Too Slow

The Dark Station

More than four thousand people are at work in the desert outside Santa Teresa, New Mexico, building a data centre that cannot yet be switched on.

Project Jupiter, at Santa Teresa in Doña Ana County, New Mexico, is meant to become one of the largest computing campuses on earth: roughly 2.5 gigawatts, developed by STACK Infrastructure for Oracle, with OpenAI reported as the end user. The buildings are nearly finished. Earlier this week the developers gave the press its first official tour of the site, and the cooling system was on display. What could not be displayed was electricity, because there isn't any yet.¹

The campus is meant to run on its own fuel-cell power plant, and the plant needs gas. The pipeline to deliver it was rejected by the state's land commissioner, rerouted across federal land, and is now not expected before February 2027. New Mexico's attorney general wants a full environmental review. On 13 October Oracle wrote to the federal regulator asking it to hurry, warning that further delay "threatens to delay Project Jupiter's operations."² Two days later it revised the design of the power plant to cut its emissions by about a third, and a hearing on a motion to dismiss the air-quality permit application is expected on 26 October.³ The tax agreements with Doña Ana County committed the first phase to operation between October and December of this year.⁴ It is October.

In September, Oracle sent its landlord a force majeure notice. If both sides accept that a power-related force majeure event has occurred, Oracle could defer rent for up to three years once payments begin.⁵ Meanwhile the construction loans, roughly $18 billion from a syndicate of about twenty banks, were already being quoted around 89 to 91 cents on the dollar in September.⁶

The September series called this an act of Jupiter. A month on, the station has been built. The trains have not been told when they may run.

Jupiter is not a failed project. It may well be finished, powered and full of Oracle's customers in due course. That is precisely the point. The technology can be right, the tenant can be real and the buildings can be standing, and the financing can still be early.

This is the first of three parts on that problem. The thesis is simple enough to state in one sentence: the AI build cannot be completed on the timetable its financing assumes, at the cost its financing assumes.

The Catch

In Joseph Heller's novel, a bomber pilot can be grounded if he is mad, but asking to be grounded proves he is sane. The rule is perfectly consistent. It simply leaves no way out.

The AI build has a Catch of its own, and it comes in three parts.

Too slow if demand is real. Suppose the demand for computing is every bit as large as its champions believe. Then the industry needs more power, more transformers, more turbines, more memory and more finished buildings than the supply chain can deliver on time. Costs rise and projects slip. Interest accrues on the loans while the revenue waits for the electricity. A developer can be right about demand and still earn an inadequate return on the capital it borrowed to meet it.

Too soon if it is not. Suppose instead that models become efficient faster than expected, or that enterprises take longer to turn experiments into paying use. The equipment already ordered still arrives. The buildings are still finished. But the contracted demand, at the contracted price, is smaller or later than the financing assumed. The asset exists. The projected cash flow does not.

Too cheap when it arrives. Suppose everything goes to plan, and competition does what competition does. The price of a unit of intelligence keeps falling. Volumes rise, perhaps enormously, but the revenue earned per dollar of capital installed falls with them. The buildings and the debt last for fifteen or twenty years. The commercial advantage of the chips inside them may last for two or three.

The technology can succeed in all three cases. The returns on the capital financing it need not. That is the Catch.

In August, A Credit Cycle Wearing AI's Costume set two of the industry's most bullish voices against each other. Dario Amodei worries about building compute too far ahead of the revenue. Elon Musk worries about the physical infrastructure arriving too slowly to serve the demand. Both can be right at the same time, and if they are, the gap between them is where the money is lost.

The simplest way to picture that gap is as two clocks. The first measures how quickly intelligence improves and becomes cheaper. It runs in months. The second measures how quickly power stations, substations, buildings and loans can be arranged, built and repaid. It runs in years. The question this series asks is not whether AI is a bubble. It is how far apart those two clocks have drifted, and who is holding the debt in between.

This first part deals with the physical clock: how much infrastructure is needed, how much is being financed, what a gigawatt is made of and how long it takes to build. Part II, Too Soon, follows the bill: cost overruns, three real projects and who carries the timing. Part III, Too Cheap, asks who gets paid when intelligence gets cheaper.

What Demand Measures

The first difficulty is that "AI demand" is measured in several different units, and they are routinely treated as though they were the same thing.

They are not.

Each stage in that table describes genuine activity. A model developer reports surging usage. Forecasters extrapolate it. A cloud provider signs a contract to supply the capacity. The contract supports a loan for a data centre. The developer orders chips, memory and electrical equipment, and the chipmakers book the revenue on delivery. Every step is real. Yet the return on the whole chain still depends on whether the final customer earns, or saves, enough to pay for it.

A promise to consume compute tomorrow has become a financeable asset today.

Tokens, the units of text a model reads and writes, are the industry's favourite measure of success, and the least reliable. At Vercel's AI gateway, open-weight models, which anyone can download and run, accounted for 56% of token volume in August, up from 7% last December. They accounted for 14% of the spending.⁷ The volume moved. The money mostly did not.

Some of the volume is not even demand in the ordinary sense. Earlier this year, insiders said, Meta ran an internal leaderboard ranking staff by the AI tokens they consumed, before closing it within days amid some embarrassment. The practice has acquired a name, tokenmaxxing, treating consumption itself as evidence of productivity. Its predecessor was the dot-com habit of counting eyeballs.

Hock Tan, whose company Broadcom supplies the chips and builds the networks, put it more bluntly in Singapore this month. Companies, he said, are "wasting money, spending tokens", like a child who empties ten dollars of tokens into an arcade in ten minutes and learns nothing.⁸ Coming from one of the largest beneficiaries of the spending, that is worth noting.

None of this means demand is weak. It means that demand measured in tokens cannot be converted directly into gigawatts, still less into the dollars needed to service the debt on them. The measure that would settle the question, cash earned per useful task completed, after the full cost of the capital behind it, is not yet disclosed by anyone in a consistent form. The absence of that number is itself a reason for caution.

