In the first essay of this series, I argued that AI has become an industrial race driven by predictable scaling laws. In the second, I argued that this race has entered a new phase: AI is now constrained by energy, not algorithms. But there is another hard limit — quieter, narrower, and far more fragile.
It is the global supply chain for advanced semiconductors. If energy is the fuel of AI, chips are the engine. And today that engine depends on a remarkably small number of companies, factories, and geographies. This is the global compute bottleneck — and it may ultimately determine how far and how fast the AI race can run.
A Thin Industrial Base
Training frontier AI models requires tens of thousands of the most advanced chips ever manufactured. These are not generic components. They rely on cutting-edge process nodes, advanced packaging, high-bandwidth memory, and extremely high manufacturing yields.
Despite the scale of global demand, only a handful of firms are capable of producing leading-edge AI accelerators at volume. At the center of this system sits Taiwan Semiconductor Manufacturing Company (TSMC). Today, most leading-edge AI accelerators, whether designed by NVIDIA, AMD, or hyperscalers, ultimately depend on TSMC’s most advanced manufacturing processes.
This is a choke point.
TSMC’s Singular Position
TSMC is uniquely positioned to manufacture advanced chips at scale with consistent yields. Other foundries are investing heavily, but matching TSMC’s combination of process maturity, reliability, and volume remains difficult.
This concentration introduces systemic risk. AI progress may be predictable in theory, but in practice it flows through a single industrial funnel.
That funnel is vulnerable to geopolitics, natural disasters, labor disruptions — and even basic resources. During drought periods in Taiwan, chipmakers including TSMC have relied on water-trucking and conservation measures to keep fabs running. The future of global AI capability now depends, in part, on whether enough ultra-pure water reaches a small number of facilities.
This is what industrial fragility looks like.
ASML: The Chokepoint Behind the Chokepoint
Even TSMC has a dependency. Advanced chips require extreme ultraviolet (EUV) lithography systems produced by a single supplier: ASML, based in the Netherlands.
EUV machines are among the most complex industrial tools ever built. They cost hundreds of millions of dollars, take years to manufacture, and have no commercial substitute. Without access to EUV, producing cutting-edge AI chips at scale becomes extraordinarily difficult.
This is why export controls on advanced lithography equipment have become one of the most consequential geopolitical levers of the decade. They do not merely slow competitors; they shape who can advance fastest at the technological frontier.
Why Capital Alone Cannot Fix the Bottleneck
It is tempting to assume that shortages can be solved with investment. If chips are scarce, build more fabs. If capacity is tight, spend more money.
But semiconductor manufacturing does not scale like software.
A leading-edge fab typically takes three to four years to build, followed by additional time to install tools and ramp production to high yields at the frontier. Skilled labor is limited. Equipment supply is constrained. Advanced packaging and high-bandwidth memory introduce additional bottlenecks.
Even companies with enormous capital reserves cannot simply buy their way out of these constraints. AI progress may follow predictable curves; the industrial systems behind it do not.
The Compute Divide
This bottleneck helps explain a widening divide in AI capability. Countries with strong universities, talented researchers, and abundant data still find themselves unable to train frontier models because they lack reliable access to advanced compute.
In the AI era, talent without compute is like oil expertise without drilling rigs. Capability depends not just on ideas, but on deep integration with industrial supply chains.
Why Diversification Matters
Against this backdrop, efforts to diversify advanced chip manufacturing take on strategic importance. Intel’s renewed push into foundry services, while still in progress, represents an attempt to add leading-edge capacity outside Taiwan. Even partial diversification reduces systemic risk.
This is why semiconductor policy has become inseparable from national security policy. The AI race is forcing governments to reconsider industrial strategies many assumed belonged to the past.
Fragility Is the Real Constraint
Scaling laws explain how to improve AI models. Energy policy determines how much compute can be powered. But supply chains decide whether either ambition can be executed.
Today’s AI ecosystem rests on a narrow industrial base: a dominant foundry, a single lithography supplier, and a small set of critical packaging and memory providers. That concentration, not intelligence, may be the binding constraint on future progress.
The first essay argued that AI has become an industrial race. The second showed that energy is the battlefield. This third makes the next point clear:
The fragility of the compute supply chain may ultimately limit how fast — and how far — AI can scale.
Intelligence may now be predictable. Manufacturing reality is still not.