in short
Nvidia has locked up more than half of the world's advanced chip-packaging capacity through 2027. A war on the other side of the planet can still threaten a fab's supply of a gas most people have never heard of. The industry needs as many as 157,000 workers it does not currently have. And building the chips that run AI requires more pure water, per day, than entire cities consume, in some of the driest places on Earth. None of these problems can be solved by writing a bigger check. That's what makes them worth understanding.
Every few months, a new headline declares that the AI hardware shortage is about to end, a new fab breaks ground, a chipmaker announces a multi-billion-dollar expansion, a government unveils another subsidy package. And yet the shortage never quite ends. It moves. The place where AI's growth actually gets throttled has shifted at least four times in the last three years, from raw transistor supply, to advanced packaging, to memory, and now toward an even more fundamental set of constraints that no press release has fully reckoned with: the machines that make the machines, the chemicals those machines depend on, the people who run them, and the water they drink by the millions of gallons.
This migration pattern is itself the most important thing to understand about AI hardware supply, more important than any single statistic in this piece. Every time one constraint gets meaningfully addressed, a new fab comes online, a packaging facility expands, a memory maker breaks ground on new capacity, the bottleneck doesn't disappear. It reveals whatever was previously hiding one layer beneath it. Solve the transistor problem, and packaging becomes visible. Solve packaging, and memory becomes visible. Solve memory, and what's left is the set of constraints this piece covers: a single company's annual machine output, a handful of chemicals nobody diversified, a training pipeline that takes years, and a finite regional water supply. There is no reason to expect this pattern to stop once these four are addressed either, which is exactly the question the fifth companion piece linked below takes on directly.
This piece is a map of those four constraints, not because they're the only ones that matter, but because they are, right now, the ones capital cannot simply outrun. Understanding them is the difference between reading the AI buildout as an unstoppable exponential curve, and reading it correctly: as an exponential curve of demand colliding with a set of supply curves that move on the physical timescales of chemistry, construction, and human training, timescales that have not sped up, no matter how fast the demand curve has.
Bottleneck One: The One Machine on Earth That Can Print an Advanced Chip
Picture a chip's design as a blueprint that has to be printed at a scale smaller than a virus, features just a few nanometers wide, on a surface that will become a silicon wafer. The machine that does this printing uses an extraordinarily precise beam of light, called extreme ultraviolet light (EUV), fired through a stencil and onto a light-sensitive coating. Wherever the light lands, a transistor gets etched. Layer this dozens of times, and you have a chip.
Here is the detail that turns an engineering marvel into a genuine chokepoint: one company on Earth makes the machines capable of doing this at the most advanced level. Not "the market leader." The only maker. ASML, based in the Netherlands, is the sole global supplier of High-NA EUV lithography, the newest, sharpest tier of the technology, the one every future generation of advanced chip depends on. There is no second source. If you want the sharpest chips in the world, you are, by definition, waiting behind everyone else who also wants one, because there is no other line to join.
The Machine Itself
The scale of that machine matches the scale of its scarcity. Each one is roughly bus-sized, costs several hundred million dollars, and travels to its buyer across multiple cargo flights because it's too fragile and precise for any other transport method. Inside, it fires a droplet of molten tin, smaller than a grain of sand, and blasts it with a laser twice, fifty thousand times per second, vaporizing it into the exact plasma wavelength needed to generate usable light. Because that light gets absorbed by ordinary glass, the machine uses mirrors instead of lenses, mirrors so smooth that, scaled to the size of Germany, the tallest bump on the surface would be under a millimeter. ASML sources those mirrors from essentially one supplier. It took the company roughly three decades and, across ASML and its partners, more than $10 billion in cumulative research to get this technology from concept to a working commercial machine, which didn't ship until 2019.