So the series uses a different unit.

Paper Gigawatts

Electricity is the one input every model, chip and building has in common. It is measured, it has a history, and it cannot be annualised from a good month. Power is the honest unit.

What a Gigawatt Is

Megawatts and gigawatts are now thrown around in AI announcements as casually as market capitalisations, and they are just as hard to picture. It is worth pausing on what they mean.

A watt is a rate, not a quantity. It measures how fast electricity is used, much as miles per hour measures speed rather than distance. A megawatt is a million watts; a gigawatt is a thousand megawatts. Run a gigawatt for an hour and you have used a gigawatt-hour of energy; run it for a year and you have used 8.76 terawatt-hours.

A single megawatt, running all year, uses about as much electricity as 800 American homes, or more than 3,000 British ones, which use far less.⁹ A gigawatt is a thousand times that: roughly 800,000 American homes, or more than three million British homes. It is about the output of one large nuclear reactor.

GW_SCALEGW_SCALE · Illustration · EIA; Ofgem; author's conversions

What a Gigawatt Is

Capacity in megawatts, with the homes each would supply running all year. Four rungs, from one megawatt to Project Jupiter.

Illustration

Now scale up. Jupiter's 2.5 gigawatts is more than the capacity of the Hoover Dam. At full load it would draw as much electricity as about two million American homes, more than twice the number of households in New Mexico, about 840,000.⁹ A single campus, in other words, built to consume the electricity of a state.

And the industry's plans are not counted in single campuses. The US pipeline described below runs to about 509 gigawatts. Britain's entire electricity system meets a peak of about 60 gigawatts in an average cold winter.¹⁰ The plans on paper amount to more than eight Britains.

EIGHT_BRITAINSEIGHT_BRITAINS · Illustration · Van Nieuwerburgh (2026); NESO Winter Outlook 2025/26

More Than Eight Britains

US planned data-centre capacity, gigawatts, drawn in blocks of Great Britain's peak electricity demand in an average cold spell (~60 GW).

Illustration

That is the enormity of what is being proposed. Every one of those gigawatts must be generated, transmitted, transformed, switched, cooled and paid for, and every one of them needs equipment that is already in short supply. Keep that scale in mind, because the rest of this series is an attempt to work out how much of it is needed, how much of it can be built, and who has lent against the difference.

How Much Has Been Needed

The International Energy Agency's latest assessment, published in April, puts global data-centre electricity consumption at about 485 terawatt-hours in 2025, up 17% in a single year. In its central case, consumption roughly doubles to about 950 terawatt-hours by 2030, with the AI-focused share tripling.¹¹ The United States accounts for the largest part, about 224 terawatt-hours last year, rising to 426 by 2030.¹¹

IEA_DCEIEA_DCE · 2025 estimate, 2030 base case · IEA, Key Questions on Energy and AI (Apr 2026)

The Demand Forecast: 485 to 950 TWh

Global data-centre electricity consumption, terawatt-hours. IEA 2025 estimate and 2030 base-case projection.

Reported

Those figures are energy, not power, and the difference matters. Energy, measured in terawatt-hours, is how much electricity is used over a year. Power, measured in gigawatts, is how much is being drawn at any moment. A gigawatt drawn continuously for a year uses 8.76 terawatt-hours, so the IEA's forecast can be turned into power. By 2030, it means, the world's data centres would draw on average about 108 gigawatts, every hour of every day: the output of roughly 108 large power stations running without a break. About 49 gigawatts of that would be in the United States.

Three things follow. First, these are all data centres, not just AI. Cloud computing, storage, streaming and every other digital service are in the total, so not all of it can be attributed to the AI build. Second, 49 gigawatts is total demand, not new construction. It includes facilities already running, and the addition the IEA expects in the United States by 2030 is about 23 gigawatts of average load. Third, it is average load, not installed capacity. Data centres are built to carry their peaks, their redundancy and their future tenants, not their average day, so installed capacity has to exceed average load. That is why the 57 gigawatts already operating, counted below, can exceed the IEA's 49 gigawatts for 2030 without contradiction: one measures what is installed, the other what is drawn on average. (Think of a restaurant. It needs enough seats for the Saturday-night rush, not for an average Tuesday lunch. Data centres are the same: they are built for their busiest moments, for backup equipment that sits idle until something fails, and for tenants still moving in. So today about 57 gigawatts of buildings are installed, but on average they draw only about 26 gigawatts, a little under half. If that stays true, the 49 gigawatts of average demand the IEA expects in 2030 would need about 107 gigawatts of buildings, roughly 50 gigawatts more than exist today.)

For the same reason, 49 gigawatts cannot be set directly against the 509 gigawatts in the development pipeline, which measures proposed capacity rather than electricity consumed. The question is how much of that proposed capacity will be needed, how much can realistically be delivered, and whether the revenue will justify the capital committed to building it. Even allowing generously for headroom, the IEA's central path implies a US requirement measured in tens of gigawatts this decade, not hundreds.

The constraint is also local. The IEA expects data centres to use only about 3% of the world's electricity in 2030.¹² The strain comes from where they cluster: a handful of grids must find the connections, substations, transformers and dependable generation for most of the growth. The question is not whether the world can produce enough electricity. It is whether enough deliverable power can be made available in the right places, on the promised timetable, at a cost the economics of the build can bear.

Read together, the two halves of this argument can look contradictory. If the plans far exceed what the IEA expects to be used, how can the build be too slow? The answer is that the surplus and the shortage are in different places. The surplus is in the plans: hundreds of gigawatts announced against the tens that independent forecasts require, much of it never to be built. The shortage is in deliverable power at the particular sites where buildings have already been financed, leased and filled with orders for chips, and where the debt assumes the electricity arrives on a given date. A national surplus of plans does not connect a single campus sooner. Too much is planned; too little of what is needed can be delivered in time.

