You can't buy your way out
The richest companies in the world can't buy their way to more AI any faster. Money buys a place in line. Only building makes the line shorter, and the lines are now measured in years while AI models arrive every few months.
trying to answer →What becomes scarce when intelligence becomes cheap?What is a moat in an AI world?
In 2024 Microsoft agreed to pay for a shut-down nuclear reactor at Three Mile Island to be switched back on. In 2026 companies were reserving gas turbines for delivery in 2031, and Meta agreed to pay in advance for electricity from a reactor site whose first phase is targeted for 2030. These are not the moves of companies short of money. They are the moves of companies that have found out money isn’t the thing that’s short.
The rule underneath is simple, and most of the AI build-out follows from it.
money can buy
your place in line
allocation · priority · reservations · paying up front
only building can shorten
the line itself
factories · clean rooms · qualified suppliers · machines · trained people · grid connections
Every part of an AI data centre has a line in front of it: the time it takes someone to make more. Some lines are weeks, some are years. The useful way to measure them isn’t in years but in AI models, because that is the race the buyer is running: a new frontier model arrives roughly every six months. So walk out from the chip to the power line, and count how many models ship while you wait at each step.
The chip
The memory next to the chip
Advanced packaging (chip-to-chip)
The transformer
The power plant
Plugging in
less than one new AI model ships while you wait for more chipsabout 6 new AI models ship while you wait for a new memory factory2 to 3 new AI models ship while you wait for a new packaging line5 to 8 new AI models ship while you wait for a big transformer9 to 10 new AI models ship while you wait for a new gas turbine8 to 14 new AI models ship while you wait for a grid connection
each blue tick: a new frontier model, about every six months
step 1 of 6
The chip
Getting more AI chips takes weeks once you have been given an allocation. The chip itself stopped being the hard part a while ago.
step 2 of 6
The memory next to the chip
An AI chip spends much of its time waiting for data from a special stacked memory that sits right beside it. Three companies make it, and the leader’s next new factory opens in 2029.
step 3 of 6
Advanced packaging (chip-to-chip)
Chip and memory are bonded into one package, a step done mostly by TSMC. New lines have been quoted at one to one and a half years. More are being added now, which moves the wall rather than removing it.
step 4 of 6
The transformer
The building needs transformers to bring grid power down to a usable voltage. A handful of factories make the big ones, all short of the same special steel. The wait is two and a half to four years.
step 5 of 6
The power plant
More electricity means more power plants. A large gas turbine ordered in 2026 arrives around 2031.
step 6 of 6
Plugging in
And before any of it can be switched on, it has to be allowed onto the grid: years of studies and upgrades, in queues thousands of projects long.
Look at where the long lines are. The quick fixes near the chip have mostly been made; money is good at those. What’s left are the waits measured in years, and the longest of all are at the far end, in power plants and the grid.
That isn’t a coincidence. It’s how the wall moves. Anything that can be fixed in months gets fixed, by someone with money. So the wall never stays where it is. It moves outward, one step at a time, until it reaches the thing money can least hurry. That’s why, over three years, the thing everyone was short of went from the chip, to the memory beside it, to the packaging, to the cooling, and out to the power.
And the famous names are only the first layer. Behind each one sits a supplier almost nobody has heard of, with a line of its own.
A food company
Ajinomoto
Ajinomoto has made the flavour enhancer MSG since 1909. In the 1990s its chemists, working from the same chemistry, made a plastic film that insulates better and lies flatter than anything else. Almost every high-end chip package now has layers of it inside. TSMC can build more packaging lines. It can’t build more of a food company’s film.
A mirror polisher
Zeiss
Every advanced chip is printed by machines from one Dutch company, ASML. The mirrors inside those machines come from another company, Zeiss, which polishes them to within about a nanometre. The famous bottleneck has its own bottleneck.
A capacitor
High-capacitance capacitors
A component smaller than a grain of rice, hundreds to a board. On NVIDIA’s newest boards one type went from 320 per board to 500, and the wait for some of them from eight weeks to twenty. Nobody plans a data centre around a capacitor, until it is the thing missing.
TrendForce, structural shortages of high-end MLCCs may emerge in 2H26 (17 Jun 2026)
What a strategist does with this
Buy time, not things. The companies ahead are the ones that reserved years ago: turbines booked for 2031, a reactor brought back, power paid for before the plant exists. When the scarce thing is time, a reservation is a moat, and the only way to shorten a wait measured in years is to have started years ago.