The actual production numbers make the scarcity concrete rather than abstract. ASML holds 100% of the global EUV lithography market, not a dominant share, the entire market, full stop (Quantum Run). It expects to ship roughly 65 of the standard "Low-NA" EUV systems in 2026, and in July said it would add about 30 percent more capacity for 2027, while its CEO acknowledged that nearly all of that expanded EUV capacity through 2027 is already fully booked (ASML, SEC filing; Reuters, via Euronext). The newest "High-NA" tier is far scarcer. ASML's own estimate is about ten High-NA scanners for 2027, at roughly $380 million each (TechPowerUp), and Reuters described one early-2026 delivery as one of fewer than a dozen such systems in the world (TrendForce). Even ASML's CEO has said the company will avoid being the bottleneck "by all possible means" (Reuters/Economic Times, via LetsDataScience), a notable thing for a monopoly supplier to feel compelled to say out loud, and a tell about how much pressure sits on that single production line.
The Atelier Analogy
Think of it like a single atelier. Imagine every major fashion house on Earth, Chanel, Gucci, Prada, depended on one small, family-run workshop in Belgium that could hand-stitch a fixed number of gowns each year, and no other workshop on the planet knew the stitch. It wouldn't matter how much money Chanel had. There would still only be forty dresses. That is ASML's actual relationship to the global chip industry: TSMC, Samsung, and Intel are all, in effect, waiting on the same small production line, and the waitlist runs years, not months.
Terafab Runs Into This Wall Already
This is precisely the wall Elon Musk's newly announced "Terafab" project is running into in real time. Musk's venture, a joint effort across Tesla, SpaceX, and Intel, has committed $16.8 billion to a Texas megafab, with plans totaling as much as $119 billion across four phases through 2036, aimed at producing chips on Intel's advanced 1.4-nanometer process (TechCrunch; ENR). Intel's plans for that node are built around High-NA EUV, the scarcest tier of the scarcest machine on Earth, though Intel has also described a Low-NA-only option (Bits&Chips). Either way the process runs on EUV, and ASML says nearly all of its EUV capacity through 2027 is already booked by existing customers. State filings put the completion of Terafab's Grimes County facility at the end of 2028 (Wikipedia), so its tool orders land in the window where ASML is still deciding whether to add capacity, competing with customers who have been ordering for years. One independent analysis calculated that hitting Terafab's stated goal of one terawatt of annual compute would require roughly 22 million advanced logic wafers a year and something like 150 fab modules, a level of advanced manufacturing capacity that doesn't currently exist anywhere on the planet (Optimusk). Money bought Terafab a real, serious first phase. It did not buy it a place at the front of ASML's line. We go much deeper on this specific collision in a dedicated piece, linked below.
The Geopolitical Chokepoint
There's a second layer to the lithography bottleneck that has nothing to do with capacity at all, and everything to do with who gets permission to buy what exists. Because ASML is the sole maker of the most advanced machines, it accidentally became one of the most consequential chokepoints in global politics. Washington pressured the Dutch government to restrict ASML from selling its most advanced tools to China, not through a treaty or a tariff schedule, but essentially by leaning on the export license of a single private company. That one decision arguably shapes China's ceiling on advanced AI chip production more than any tariff package assembled since. The lesson generalizes beyond this one case: when an entire global industry depends on a single supplier, that supplier's shipping manifest becomes a lever of state power, whether or not the company ever asked for that role. The control is not guaranteed to last, either: Reuters reported in December 2025 that a Chinese team had built a prototype EUV machine that can generate EUV light but has not yet produced working chips, with working chips targeted for somewhere between 2028 and 2030 (Engadget, summarizing Reuters; Power Electronics News). Whoever controls access to the atelier controls who gets to make the best dresses at all, and in this case, several governments have realized that simultaneously.
Bottleneck Two: A War You've Never Heard Named Can Still Threaten a $50 Billion Factory
If lithography is a scarcity of machines, the second bottleneck is a scarcity of ingredients, and the volumes involved are almost comically small relative to the damage their absence can cause.
Start with neon. It sounds decorative, the gas in old diner signs, but ultra-high-purity neon is what powers the excimer lasers at the heart of the lithography process described above. Ukraine historically supplied somewhere between half and 70 percent of the world's semiconductor-grade neon, depending on the year and the source, a by-product of industrial air separation at its steelworks (USITC; MIT Sloan Management Review; SpecGas). When Russia's invasion halted shipments from two major Ukrainian suppliers, it wasn't a symbolic disruption, it threatened a genuine chokepoint in a global industry worth hundreds of billions of dollars, because there was no ready alternative source at the volume required (Tom's Hardware). Neon is also the one material where the industry has since done something about it: China is now the largest active producer, and Air Liquide and Linde added US purification capacity and supply contracts in early 2026 (SpecGas). The companion piece on materials tells that story in full.