That shortage is the heart of this part, and it does not depend on the 509 gigawatts at all. Strip out every speculative project, every duplicated queue entry and every announcement that will never be built, and the problem remains. Much of what the IEA's central case needs is already being built: Cushman counts 37.7 gigawatts under construction across the Americas, most of the roughly 50 gigawatts of extra buildings that 2030 requires.¹³ But under construction is not switched on. Every one of those buildings must still pass through the same grid queues, transformer factories and turbine order books as everything else. Connecting a large load to the grid takes about five years. Transformers and switchgear are on lead times measured in years. Turbine orders run well ahead of the factories that make them. The 2030 total may well be reached. What will struggle is the timetable in between: the dates on which these buildings were promised power, and on which their loans begin to fall due.

Even the low estimate runs late against its debt.

How Much Has Been Committed

Now set that against the plans.

Stijn Van Nieuwerburgh's paper for the Brookings Papers on Economic Activity, presented in September, counts about 57 gigawatts of US data-centre capacity operating in July and about 509 gigawatts planned. In his central scenario, 182.7 gigawatts are built by 2032, a further 117.2 gigawatts later, and 226.9 gigawatts never.¹⁴ Cushman & Wakefield's September guide counts 50.4 gigawatts operating across the Americas, 37.7 gigawatts under construction and 285.4 gigawatts planned.¹³ The two sources use different scopes and dates, but they agree on the shape: the planned pipeline is several times the capacity that exists, and several times what the IEA's central case would need. AI, in Cushman's phrase, "has become a rock to the data center windshield". And the excess is largely American. Cushman's global comparison, published in May, puts total capacity in the Americas, once everything under construction and planned is built, at six times today's, against little more than twice in Europe and Asia, and its count leaves out early-stage projects.¹⁵

REGION_MULTREGION_MULT · Capacity by status, GW · Cushman & Wakefield, Global Data Center Market Comparison (May 2026)

The Excess Is American

Data-centre capacity by status, gigawatts, and total as a multiple of today's. Cushman & Wakefield, May 2026.

Reported

Multiple = (operating + under construction + planned) ÷ operating. Pipeline excludes early-stage projects.

PAPER_GWPAPER_GW · 2025–2032 · IEA, Key Questions on Energy and AI (2026); Van Nieuwerburgh (2026); Cushman & Wakefield (2026)

Paper Gigawatts

US data-centre capacity, gigawatts: planned, built by 2032 in the central case, and the IEA-implied average US load in 2030.

Reported and Modelled

That is the first gap, and it is wide. It is not a forecast of a tenfold surplus of electricity. It is a measure of how far the construction being financed has run ahead of independent forecasts of use, realistic timetables and demonstrated demand. And the Catch-22 survives either outcome: if demand disappoints, the buildings arrive too soon; if demand accelerates, they arrive too slowly. Some of it is ordinary development optimism. Some of it is speculation; Cushman reports that the industry now talks of "phantom" or "fake" data centres, land bought in the hope of finding a tenant, and that some states and utilities have begun charging fees to stay in the queue for power.¹³ And some of it is double counting, as the same gigawatt appears in a developer's pipeline, a utility's connection queue and a hyperscaler's plan.

I call the result paper gigawatts: capacity that exists in leases, purchase orders, announcements and queues, but not on the grid. The industry has its own name for the speculative part. It calls them "phantom" data centres, and Cushman now gives more weight to the preleased part of the planned pipeline when judging a market's prospects.¹³

There is a third reason to discount the plans: approval. In the main markets, Cushman expects projects already under construction to deliver, "while tighter regulation will reduce approvals for new developments and keep growth incremental", and it records cancellations of developments that had already been approved.¹⁵ The plans are largely unleased, concentrated where the grid is growing slowest, and increasingly unlikely to be approved as drawn. Even approval is no guarantee of a building. In Loudoun County, the heart of the industry, 35.7 million square feet of data centres is approved but not yet built, nearly two-thirds as much again as the 56.5 million built or under way, and about 115 more buildings by the county's own estimate.¹⁶ Approval is a permission, not a power connection. Nscale’s prospectus shows the same gap inside a single company: $103.4 billion of contracts, of which $2.6 billion are tied to capacity actually running; 55 megawatts live, most of it in other operators’ buildings, against about 1.37 gigawatts active and contracted. Microsoft and Anthropic account for about 85% of the backlog. The financing for Anthropic’s West Virginia campus had yet to be secured when the company filed, and the contract itself lets either side walk away if it is not secured by a deadline, and lets Anthropic cancel delayed capacity without paying for it.¹⁷

Paper gigawatts would be harmless if they stayed on paper. The difficulty is that a growing share of them has been financed. Four hyperscalers alone carry about $904 billion of future lease payments and $1.52 trillion of purchase commitments, by the Wall Street Journal's count.¹⁴ A purchase commitment is a signed, largely non-cancellable obligation to take and pay for equipment, capacity or services that has not yet reached the balance sheet. The printed figure is a floor on what has been promised and a ceiling on the computing it will buy, because the same demand appears again one tier up, in the commitments of the chipmakers supplying it. Add the tiers together and the same gigawatt is counted several times.

Even so, the order of magnitude is clear. Roughly $2.4 trillion of obligations at four companies, and far more across the chain, rest on a single assumption: that the planned gigawatts will be built, energised and filled with paying customers on schedule.

Who Is Forecasting the Demand?

Before going further, it is worth asking a simple question. Who is actually forecasting the demand that these gigawatts are meant to serve?

My working belief, when this research began, was that much of the demand had not been forecast at all, but assumed: inferred from the spending itself, or from the players' own statements, and then quoted back as proof. That belief deserves testing rather than assertion. Sorting the published numbers by what they measure is the test.

