# The Constraint Money Can't Buy

For two years, the question hanging over AI was whether the world could make enough chips. That question has been answered. It was also the wrong question. The chip was never going to be the binding constraint for long. Fabs are the most complex facilities humanity builds, but they are buildable — with enough capital, enough time, and enough will. And this industry has all three right now. What's happened instead is the thing anyone who has run a factory would have predicted. The bottleneck moved. Then it moved again. And each time it moved, it landed somewhere less responsive to money than the place it left. That trajectory — not any single shortage — is the story worth understanding. ## You don't remove a bottleneck. You relocate it. Every system has one constraint that sets the pace for the whole thing. Relieve it and output rises — right up until the next constraint binds. You haven't eliminated the limit. You've promoted the next one. That's not a theory. It's what the last twenty-four months look like when you line them up. **Advanced packaging** was the first hard stop. The step that assembles logic and memory into a working AI accelerator was sold out for years. So TSMC poured concrete at it — new packaging fabs, tens of billions of dollars, capital expenditure guidance raised mid-year to $60–64 billion and another $100 billion committed to Arizona. Money worked. Not instantly, but it worked. **Memory** was next. HBM demand pulled capacity away from everything else, and the shortage cascaded down through the market — conventional DRAM, then display drivers and power management parts, the cheapest silicon on the board. By this month it had reached the finished product: reports of a completed GPU held back from launch because the memory it bolts onto costs too much. Money is working here too. Capacity is being built. But slower — and the arithmetic underneath is unforgiving. Wafer supply for 2027 grows around 12%, roughly half of which HBM consumes. That leaves conventional DRAM bit growth near 15% against demand growing 22%. Seven points short, in a year the industry's own executives are describing as the worst on record from a supply perspective. **Silicon wafers** came after that — the raw material itself. When Micron signs a ten-year agreement with a wafer supplier, backed by half a billion dollars in support, that's a company reaching further upstream than the shortage has yet traveled and locking the input before it binds. So far, so ordinary. Each of these constraints has a price. You can spend your way through them, given time. Then the ladder changes character. ## The first hard stop Every leading-edge chip in the world is printed by a machine that comes from exactly one company. ASML's capacity for its workhorse EUV systems is roughly 65 units this year, rising perhaps 30% next year, with a further increase still under evaluation for 2028. Sixty-five machines. That number governs the ceiling of the entire leading edge. There is no second source. There is no substitute. The optics come from essentially one supplier, the build takes quarters, and the engineers who install and maintain them take years to train. When ASML says it is *investigating* an expansion for 2028, that's the honest word — adding capacity to a build rate like this isn't a decision, it's a program. This is the first constraint on the list where more money does not produce more units on any timescale that matters. It's also the point where the character of the problem changes from financial to physical. Below it, things get harder still. ## Power, water, permission, people **Power.** Korea's Yongin cluster alone needs something on the order of 15 gigawatts — roughly ten reactors' worth. Its southwestern cluster and planned AI data centers need tens of gigawatts more. In Taiwan, semiconductors already consume around 17% of national electricity, on a path toward 30% by 2035. A fab takes three to four years to build. A nuclear plant takes nine or ten. Transmission lines, substations and industrial water systems can take more than a decade just to break ground. **Water.** Unlike carbon, water is not fungible. A ton of carbon emitted anywhere has the same effect anywhere, which is what makes global net accounting coherent. A stressed aquifer in Arizona is not replenished by irrigation efficiency in Iowa. Water binds one catchment at a time — which means the decision that matters isn't the measurement framework, it's the siting. And siting optimizes for power, land, fiber and tax. Catchment stress usually gets assessed after the site has effectively been chosen. **Permission.** New York became the first US state to pause new large data centers, freezing state environmental permits for facilities drawing 50 megawatts or more. Australia has moved toward requiring data centers to generate as much power as they consume. Neither of these is a grid running out. It's the *permission* running out — and polling suggests it isn't a passing mood. **People.** Underneath all of it: engineers. Every fab on earth is short of them. India's national semiconductor program funds training for tens of thousands of students, which is the one input capital cannot compress. You don't buy a trained workforce. You grow one, and it takes years. ## The pattern Read the ladder from top to bottom and one thing is consistent: **each rung responds less to money than the one above it.** Packaging responded to capital. Memory responds slowly. EUV machines don't respond at all on a useful timescale. Grids, catchments, permits and people respond to something else entirely — political will, physical time, and public consent. This matters because capital is the one input nobody in this industry lacks. Every major player is raising or spending at record scale simultaneously. An input nobody is short of cannot be a source of advantage. Advantage lives in scarcity — and the scarce things are now sitting outside the semiconductor industry altogether. ## The part that isn't physics Here's the uncomfortable corollary. The bottom of that ladder — permits, consent, environmental review — isn't a law of nature. It's a choice. Concrete cures at the same rate everywhere. Consent doesn't. China has built more than 8,000 miles of ultra-high-voltage transmission in the 2020s. The United States has built 375. In China, when an industrial zone is designated, the power to serve it is planned and built alongside it. In the US, a project joins an interconnection queue measured in years. California authorized a high-speed rail line in 2008. Eighteen years later, the projected cost is $231 billion, no track has been laid, and the first segment — between two Central Valley cities, not the route anyone actually needs — has slipped to 2032. Over the same period, China built the largest high-speed rail network in the world. It would be easy, and wrong, to read that as an argument for authoritarian efficiency. The bill comes due on both sides. China's local governments carry an estimated $8–11 trillion in hidden debt, much of it from precisely this kind of build-first infrastructure, and a meaningful share of that concrete became stranded assets — built fast, then left to sit. Meanwhile the friction the West pays for in years is not pure waste. It's the price of not having a data center dropped on your aquifer, or a transmission line run through your land without recourse. That delay is a real thing being purchased. Both systems pay. One builds the wrong things quickly and struggles to stop. The other builds the right things slowly, and sometimes not in time. In a chip-constrained world, none of this decided anything. In a power-constrained world, it decides a great deal. ## The fab is now the fast part The final irony is the one an operations person feels hardest. The semiconductor industry got extraordinarily good at the step it controls. Samsung has pulled a completion date forward by seven years. SK Hynix by twelve. Fabs go up faster than they ever have. And it may not matter, because speeding up a non-constraint adds nothing to throughput. It just reveals whatever was standing behind it — a substation, a water treatment plant, a transmission right-of-way, a county council. For fifty years the fab was the long pole. Everything waited on it. The industry optimized it so relentlessly that it stopped being the limit at all. Which leaves the question that actually governs the next decade. It isn't whether we can build the best chip in the world. We can, and we do. It's whether we can plug it in. --- *An operations perspective on the AI buildout — what it can manufacture, and what it can't.*