Photoresist's Separate Chokepoint
Photoresist tells the same story with a different geography. This is the light-sensitive chemical coating that actually captures the pattern the lithography machine prints, the "film" in the camera analogy, more or less, and it's dominated by a small handful of Japanese firms. In 2026, a disruption to Middle Eastern naphtha supply (the raw material for the solvents photoresist is made from) rippled directly into warnings from Japanese suppliers, with the greatest exposure concentrated in the most advanced, EUV-dependent chips, the ones this whole supply chain exists to produce (TrendForce).
The Michelin-Star Analogy
Think of a Michelin-starred restaurant that cannot serve a single dish because the one supplier of a specific, irreplaceable spice ran out. The spice itself might cost almost nothing. The restaurant's entire menu, and its entire revenue, depends on it anyway, because there's no substitute that produces the same result. A modern chip fab is that restaurant, scaled to tens of billions of dollars: a few grams of the wrong gas being unavailable can shut the whole thing down, not because the gas is expensive, but because for some inputs there is no plan B (Betabanter). Switching to a secondary material isn't a simple substitution either, impurities in an alternative supply can quietly collapse yield rates, turning a materials problem into a production-throughput problem within a single quarter (Paragraph).
Rare Earths Complete the Picture
Rare earth elements complete the picture, and they concentrate the risk even further than neon or photoresist do. These are the metals used in the magnets, sensors, and specialty components scattered throughout the broader electronics supply chain feeding AI hardware, and China dominates their processing, not necessarily their mining, but the far more technically demanding step of refining raw ore into usable material, which is where the real chokepoint sits (Betabanter). It's a monopoly dynamic that echoes neon's: the raw resource itself isn't unique to one country, but the industrial capability to purify it to the standard advanced manufacturing requires is concentrated almost entirely in one place, built up over decades while nobody else found it worth the investment to compete. China has demonstrated a willingness to use this leverage directly: the Council on Foreign Relations counts export controls imposed between 2023 and 2026 on gallium, rare earths and at least seven other materials (CFR), and the International Energy Agency found that prices for these minerals had more than doubled, with gallium and heavy rare earths in Europe trading at around five times Chinese domestic prices (IEA).
Nobody Diversified the Boring Stuff
The pattern across all three materials, neon, photoresist, rare earths, is the same, and it's worth naming explicitly: each one is a case where decades of nobody bothering to diversify a cheap, boring input created a concentration risk that only became visible once a war, a regional supply shock, or a deliberate export restriction actually tested it. Neon has since begun to diversify; photoresist and the critical minerals have not. China, for its part, is racing to close this gap on the photoresist side specifically, targeting 40% domestic self-sufficiency in 2026 and already producing photoresists capable of supporting 45-nanometer chips, respectable, but still generations behind the sub-5-nanometer chips this entire AI buildout actually needs (Asia Times). Diversifying a boring, low-margin material takes just as long as building a fab, arguably longer, since almost nobody invests in redundant supply for an input that costs pennies until the day it doesn't.
Bottleneck Three: You Can Buy the Building. You Cannot Buy the Hands That Run It
The third bottleneck is the most human of the four, and in some ways the hardest to fix with any amount of capital: the United States' chip-manufacturing renaissance is short by as many as 157,000 skilled workers by 2030, according to a joint study from McKinsey and the SEMI Foundation, which puts the gap between 127,000 and 157,000 (CNBC; Bloomberg; Startup Fortune).