Most of what is described as AI demand is one of three things. The first is spending plans: the hyperscalers' capital expenditure guidance, now running at about $800 billion a year, more than their combined operating cash flow for the first time;¹⁴ McKinsey's estimate that the industry will need about $5.2 trillion of data-centre investment by 2030 for AI alone;¹⁸ and the order books of the chipmakers. These are forecasts of supply being bought, not of demand being served. The second is energy projections, such as the IEA's, which model electricity use, not revenue. The third is usage counts, chiefly tokens, which this series has already shown are a poor guide to value.

The forecasts that model what customers will actually pay are fewer, and they are the ones that show the gap. Gartner expects about $3.6 trillion of AI infrastructure revenue through 2030, against the $8.5 trillion it estimates is needed to earn a 12% return.¹⁹ Van Nieuwerburgh calculates that the capacity built by 2032 in his central case would need about $3.7 trillion of revenue a year.¹⁴

When Nvidia reports that "demand is accelerating", it is describing its order book.²⁰ When a hyperscaler raises its spending plan, it is responding partly to its rivals' plans. When a laboratory signs for more compute, it is responding to the capacity its competitors have secured. Each is rational. Together they are circular: the order book is cited as evidence for the demand that justified the order book. The players are, to a large degree, forecasting each other.

None of that proves the demand will not come. It means the plans rest less on a forecast of what final customers will pay than on a forecast of what competitors will build.

Three Paths for Demand

None of this proves the plans are wrong. The IEA's central case is a forecast, and forecasts of new technologies have a long record of being too timid. So it is worth setting out what the gigawatts would need to believe.

The S-curve: adoption follows the familiar pattern of rapid growth and then saturation, and efficiency gains, which the IEA describes as improving at a rate unprecedented in energy history, keep cutting the power each task requires.¹¹ Demand lands close to the IEA's central case. The planned pipeline is largely paper.

The base path: demand grows strongly, efficiency offsets part of it, and AI-focused consumption roughly triples by 2030, as the IEA projects. Much of what is under construction is needed. Much of what is planned is not, or not yet.

The runaway path: agents, reasoning models and video generation consume hundreds or thousands of times more energy than a simple text query, and cheaper intelligence unlocks uses no one has yet imagined. This is Jevons's paradox: in the nineteenth century, more efficient steam engines increased Britain's consumption of coal rather than reducing it. On this path the planned pipeline is too small, and the constraint is not demand but delivery.

The runaway path has one more engine, and it is the one the bulls cite most: robotics. In August, Musk's case was that the market still underestimates the compute that agents, autonomous vehicles and robots will eventually consume. Nvidia now describes "physical AI coming online".²⁰ If machines that move through the world become as common as the software that runs on screens, today's plans may prove too small. Two points temper it. Robots need data-centre compute to be trained and simulated, but much of their day-to-day intelligence will run on chips inside the machine, not in a distant hall. And the timetable for reliable physical autonomy is long, measured in years of testing and regulation. Robotics strengthens the long-run case for compute far more than it fills the buildings being financed today. Markets, for now, price it mainly through a handful of stocks.

Beyond the robots sits the furthest claim of all: recursive self-improvement, AI that improves the systems that build the next AI, each round faster than the last. It is I. J. Good's "intelligence explosion", first described in 1965.²¹ AI already helps write code and tune models, but that is not a system improving itself without limit, and nothing yet shows the second. Suppose it arrives. It would speed up the technology clock, not the physical one. Every round of self-improvement runs on more compute, which means more power, more transformers and more copper, delivered no faster than before. Deutsche Bank has priced it. A training run of the scale some laboratories expect to need for human-level AI would take about 6 gigawatts and $300 billion to $500 billion of computing hardware; done in four months on Nvidia’s newest chips, 7.8 million of them, 26 gigawatts and $900 billion. Recursive self-improvement may speed up the research. It does not speed up the grid: connection queues, turbines and transmission are “measured in years, not the months it takes to assemble and air-freight the latest Nvidia chip.”²² And it would make each generation of chips obsolete sooner and intelligence cheaper, sharpening the other two legs of the Catch.

The runaway path is the bull case, and it deserves to be taken seriously. But notice what it implies. If demand really does run away, the bottleneck moves from the customer to the supply chain, and the build becomes too slow. If it does not, much of the build is too soon. The plans are financed as though both risks could be avoided at once.

Either way, the next question is the same. What does it actually take to build a gigawatt?

Anatomy of a Gigawatt

A data centre is usually photographed from the air: a grey box the size of several football pitches, a car park and a substation. It looks like a warehouse. It is closer to a power station that consumes its own output, wrapped around the most complex manufactured objects ever made.

Follow the electricity in from the grid and out as heat, and the bill of materials looks like this.

The Land and the Licence

It starts with a site that can be connected. Land without power is farmland. Cushman & Wakefield now publishes a hierarchy of "powered land" for that reason: raw land at the bottom, then land with a secured grid connection, then zoning, fibre and ground conditions, and finally "true powered land" at the top, where every risk has been cleared and buyers bid against each other. In primary US markets, powered land averaged $584,000 per megawatt this year, 51% more than a year earlier.¹³ A gigawatt site therefore costs over half a billion dollars before anything is built on it, and the price is mostly for the right to wait in a queue. "The market has fully priced in power," Cushman writes, "presenting a tradeoff: Buyers either pay a premium to shed risk by buying powered sites or pursue riskier non-powered land to chase the upside."¹³ Jupiter took the second path: no grid connection, a power plant of its own, and a pipeline still awaiting permits. Even powered land is not risk-free; Cushman notes that energisation "delays are surfacing in highly constrained markets".

