The short version
Parts 1 to 3 covered the chain from sand to server, and the companies that control it. This part covers what happens when that chain meets the physical world, which has its own schedule and does not care about demand forecasts.
Electricity is the binding constraint
Notice how the industry talks about itself now. New sites are announced in gigawatts, not in square metres or server counts. That change in unit is the whole story. Floor space is easy. Power is not.
The numbers are stark. Globally, more than 2,500 gigawatts of projects, data centers among them, are stuck waiting to connect to electricity grids. In the United States alone, roughly 2,060 gigawatts of generation and storage capacity sit in interconnection queues, which is more than the entire installed capacity of the American grid at about 1,374 gigawatts. Two caveats the same Berkeley Lab report attaches: this is not an assessment of resource adequacy, and historically only around an eighth of queued capacity is ever built.
Berkeley Lab puts the median wait to connect at about 61 months, roughly five years, and longer in the markets carrying most of the announced buildout. Projects nationally have averaged a median of about 45 months, nearly four years, just to reach an interconnection agreement, and more after that before coming online. One widely cited estimate put a large share of large-scale data center capacity scheduled for 2026 at risk of delay, though other analysts have pushed back hard on that figure.
The instinctive response is to build more power plants, and companies are. But generation was rarely the slow part. The slow parts are:
Transformers. The equipment that steps voltage down from transmission lines to usable levels. Lead times run roughly 36 to 60 months, on the US Department of Energy's 2024 figures. There is no way to expedite this meaningfully, because the global manufacturing base for large transformers is small and already full.
Transmission lines. Getting power from where it is generated to where it is needed requires land, permits, and agreement from every community along the route. A decade is normal, and considerably longer is not unusual. Part 6 describes a Korean line that took 21 years.
Turbines. For operators building their own gas generation, turbine prices are on pace to rise about 195 percent by the end of 2027 compared with 2019, and order books stretch years out.
So the industry is routing around the grid. Microsoft, Amazon, Google, Meta and Oracle have all announced nuclear agreements, covering existing reactors, restarts of retired plants, and small modular reactors expected online later this decade. Others are building dedicated gas generation on site, accepting the emissions cost in exchange for a schedule they control.
There is an irony worth naming. An industry that markets itself as weightless and virtual has become one of the most physically constrained on earth, waiting on copper, concrete and steel.
Water and heat
Less discussed and locally more contentious. A current AI rack draws an enormous amount of power, and essentially all of that energy becomes heat that must be removed. Air cannot do it at that density, so these systems are cooled by liquid piped directly to the chips.
Some designs use evaporative cooling, which consumes water. Others use closed loops, which consume much less but require more power. The trade-off is real and site-specific, and it has become a live political question wherever data centers are proposed near communities that are already water-stressed. It is one of the most common sources of local opposition, and local opposition adds years.
The money
The spending is difficult to hold in perspective, so anchor it against something.
The four largest US cloud companies plan an enormous combined capital expenditure in 2026, a sharp increase from 2025. Reported guidance has moved upward repeatedly through the year at every one of the four: Amazon raised its figure at its Q2 2026 earnings call on 30 July 2026, citing higher memory costs; Google raised its figure at its Q2 2026 call on 22 July 2026; Microsoft revised its figure down at its 29 July 2026 call, after a lease-accounting change moved future data center leases off the capital spending line; and Meta anticipates capital expenditures of approximately USD 130 billion to USD 145 billion in 2026, per its filed Q2 2026 10-Q. Different trackers count different things and produce different totals, so treat any single number as an estimate rather than a fact.
That is annual spending by four companies exceeding the entire GDP of most countries.
Beyond the cloud companies, OpenAI's Stargate programme has announced a commitment to a very large amount of planned capacity and an enormous investment over several years, with Oracle and SoftBank. Notably, one analyst reported in March 2026 that parts of it had been scaled back amid financing complexity and revised demand assumptions; Oracle publicly denied any setback and said the site remains on track. That disagreement is itself a useful reminder that announced plans and built capacity are different things.
And a newer category has appeared. Neoclouds exist to buy accelerators and rent them out. CoreWeave, the largest, disclosed USD 98.8 billion of remaining performance obligations as of 31 March 2026. Its last full-year filing put it at 43 data centers, 850 megawatts of active power and 3.1 gigawatts contracted at end-2025.