This isn't an abstraction sitting in a report, it's already visibly slowing real projects. TSMC's Arizona fab has faced multiple construction delays directly attributed to a shortage of skilled engineers (BigGo Finance). Intel's Ohio site needed roughly 7,000 construction workers in a region labor officials said simply didn't have them (CSIS). Companies have responded with genuinely large moves, TSMC offering signing bonuses exceeding $50,000 to lure technicians away from competitors, Intel committing $50 million to scholarships across more than 80 Ohio institutions, universities like Arizona State expanding microelectronics enrollment by 400% since 2022 (The Tech Buzz; CNBC), and it still isn't closing the gap fast enough, because training a fab technician or a process engineer takes years, not the single hiring cycle that capital can compress everything else into.
The Surgical Robot Analogy
Think of it like owning the most advanced surgical robot ever built and having no surgeon trained to operate it. The machine itself is not the constraint. The decade of accumulated, hands-on expertise required to run it safely is, and that expertise cannot be purchased, airlifted, or built with a subsidy check the way a building can. One report put it plainly: this is a wall that "no amount of CHIPS Act funding can fix" (The Tech Buzz), because the constraint isn't money. It's the number of human beings on Earth who currently know how to do this job, and that number grows on a training timeline, not a construction timeline.
The Cost Dimension
The shortage also carries a cost dimension that compounds the delay problem. Labor costs account for roughly half of the total cost gap between building a fab in Taiwan versus the United States, and three-quarters of that specific gap comes down to America's simply higher wage levels for the same skilled roles (BigGo Finance). Entry-level fab technicians in the US now start above $60,000 a year; experienced process engineers command well over $120,000 (The Tech Buzz), and even at those wages, companies are still short-staffed, which tells you the constraint isn't compensation, it's the raw number of qualified people who exist at any price. TSMC has responded in part by relocating hundreds of experienced engineers directly from Taiwan to Arizona to transfer institutional know-how in person, an approach that has generated its own friction, including reported cultural and labor disputes between imported management practices and the local workforce (Tech Insider). Even the most well-resourced company on Earth at this specific skill, in other words, cannot simply teleport its way around a shortage this structural.
The Brain Drain
There's a quieter driver behind the numbers worth naming too: a meaningful share of the shortfall is described as a "brain drain" toward higher-paying, higher-prestige AI and software roles, the same AI boom creating the demand for more chips is simultaneously pulling talent away from the unglamorous, cleanroom-based work of actually manufacturing them (BigGo Finance). The industry isn't just competing with itself for workers. It's competing with the very AI sector its own output exists to serve.
Bottleneck Four: A Desert Factory That Drinks Like a City
The fourth bottleneck sounds almost absurd until you see the numbers: making an advanced chip requires water purer than what you'd find in a hospital operating room, at a volume that rivals a small city's daily consumption, and a striking share of the world's most advanced fabs sit in places already short on water.
A single large fab can use anywhere from 2 to 10 million gallons of ultrapure water per day, depending on the estimate and the fab's size and node complexity, comparable to the daily water use of roughly 33,000 US households (World Economic Forum; Manufacturing Dive). TSMC's global operations alone consumed roughly 101 billion liters of water in 2023 (Robeco). Its first operating Arizona fab uses about 4.75 million gallons a day; once all three planned Phoenix-area fabs reach full production, gross water use is projected at roughly 16 to 17 million gallons per day, though with on-site reclamation the city expects to supply only about 4 million of it (Axios; Data Center Dynamics). That is still a meaningful draw on a metro region already navigating Colorado River shortages and historic drought (AZFamily). It takes roughly 1,400 to 1,600 gallons of ordinary municipal water to produce 1,000 gallons of the ultrapure water a fab actually needs, meaning the real draw on a strained water system is larger than the "ultrapure" figure alone suggests (WEF).
The Industry's Recycling Response
The industry knows this is precarious, and to its credit has responded with real engineering: TSMC currently recycles roughly 65% of its water intake at its Arizona site and has achieved over 90% process-water recycling rates in Taiwan; Intel has restored more than 1.7 billion gallons of water through on-site reclamation facilities in Arizona (AZFamily; UltraFacility; Area Development). But recycling technology has its own physical ceiling, and Taiwan's 2021 drought, the worst in half a century, still forced TSMC to truck in emergency water shipments despite having one of the most sophisticated water-management programs in the industry (Area Development).