The transactions show what the premium buys. Among the largest data-centre land deals Cushman records since mid-2025, disclosed prices range from about $9,000 an acre, for Meta's site at Holly Ridge, Louisiana, to about $330,000 an acre, for Digital Realty's at De Soto, Kansas: a 37-fold spread for land that is, as soil, much the same.¹³ The largest disclosed price was paid by Stack Infrastructure, the Blue Owl unit developing Jupiter: $737.8 million for 2,990 acres at Danville, Virginia. Hyperscalers, Cushman adds, paid roughly seven times more per acre than land developers. The giants buy certainty. The developers buy options.

LAND_PRICELAND_PRICE · US land purchases, 2025 to 2026 YTD · Cushman & Wakefield (2026); author's calculation

Paying for Power, Not Soil

US data-centre land purchases, 2025 to 2026 YTD.

Reported; per-acre figures calculated

Then come the permits: zoning, air permits for any on-site generation, water rights and, increasingly, the consent of the neighbours. That last item is no longer a formality, as later sections show.

The Power Chain

Electricity arrives at transmission voltage and is stepped down several times before a chip can use it. Each step is a separate piece of heavy equipment, and each has its own factory, order book and queue.

• The grid connection and substation. High-voltage power transformers, circuit breakers and protection equipment, usually built for or by the utility. Some high-voltage transformers now carry lead times of up to 160 weeks.²³
• Medium-voltage switchgear and distribution transformers. The equipment that divides and routes power around the campus. Cushman's survey puts medium-voltage switchgear at 38 to 63 weeks and pad-mounted transformers at 68 to 113 weeks.¹³
• Uninterruptible power supplies and batteries. A computing hall cannot drop power even for a fraction of a second, so every megawatt sits behind banks of batteries and power electronics that bridge the gap until backup power starts.
• Backup generators. Diesel generators, typically two to three megawatts each, sized to carry the whole load if the grid fails. A gigawatt campus built to full redundancy can need several hundred of them. Large generators now take around 110 weeks to arrive in the Americas, up from 60 to 70 weeks in 2024.²⁴
• Busways, cabling and power distribution units. Miles of copper and aluminium conductors carrying power from the switchroom to every rack.
• On-site generation, where the grid cannot deliver. Gas turbines, reciprocating engines or, at Jupiter, fuel cells. The IEA estimates that serving a critical load from on-site gas can require 30% to 70% more installed capacity than the load itself.¹¹ Global turbine orders reached about 100 gigawatts by the end of 2025, against manufacturing capacity of 60 to 70 gigawatts a year.²⁵

Cushman's cost survey puts this power infrastructure at 21% of a greenfield facility's cost: the largest single category, "a logical outcome," as the guide puts it, "since without power there is no compute."¹³

The Cooling

Almost every watt that enters a data centre leaves it as heat, and AI chips are now dense enough that air alone cannot remove it. The cooling plant includes chillers, cooling towers and dry coolers, pumps, piping and valves, heat exchangers, and coolant distribution units that pump liquid directly to plates mounted on the chips. Some designs draw heavily on water; closed-loop systems, such as Jupiter's, recirculate it. Cooling is 12% of the facility cost in Cushman's survey, and cooling equipment prices have risen 32% since December 2021.¹³

The Building

The shell is the least exotic part, and still enormous: structural steel, concrete foundations and slabs, roofing, sitework and roads, security, fire suppression, and the fibre routes that connect the campus to the outside world. Construction of the core and shell is 17% of the facility cost.¹³ It also consumes skilled labour: electricians, pipefitters and mechanical trades, whose wages have risen about 5% a year since 2021 and which Cushman finds in short supply in many of the markets where land and power are cheapest.¹³

The Contingency

The second-largest line in Cushman's breakdown is not a piece of equipment. It is contingency, 16% of the facility cost, with a further 3% for escalation.¹³ That is the industry's own estimate of what it does not yet know: late deliveries, price rises during procurement, utility changes and new rules. Nearly a fifth of the budget is set aside for surprises before the first surprise arrives.

What the Facility Costs

Put together, Cushman puts the all-in cost of a greenfield facility at $8.9 million to $23.3 million per megawatt, excluding the chips. The high end is for fully redundant facilities built for AI and high-performance workloads, which is what most of the new build is.¹³ Costs per megawatt have risen 21% since late 2024, against general inflation of about 5% over the same period.¹³

That is only the building. The expensive part goes inside it.

The Chips

A modern AI data centre is filled with racks, each a tightly packed cabinet of accelerators, processors, memory, networking and power and cooling hardware. The newest racks draw well over 100 kilowatts each, perhaps fifty times a conventional server rack, so a gigawatt of computing capacity means several thousand of them.

Each accelerator is itself the end of a chain of manufacturing that crosses half the world and takes the better part of a year.

The design. The chip is designed by Nvidia, AMD, or increasingly by the hyperscalers themselves with partners such as Broadcom. Design takes years, and the leading designers now release a new generation roughly every year.

The lithography. The circuits are printed on silicon wafers using extreme ultraviolet lithography, which projects patterns of light onto the wafer at a scale measured in nanometres. One company, ASML in the Netherlands, makes the machines. Each costs a few hundred million dollars and is shipped in many containers. Without them, there are no leading-edge chips.

The wafer. The wafers are made in fabrication plants, overwhelmingly at TSMC in Taiwan. A leading-edge fab costs tens of billions of dollars and takes years to build. Each wafer passes through hundreds of process steps over roughly three months before the chips on it are finished. The fabs also consume ultra-pure water, speciality gases and chemicals in quantities that make a data centre look frugal.

The memory. AI accelerators are surrounded by high-bandwidth memory, stacks of memory chips bonded vertically and placed alongside the processor. Three companies make it at scale: SK Hynix, Samsung and Micron. It has been the tightest component of all. Memory prices rose more than five times in a year, by Bloomberg's account.²⁶

The packaging. The processor and its memory stacks are mounted together on a silicon interposer and a substrate, a step known as advanced packaging. TSMC's version, CoWoS, has been a bottleneck in its own right, and the substrates depend on a specialised insulating film supplied by very few manufacturers. A chip that has been printed but not packaged cannot be used.