Own a place in line, not a product. The money stays with whoever already holds the capacity: the three memory makers, TSMC, the owners of land that already has a power connection. Standing next to a shortage is not the same as holding it. Who got paid follows that money rung by rung.
Watch the longest line. The wall moves toward whatever money can least hurry. Today’s longest wait is where the constraint is heading, and where the pricing power will be when it gets there.
The ledger every wait, every supplier, every source, for checking the story
Every wait
Chip factories new leading-edge (3 nm) wafer capacity
no wait published: it's a new factory“3nm capacity, dominated by TSMC, has become “a scarce resource fiercely contested by global tech giants” (TrendForce). No lead time for a new leading-edge line has been published; it is a fab.” not quoted TrendForce, AI competition tightens advanced packaging and 3nm capacity (30 Apr 2026)
The AI chip an allocation of accelerators
≈ 1–0 model“Allocation first, then weeks. The wait is commercial, not technical.” our estimate NVIDIA H100 product page · NVIDIA GB200 NVL72
Memory next to the chip new HBM output, counted from September 2026
≈ 6 models“SK Hynix’s new Yongin Y2 cleanroom is targeted for June 2029, and its Indiana HBM plant for volume production in the third quarter of 2029.” as reported SK hynix, fab and facility investment (7 Aug 2026) · Quartz, SK Hynix breaks ground on $4 billion Indiana HBM plant (27 Aug 2026)
Advanced packaging (chip-to-chip) CoWoS capacity at TSMC
≈ 2–3 models“52 to 78 weeks quoted, with lines sold out through 2026 into 2027. Allocation is committed two to three years ahead. Capacity is expanding: TrendForce expects the supply gap to narrow from about 20% to about 10% by the end of 2026, which moves the bottleneck rather than removing it.” as reported Tom's Hardware, TSMC: shortage of Nvidia's AI GPUs to persist for 1.5 years (Sep 2023) · TrendForce, CoWoS supply-demand gap seen narrowing from 20% to 10% by end-2026 (15 Jun 2026) · TrendForce, CoWoS supply-demand gap seen narrowing from 20% to 10% by end-2026 (15 Jun 2026)
Wiring between chips qualified optical modules and lasers
≈ 1–4 models“Laser and module lines take quarters to years to qualify, and the photonics startups are years from volume.” our estimate Invezz, Coherent and Lumentum continue surge (13 May 2026)
Assembly into racks a qualified systems-integration slot
≈ 1–2 models“Quarters. A new assembly line is a building and a workforce, not a fab.” our estimate Yahoo Finance, Dell stock surges on record orders for AI servers (Sep 2026) · Supermicro, fourth quarter and full fiscal year 2026 financial results · Wiwynn, first quarter 2026 financial results · Hon Hai (Foxconn), second quarter 2026 financial results
Cooling a new cold-plate and coolant-distribution line (the building’s own cooling is part of the site)
≈ 1–4 models“Manufacturing lines are added in months to a year or two, far faster than a fab.” our estimate Motley Fool, data center demand drove Vertiv’s earnings up 83% (1 May 2026)
Transformers and power gear high-capacitance capacitors (MLCCs) for accelerator boards
≈ 1 model“On NVIDIA’s Vera Rubin platform, demand for one high-capacitance MLCC rose from 320 to 500 per board, and lead times for certain high-capacitance parts stretched from eight weeks to as long as twenty (TrendForce, June 2026).” as reported TrendForce, structural shortages of high-end MLCCs may emerge in 2H26 (17 Jun 2026)
Transformers and power gear a large power transformer
≈ 5–8 models“Power transformers averaged 128 weeks to deliver and generator step-up transformers 144 weeks in 2025; substation transformers passed 160 weeks in 2026. US transformer lead times reached as long as four years by 2026.” as reported POWER Magazine, Transformers in 2026: shortage, scramble, or self-inflicted crisis? · pv magazine USA, transformer lead times extend to four years (11 May 2026)
Buildings and land a new hall, from site control to first power
≈ 4–8 models“Two to four years from site control to first power.” as reported IEA, Energy and AI (2025)
Power plants a heavy-duty gas turbine, ordered in 2026
≈ 9–10 models“A heavy-duty gas turbine ordered in 2026 is delivered around 2031. A new reactor takes longer.” as reported Utility Dive, GE Vernova gas turbine backlog climbs to 116 GW (2026)