The circular financing question
Here is where reasonable people disagree, and where it is worth being careful.
A growing share of the deals that generate this spending involve the supplier financing the buyer. Nvidia was reported in 2026 to be discussing a very large backstop for OpenAI, supporting a planned data center campus in Ohio, alongside reported arrangements covering a further large volume of chip purchases. In one reported week, Nvidia signed or opened talks on an enormous combined value of deals, including a letter of intent with SK Group and an investment into an AI research company in return for it buying substantially more Nvidia compute.
The concern is straightforward. If a chip company funds a customer who then buys chips from it, revenue is recorded, but part of that revenue began as the seller's own money. Critics draw a direct line to the vendor financing that preceded the dot-com collapse. Markets have paid attention: Nvidia's shares fell sharply on one day in late July 2026, briefly putting its market capitalisation below Apple's for the first time in over a year.
The counterargument is also serious. Infrastructure has always been financed by parties with an interest in it existing. Railways, telecoms and power generation were all built this way. The customers here have genuine demand they cannot currently satisfy, and the constraint is capital and construction time rather than appetite.
Two things are widely reported and rarely contested, even though neither comes from a filed or audited source. OpenAI is reported to be on track to lose a very large amount of money in 2026, a sharp increase from its 2025 losses, while projecting an enormous revenue figure by 2029. And a large share of the capital being deployed assumes AI hardware remains useful and valuable for a number of years, an assumption that has not yet been tested through a full hardware generation.
Where Korea sits in this layer
Korea faces every constraint in this part, in a compressed form, because it is trying to build one of the world's largest semiconductor clusters on a small and crowded peninsula.
The Yongin (์ฉ์ธ) cluster is projected to need roughly 15 to 16 gigawatts at full operation. The site can supply only a small fraction of that on its own. Closing that gap requires long-distance transmission from the east coast and the southwest, and Korea Electric Power is pursuing a 1,153 km, 345 kV transmission network to do it, in three stages running to 2030, 2036 and 2042; the 2036 date is the second stage, not the finish line. The government has committed to completing power supply within the cluster by 2042, already pulled forward by more than a decade.
Korea also has a specific problem the United States does not: industrial electricity now costs more than household electricity, an inversion of the usual pattern, which is pushing data center operators toward direct contracts with gas generators rather than the public grid. Part 6 covers all of this, including what it means for people who pay Korean electricity bills.
The escape hatch: compute in orbit
Worth covering because it is being seriously funded, and worth treating skeptically because almost none of it exists yet.
The pitch is that several of the constraints above disappear in space. Sunlight is continuous and unfiltered in the right orbit. There are no neighbours to object, no land to buy, and no water needed for cooling.
The filings are real. SpaceX, which completed a merger with xAI on 2 February 2026, applied to the FCC on 30 January 2026 for up to one million orbital data center satellites at altitudes between 500 and 2,000 km, projecting that launching a million tonnes of satellites annually would yield 100 gigawatts of compute capacity. Announcing the SpaceX and xAI merger in February 2026, Elon Musk said that within two to three years the cheapest way to generate AI compute will be in space. SpaceX's own registration statement, filed four months later, is more guarded: it says orbital data centers can eventually achieve a lower cost than terrestrial data centers over time, and that the timeline for reaching 100 gigawatts in orbit may be difficult or impossible to determine. Days later, Starcloud filed for an 88,000-satellite constellation, having raised a substantial funding round in March 2026 at a valuation over a billion dollars. Google's Project Suncatcher, announced in November 2025, plans to launch two prototype satellites carrying TPUs by early 2027.
The unsolved problems are substantial. Heat rejection is the hardest: in a vacuum there is no air or water to carry heat away, so everything must be radiated, which requires large radiator surfaces. Launch costs must fall much further to make the economics work. Radiation degrades chips. And hardware in orbit cannot be repaired or upgraded the way a data center can.
Treat announced constellation capacity as a statement of intent. As of now, the world's AI compute runs in buildings, connected to grids, waiting in queues.
Part 5 turns to the force that has recently done more to reshape this industry than any technology: government policy.