The system. Packaged chips are assembled onto boards, joined to processors and networking, and built into racks by contract manufacturers. Tens of thousands of accelerators must then be wired together: switches, optical transceivers and copper cables in volumes the networking industry has never before supplied.

Each link in that chain has a small number of suppliers, long lead times and capacity that takes years to expand. Most of the value sits here. In Bloomberg's estimate for a one-gigawatt campus at Abilene, Texas, chips and servers account for about half of a total cost of roughly $30 billion. Van Nieuwerburgh's all-in estimate is about $41 billion per gigawatt, with information technology 68% of it.^{14 26}

Four Estimates, Four Definitions

The cost of a gigawatt is quoted everywhere and defined almost nowhere. The estimates differ mostly because they count different things.

The honest answer is somewhere between $30 billion and $40 billion for a modern AI gigawatt, built and equipped, with the facility the smaller part and the chips the larger. Multiply by the gigawatts in the plans, and the sums run into trillions.

GW_ANATOMYGW_ANATOMY · Illustration · Cushman & Wakefield (2026); Bloomberg (2026); Van Nieuwerburgh (2026)

Anatomy of a Gigawatt: $37.6 Billion

Initial capital for one AI gigawatt, $ billion, by component, from grid to chip. Reference case.

Modelled

There is one more thing to notice about that bill of materials. The building lasts for decades, the lease for fifteen or twenty years, and the debt for five or more. The chips inside it are superseded every year or so. The most expensive component has the shortest commercial life. We will return to that in Part II.

The Materials

A gigawatt is also a quantity of metal, and every paper gigawatt implies an order for it. Translating the plans into materials was always the point of measuring in power, because materials have suppliers, and suppliers have limits.

The list is longer than most people imagine. Follow the electricity in again, layer by layer.

The Power Chain

• Copper in transformer windings, busbars, cables, switchgear and grounding, in every building and in every line that feeds it.
• Aluminium in high-voltage conductors, busways, cable and enclosures, and increasingly as a cheaper substitute for copper where weight and space allow.
• Grain-oriented electrical steel, the specialised steel that forms the core of every transformer. It is made by very few mills worldwide, one of the quieter reasons transformers are scarce.
• Silver in the contacts of breakers, relays and switches, where its conductivity and resistance to wear are hard to replace.
• Tin in solder, zinc for galvanising steel, and insulating oils, papers and gases in the transformers and switchgear.

Backup Power and Storage

• Lead in the lead-acid batteries still common in uninterruptible power supplies, and lithium, nickel, cobalt, manganese and graphite in the lithium-ion systems replacing them and in the battery banks now built beside many campuses.
• Nickel and cobalt superalloys in the hot sections of gas turbines, which is one reason only a handful of firms can build them.
• Diesel for hundreds of standby generators, and natural gas, by pipeline or now by truck, for on-site generation.
• At Jupiter, fuel cells, which bring their own ceramics and rare materials.

Cooling

• Copper and aluminium in heat exchangers, cold plates and coils.
• Stainless steel in pumps, valves and piping.
• Rare earths (neodymium, praseodymium, dysprosium and terbium) in the permanent magnets of the motors that drive fans and pumps.
• Water, in some designs a great deal of it.

The Building

• Steel for the frame, concrete (cement, sand and aggregate) for foundations and slabs, glass, roofing, and the fibre-optic routes that connect the campus to the world.

The Chips and Servers

• Silicon, refined to purities measured in nines, grown from high-purity quartz.
• Copper, cobalt, tungsten and tantalum in the microscopic wiring and barriers inside each chip; hafnium in its transistor gates.
• Gold, silver, tin and palladium in bonding, plating, solder and connectors.
• Gallium in the gallium-nitride power electronics that feed the chips; germanium in fibre optics; indium in the indium-phosphide lasers inside optical transceivers.
• Helium, neon and fluorine-based gases, consumed in the fabs that make the chips.
• High-purity alumina and speciality ceramics in substrates and packaging.

The International Energy Agency has made the first serious attempt to size the demand. By 2030, it estimates, data centres could consume more than half a million tonnes of copper and 75,000 tonnes of silicon a year, about 2% of world demand for each, and more than a tenth of the world's gallium.²⁷ Those shares sound modest. Two features of the supply make them anything but.

The first is concentration. China controls 80% to 90% of global refining of silicon, gallium and rare earths. It restricted exports of gallium and germanium in 2023, and since late 2024 has added curbs on tungsten, tellurium, bismuth, indium and molybdenum, all of them inputs to microprocessors and server hardware.²⁷ The same handful of turbine, transformer, memory and packaging suppliers recur at every stage.

The second is inelasticity. Silver is the clearest example. About three-quarters of the world's mined silver comes as a by-product of mining copper, lead, zinc and gold, and its primary share has fallen to a new low of 26%.²⁸ A higher silver price does not, by itself, produce much more silver. Data centres, now more than a tenth of electrical and electronics demand, are not a large enough buyer to be the bottleneck for the build. They are a price-insensitive buyer at the margin of a market already in its sixth year of deficit, which is a different and more interesting thing.²⁸ A developer racing to energise a campus does not stop to haggle over the silver content of its switchgear.

GW_MATERIALSGW_MATERIALS · Illustration · IEA; IMF Finance & Development; Silver Institute; Cushman & Wakefield

What a Gigawatt Is Made Of

One gigawatt in five layers, from grid connection to chip, with principal materials and sourced supply constraints.

Illustration

Scarcity raises prices. Higher prices do not quickly produce more supply. And the buyers, spending borrowed money against a timetable, are the least price-sensitive these markets have ever had. Of all the materials on the list, one matters more than the rest, because it runs through every layer, from the transmission line to the chip.