The Hospital-in-the-Desert Analogy
Think of a hospital operating room that needs to run at the water consumption of an entire city, located in the middle of a desert. You can build the operating room. You can recycle a large share of its water. You cannot make it rain, and you cannot manufacture more Colorado River. This bottleneck is different in kind from the other three: it is the one most directly gated by geography and climate rather than engineering or capital, which is exactly why it's the hardest of the four to fully solve rather than merely manage.
Manufacturing Water From the Ocean
The industry's answer, where recycling alone isn't enough, has been to manufacture water from the one source that never runs dry: the ocean. Taiwan's government is building a major seawater desalination plant specifically to secure supply for Hsinchu Technology Park, the country's most important chipmaking hub, and a separate €320 million project is underway to guarantee uninterrupted water for the broader semiconductor cluster as demand keeps climbing over the coming decades (Robeco). Desalination is also, notably, expensive relative to ordinary municipal water, one Harvard-published estimate put a modest four-million-gallon-per-day desalination plant at under $100 million, which sounds almost affordable next to the cost of relocating a multi-billion-dollar fab, but is still a permanent new line item added to every future facility built somewhere water can't otherwise be guaranteed (Harvard Digital Innovation). Arizona has its own backstop: the state has reportedly stored nearly 3 trillion gallons of water in underground aquifers as a hedge against exactly this kind of squeeze (Area Development), a genuinely large buffer, though one that draws down rather than replenishes under sustained industrial demand.
Why Water Is Different From the Other Three
What makes water different from the other three bottlenecks is the direct collision with an entirely separate constituency: the ordinary residents of the same water-stressed regions these fabs are built in. Unlike a lithography allocation dispute, which plays out between corporations and a single supplier, a fab's water draw is a dispute with the public, in public, over a resource everyone in the region needs regardless of whether they've ever thought about a semiconductor. That's a fundamentally different kind of pressure than the other three bottlenecks face, and it's one that tends to generate political and regulatory friction on a timeline measured in permitting fights and local elections, not supply-chain contracts.
Why These Four Compound Each Other
None of these four bottlenecks operates in isolation, and treating them as four separate line items understates the real risk. A new fab needs all four simultaneously to reach production: it needs an ASML machine allocation, a stable materials supply chain, a fully staffed technical workforce, and a secured water source, all at once, all on overlapping timelines. A shortfall in any single one delays the whole facility, which is exactly why TSMC's Arizona buildout has been slowed by labor even where its lithography allocation was secure, and why Terafab's capital is real even though its lithography access is not. The four bottlenecks don't queue up one after another for a company to solve sequentially. They all have to clear at the same time, and the slowest one sets the pace for the entire project regardless of how far ahead the other three have progressed.
A Single Fab, Walked End to End
Consider what this looks like end to end for a single hypothetical fab, because the sequencing is where the real risk hides. A chipmaker secures its ASML allocation years in advance, the single hardest reservation to get, so naturally it gets locked in first. Groundbreaking happens on schedule. Then construction stalls for eighteen months because the region can't supply the 7,000 specialized workers needed to build out cleanrooms to spec. By the time construction catches up, a regional drought has tightened water permitting, adding another year of environmental review before the facility can secure its allocation. And by the time all of that clears, a fresh materials disruption, a new export restriction, a new regional conflict, has hit the specific photoresist or gas supplier the fab's process was designed around. Every one of the last three years' announced mega-fabs has run into at least one of these four walls after breaking ground, not because the companies involved were poorly run, but because the walls exist regardless of how well-run the company is, and because securing one of the four in advance offers no protection against the other three.
The Challenge: Isn't This Just How Every Capacity Cycle Looks From the Inside?