Plugging into the grid a connection to the transmission grid
≈ 8–14 models“US interconnection processes can stretch across years: at the end of 2025 about 8,200 projects were waiting in the queues. Those are queues for power plants, not data centres, so the count of models here is this drawing’s translation of four to seven years, not an average wait.” as reported Berkeley Lab, Queued Up: 2026 Edition · LBNL, Queued Up: interconnection queues
Who makes each part, and who decides how many
in black, who sells the part. in pen, the companies that don't sell it, but decide how many can be made
Chip factories
TSMC
- ASML EUV scanners. Nothing below 7 nm is patterned without one, and they ship on the order of tens a year
The AI chip
NVIDIA · AMD · Broadcom, Marvell
- Applied Materials, Lam, Tokyo Electron deposition and etch, the steps that actually build the transistor
Memory next to the chip
SK Hynix · Samsung · Micron
- Hanmi the MR-MUF bonders SK Hynix stacks on
- ASMPT, Kulicke & Soffa thermal compression bonders, the route Micron and Samsung took instead
- DISCO grinds each DRAM die under 40 um so twelve fit in the height budget
- Lam Research, Tokyo Electron etches the through-silicon vias, 8:1 aspect ratio and deeper
Advanced packaging (chip-to-chip)
TSMC · Intel Foundry · Amkor, ASE
- Ajinomoto coats the build-up film almost every substrate uses
- Ibiden, Unimicron the substrates, on qualification cycles of their own
- BESI, Applied Materials hybrid bonders, sub-micron placement
- ASMPT, Kulicke & Soffa thermal compression bonders
Wiring between chips
NVIDIA · Broadcom, Marvell · Coherent, Lumentum, Innolight
- InP epitaxy capacity every optical port needs a laser, and laser wafers are a thinner supply chain than the transceivers built on them
Assembly into racks
Foxconn, Quanta, Wistron · Dell, Supermicro, HPE
- Amphenol, Samtec, TE Connectivity the backplane connectors carrying 224G PAM4 across thousands of channels
Cooling
Vertiv, CoolIT, Boyd, nVent · Motivair, LiquidStack
- Danfoss, Staubli, CPC dripless quick-disconnects. Every one is a joint that must not leak for the life of the loop
- Vacuum braze and diffusion bonding capacity how a cold plate is actually made, and it is not a large industry
- Indium Corporation, Henkel, Honeywell the interface materials themselves: liquid metal alloys and sintered silver preforms
- Nordson, Musashi the dispensers that place them to a few microns, repeatably
Transformers and power gear
Vertiv, Eaton, Schneider, ABB · Delta, Flex, Monolithic Power · Hitachi Energy, Siemens Energy
- Grain-oriented electrical steel the core material, rolled and annealed so the crystal grains line up with the flux. A handful of mills make it, which is why paying more does not make a transformer arrive sooner
- Cycloaliphatic epoxy the cast insulation in switchgear. When it is short, the whole line lengthens behind it
Buildings and land
Equinix, Digital Realty, Vantage, QTS · Microsoft, Amazon, Google, Meta
- Caterpillar, Cummins, Kohler multi-megawatt gensets, ordered years ahead
- Baltimore Aircoil, EVAPCO, SPX towers and fluid coolers, the heat rejection side of the building
Power plants
Independent power producers
Plugging into the grid
Regulated utilities and system operators · GE Vernova, Siemens Energy, Mitsubishi
- Southwire, Hitachi Energy EHV breakers and conductor for the line itself
- Line crews and commissioning labour the constraint nobody quotes, because it cannot be ordered
More names underneath the names
Ajinomoto
A food company sits unexpectedly deep in the package-substrate supply chain.
Zeiss SMT
The constraint behind ASML: it polishes the EUV mirrors to about a nanometre.
Hanmi
A niche Korean toolmaker whose bonders SK Hynix stacks its HBM on.
DISCO
A near-monopoly on grinding and dicing, a step nobody writes about.
BESI
Hybrid bonders, the tool that decides whether 3D stacking scales.
High-capacitance MLCCs
A capacitor nobody thinks about: one specification went from 320 to 500 per board on NVIDIA’s Vera Rubin, and lead times from eight weeks to twenty.
Grain-oriented electrical steel
Why paying more cannot make a transformer arrive sooner.
Stäubli / Danfoss
A hose coupling that gates a data hall.
Photoresist (JSR, Tokyo Ohka)
The 2019 Japan–Korea export controls showed how sharp this one is.