.

As Langdon Winner asked: do artifacts have politics? (Winner, 1980). The semiconductor supply chain answers emphatically: yes. The political choices embedded in chip design, manufacturing location, and export policy shape who can participate in AI development.

The redistribution question for hardware is: who should have access to the physical substrate of AI, and how? Market allocation gives access to whoever can pay. Public infrastructure provides access based on policy priorities. Export controls allocate access based on geopolitics.

Right now, all three mechanisms concentrate access — markets favor the wealthy, public investment is nascent, and export controls draw lines between allies and adversaries. A different configuration could distribute access more broadly. But it would require treating semiconductors as infrastructure with public interest obligations, not just commercial products.

The chips that run AI are manufactured in fabs that cost billions, from materials that are geographically concentrated, using equipment that only one company can make. That supply chain is now a site of great power competition. The hardware bottleneck is not a temporary constraint that software can route around. It is a structural feature of how AI works — and who gets to use it.

Sources

  • Birhane, Abeba. "Algorithmic Colonization of Africa." SCRIPTed, 2020. SCRIPTed
  • Bureau of Industry and Security. "Implementation of Additional Export Controls on Advanced Computing and Semiconductor Manufacturing Items." Federal Register, 2022. Federal Register
  • Crawford, Kate. Atlas of AI. Yale University Press, 2021. Publisher
  • CSET (Center for Security and Emerging Technology). "The Semiconductor Supply Chain." Khan, Saif M., 2021. CSET
  • Dettmers, Tim et al. "LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale." NeurIPS, 2022. arXiv
  • Epoch AI. "Trends in Machine Learning Hardware." 2024. Epoch
  • IEA (International Energy Agency). "Critical Minerals Market Review." 2024. IEA
  • Miller, Chris. Chip War: The Fight for the World's Most Critical Technology. Scribner, 2022. Publisher
  • SemiAnalysis. "AI Datacenter GPU Market Share." 2024. SemiAnalysis
  • Winner, Langdon. "Do Artifacts Have Politics?" Daedalus, 1980. JSTOR
← Back to all articles
On this page