Why Copper Cannot Arrive in Time

Robert Friedland, who has spent half a century finding copper, put the scale of the problem bluntly at the Mining Forum Americas last month: "Just to maintain three percent GDP growth, absent the AI phenomenon, data centers, electric cars, we have to mine the same amount of copper in the next eighteen years as we mined in the last ten thousand years."²⁹ The qualification is the alarming part. He has been making the same point for a year, at the University of Southern California and to BMO's mining conference.

It is the kind of line that sounds like a salesman's. Friedland, after all, runs a copper miner. But the evidence behind it does not come from him, and it is worth setting out, because copper is where the physical clock runs slowest of all.

How Much a Gigawatt Needs

Estimates of copper per gigawatt vary with what is counted. BMO puts it at about 35,000 tonnes per gigawatt of data-centre capacity, a figure Robert Friedland cited this week.²⁹ Much of it sits not inside the building but in the transmission and distribution lines that feed it. The IEA's narrower count, inside the data centre itself, reaches more than half a million tonnes a year by 2030.²⁷

Apply that to the plans and the arithmetic is uncomfortable. Van Nieuwerburgh's central case of 182.7 gigawatts built by 2032 implies about 6.4 million tonnes of copper. The full 509 gigawatts of US plans would need about 18 million tonnes, three-quarters of a year of the world's entire mine output, about 23.5 million tonnes in 2026 on Cochilco's estimate.³⁰ And that output is barely growing; the International Copper Study Group now expects world mine production to rise just 1.3% this year, revised down from 1.6% in April, held back by Grasberg and Kamoa-Kakula.³¹ Paper gigawatts become paper tonnes.

Falling Grades

The first reason supply cannot simply rise to meet that is geology. The richest, shallowest ore was mined first.

Chile, which produces about a quarter of the world's mined copper, has seen its weighted average ore grade fall from 1.13% in 2002 to 0.62% in 2024, according to the Chilean Mining Council, a decline of nearly 45% in little more than two decades.³⁰ Across the world, the IEA puts the average grade of copper being mined 40% below its 1991 level.³²

A lower grade means more rock moved, crushed and processed for every tonne of metal, and so more energy, more water and more capital simply to stand still. Friedland again: since 1900, he says, the energy needed to produce a unit of copper has risen sixteenfold, and the water needed has doubled.²⁹ Escondida, the largest copper mine in the world, expects its output to fall by about 300,000 tonnes to 1.1 million tonnes in 2027 as its grades decline.³⁰ The industry is on a treadmill, running harder to stay in the same place.

CU_GRADECU_GRADE · 2002 and 2024 · Consejo Minero (Chilean Mining Council), as reported by Reporte Minero and UPI; IEA (Mar 2026)

Running Harder to Stand Still

Chile's average copper ore grade, % copper. 2002 and 2024.

Reported

Fewer Discoveries, Longer Lead Times

The second reason is time. S&P Global finds that mines which began production between 2005 and 2009 took 12.7 years on average from discovery to first output. Mines that started between 2020 and 2023 took 17.9 years, lengthened by exploration, permitting, studies, financing and construction permits.³³ Major new discoveries have become scarce; the IEA notes that only 5% of the copper deposits found in the last 35 years were discovered in the last decade.³² Grassroots exploration fell to 25% of base-metal exploration budgets in 2024 and 2025, an all-time low.³⁴

The consequence follows directly. S&P expects global mine supply, including every operating mine and every project in the pipeline, to peak at 27.3 million tonnes in 2030 and decline to 25.1 million tonnes by 2035.³⁴ The IEA projects a primary supply shortfall of about 30% by 2035 on current policies.³²

A copper mine discovered today would, on the recent average, begin producing around 2044. The data centres financed this year expect to be running by 2028.

CU_LEADTIMECU_LEADTIME · Mines by start period · S&P Global Market Intelligence

The Copper Clock

Years from discovery to first production for copper mines, against the grid connection wait and the data-centre build.

Reported · last row Illustrative

Fragile Supply

The third reason is fragility. A concentrated supply base is exposed to single events. In September 2025 a mud rush at Grasberg in Indonesia, one of the world's largest copper mines, killed workers and forced Freeport to declare force majeure; the main block cave is being restarted in phases through 2026 and 2027.³⁵ In May 2025, seismic activity flooded parts of Ivanhoe's Kamoa-Kakula in the Democratic Republic of Congo; it is still producing, but dewatering and recovery continue. Cobre Panamá, which produced 350,000 tonnes in 2022, has been closed since November 2023; Panama approved processing of stockpiled ore in April 2026, which First Quantum stresses is not a reopening of the mine.³⁵ Each disruption is a reminder that the supply curve has no slack.

The Strongest Counter

There are good answers to all of this, and they deserve a hearing.

Data centres are still a small part of copper demand, about 2% by 2030 on the IEA's count. Prices ration demand: at high enough prices, other buyers defer and aluminium replaces copper wherever it can. Recycling is rising. And the industry has met large demand shocks before, in China's boom of the 2000s. And copper wire itself still arrives in four to eight weeks.¹³

All true. But notice what each answer implies. Rationing by price means the data centres outbid someone else, which raises their costs further. Substitution has limits in transformers, motors and chips. Recycling cannot supply metal that has not yet been used. Wire arrives quickly because it is drawn from metal already mined, at whatever price clears, and copper is already 46% dearer than in 2021. A short delivery time says nothing about the metal not yet mined. The constraint is not this year's wire but the next decade's mines. And the China boom was met over a decade, by mines discovered in the 1990s; the discovery pipeline that would serve this boom is the thinnest in a generation.

So the question is not whether copper will be found for the data centres. At a high enough price, it will. The question is when, and at what cost to a project financed on 2025 prices and a 2028 timetable. BMO has raised its long-term copper price forecast to $18,000 a tonne.²⁹ Every point of that rise is a cost the builders did not budget for, paid in a metal that cannot arrive in time.