The fair counter-argument deserves real weight, and it applies to all four bottlenecks at once. Every prior semiconductor upcycle has looked exactly like this from the inside: acute shortages, urgent headlines, breathless claims that "this time is structural", and every one of them eventually resolved as new capacity came online, alternative suppliers emerged, or demand growth simply normalized. This isn't a hypothetical pattern, either, it has a direct precedent in exactly this piece's own material. Neon prices spiked after Russia's 2014 annexation of Crimea, the industry largely returned to business as usual, and it took the 2022 invasion to revive the same panic (USITC); industry trackers have written about the rare gas market as a recurring, cyclical phenomenon since at least 2017 (JRC). And this time the market did respond, with new neon capacity in China and the United States. ASML is expanding output. New water reclamation technology keeps improving recycling rates. Training pipelines, however slow, are genuinely scaling. Assuming that all four bottlenecks stay permanently binding risks repeating the oldest mistake in this industry's history: mistaking a temporary squeeze for a permanent ceiling.
Why the Four Aren't Actually Equivalent
The honest rebuttal to that rebuttal is that these four constraints are not identical in kind, and lumping them together risks understating the ones that are genuinely different this cycle. Lithography and labor are capacity problems, they scale, slowly, with investment and time. Materials and water are closer to hard physical limits, you can recycle water, but you cannot make a desert wetter, and you can diversify neon sourcing, but you cannot manufacture ultra-high-purity byproducts of a steel industry that doesn't exist at the required scale anywhere else. Treating all four as equally solvable-with-enough-time would be its own kind of overconfidence.
Conclusion
Efficiency and pricing move at the speed of software. These four bottlenecks move at the speed of chemistry, geology, and human training, and none of them have sped up to match the pace of AI demand, no matter how much capital has been thrown at them. That mismatch, more than any single company's roadmap or any single government's subsidy package, is the real constraint on how much AI compute the world can build this decade. Understanding where the actual wall sits is worth more than another headline about a new fab breaking ground.
This piece is the map. Five companion pieces go deeper on each thread:
Terafab vs. Taiwan: Can Money Buy What TSMC Took Fifteen Years to Build?, whether Musk's venture can out-execute TSMC on the thing TSMC has a fifteen-year head start on: allocation priority with ASML itself
Cheap Inputs, Expensive Failure: The Materials Chokepoint Behind Every AI Chip, why neon diversified, why photoresist hasn't, and why critical minerals have become a lever
The Talent Gap: The Bottleneck That Money Moves Around but Doesn't Create, why a 157,000-person gap is really two shortages, and why the engineering half is the hard one
Thirsty Silicon: The Water Economics of Building Chips in the Desert, why a cheap input can still decide where a fab gets built
Will the Race Ever End, and Where?, whether the buildout ends, and who benefits when it does
source
covers
Terafab's $16.8B first-phase investment
Terafab scale and site detail
Terafab's terawatt target vs. required fab modules
ASML's 100% EUV market share
ASML's ~10 High-NA scanners planned for 2027; 2026 EUV shipment plans
2026 Low-NA capacity (~65); +30% for 2027
ASML capacity boost; EUV capacity booked through 2027
ASML CEO on avoiding being the bottleneck
Fewer than a dozen High-NA systems worldwide
Intel 14A plans and the Low-NA-only option
Terafab's four-phase, $119B plan through 2036
Terafab timeline; state-filing completion date
China's EUV prototype (Dec 2025)
Status of China's EUV efforts
Chinese export controls, 2023-2026
Critical-mineral price doubling and divergence
China as largest neon producer; Air Liquide/Linde capacity
127,000-157,000 workforce shortfall range
Ukraine's share of global neon supply
Neon's role in lithography lasers
Neon market as a recurring, cyclical pattern
Ukrainian neon supplier disruption
Japan photoresist/naphtha shortage
Materials concentration risk framing
Yield impact of alternative materials
157,000-worker shortage; industry training response
Labor shortage vs. planned investment
TSMC Arizona delays tied to labor
Construction labor shortfall, Ohio/Arizona
Signing bonuses; training investment
Ultrapure water volumes; municipal water ratio
Water as largest-volume fab chemical
TSMC's total 2023 water consumption
Arizona-specific TSMC water use and recycling
Gross vs. net Phoenix water draw
Phoenix fab water use; reclamation plant timing
Taiwan's 2021 drought response
Water recycling rate comparisons
Desalination plant cost economics
China's photoresist self-sufficiency push
TSMC Arizona labor transfer, cultural friction