The Critical Path

Engineers plan a large project around its critical path: the longest chain of tasks that must happen in sequence, which sets the earliest possible finish. Everything else can run in parallel. For an AI data centre, the critical path is rarely the building. It is the power.

The Queue for Electricity

Before a site can draw a gigawatt, the utility must study the request, build or upgrade substations and transmission lines, and connect it. Cushman's global survey puts the average expected wait for energising a new large load at about five years in the Americas, and 4.4 years globally.¹⁵ Some powered land now comes with delivery dates of 2033 or later.¹³ That is longer than many of the loans financing the buildings that will wait for it. Cushman's own rankings make the mismatch plain: the ten markets with the most capacity planned are all in the Americas, and not one of the ten with the fastest-growing electricity production is in the United States.¹⁵

Britain shows how far the queue can stretch. Nscale's £2 billion site at Loughton in Essex, billed as the country's largest AI supercomputer, with more than 23,000 Nvidia GPUs and Microsoft as anchor tenant, won planning permission and a 90-megawatt grid connection, then learned that the power would not arrive until the early 2030s.³⁶ On 18 October the Telegraph reported that Microsoft had cancelled the deal, as Nscale prepares for a $35 billion listing.³⁷ Chips bought for 2027 and a tenant gone before the electricity arrives: too slow, in a single site.

The hyperscalers are already paying for the slow clock. Morgan Stanley estimates that a substantial portion of this year’s capital spending, “in some cases 30-40%”, is going on land, buildings and power for data centres that will not open until 2027, 2028 or 2029.³⁸

The Queue for Equipment

Then there is the equipment, each item with its own order book:

• Large generators: 60 to 100 weeks, and about 110 in some surveys.^{13 24}
• Pad-mounted transformers: 68 to 113 weeks; some high-voltage transformers up to 160 weeks.^{13 23}
• Medium-voltage switchgear: 38 to 63 weeks.¹³
• Uninterruptible power supplies: 36 to 42 weeks.¹³
• Gas turbines: orders running well ahead of a year's manufacturing capacity.²⁵
• Accelerators, memory and packaging: allocated by suppliers, with a new generation roughly every year.

These lead times do not simply add up. A well-run project orders equipment while the building rises and the utility works. The longest single path sets the date. But the longest path is now measured in years, and the queue has favourites. Buyers with long supplier relationships and a large annual spend, Cushman notes, "often secure significantly shorter lead times" than new entrants.¹³ The longest waits fall on the newcomers, the neoclouds and former miners, who borrowed the most and can least afford to wait. Labour is a queue of its own; the shortage of electrical and mechanical trades, Cushman notes, is producing "higher-than-expected wages and longer project timelines".¹³

LEAD_TIMESLEAD_TIMES · Weeks, order to delivery · Cushman & Wakefield (2026); Gilbane

The First Queue: Equipment

Lead times from order to delivery, weeks, quoted range. North America, 2026.

Reported

The Neighbours

A newer item has joined the critical path: consent. Even at the centre of the industry, the queue is now political. Loudoun County in Virginia, home to Ashburn's "Data Center Alley" and the largest concentration of data centres in the world, has about 56.5 million square feet of them built or under construction, and a further 35.7 million approved but not yet built.¹⁶ In September its board voted to move towards pausing, for up to twelve months, its final votes on data-centre and substation applications that need its approval, the longest pause Virginia law allows; the county attorney had advised that an outright moratorium would be illegal. The formal resolution comes before the board on 20 October.^{16 39} It is not a ban; existing approvals stand and no application lapses. But in the market that built the industry, new approvals may now wait a year. This month a Virginia state senator proposed halting approvals statewide for three years from July 2027.³⁹ Statewide moratoriums have been introduced in New York, Texas and Oregon, and Arizona's attorney general has called for a pause.³⁹ Data Center Watch counted 45 projects worth $68 billion blocked or delayed by local opposition in the second quarter alone.⁴⁰ Cancellations are now running faster than blockages; Heatmap counts at least $260 billion of planned data-centre investment cancelled this year after sustained local opposition, roughly half of it in the three months to September.⁴¹

The industry has noticed. Cushman's own cost guide now calls community engagement "central to site viability, schedule certainty and capital planning".¹³ The neighbours have joined the grid and the transformer factories on the critical path. The industry's answer is engineering and better rules: concerns over noise, water, emissions and the grid are, Cushman argues, "increasingly manageable through proven engineering solutions", and are being written into zoning and permitting standards. Over time that should make approvals more predictable. In the meantime, each new standard adds a condition to the timetable and a line to the budget.

The Timetable in Practice

The evidence is mixed, and fairly so. The flagship Stargate campus at Abilene had delivered 618 megawatts of its 1.2 gigawatts by September, roughly six months behind its permits' original timetable.⁴² Nscale's flagship site in Norway is six months late, after cooling leaks, regulatory delays and a fire.⁴³ Jupiter has its buildings and no power.

Yet Applied Digital brought another 75 megawatts into service in North Dakota this month, taking its first campus to 250 of a contracted 400 megawatts. Its chief executive put the lesson plainly: “Securing power is only the starting point.” The builders delivering are the ones who secured their power and equipment early.⁴⁴ Galaxy delivered the first phase of its Helios campus in Dickens County, Texas, on time.⁴⁵

That is not a contradiction. It is the critical path at work. The builders who secured their place in the equipment queue early are delivering. Those who did not are waiting. The queue itself is the bottleneck, and it is longest for the newest entrants with the most borrowed money.

The September series used the London Underground to describe how a fault on one line spreads across the network. The same picture describes the timetable. Every line on a data-centre project runs to its own clock. The chip line runs a new train every year. The power line, held at signals for turbines and transformers, runs every two or three. The construction loan runs for five, the tenant's contract for three, and the building's lease for fifteen or twenty. Every financing in this boom assumes the connections will be made. When the power train is late, the revenue misses the debt.

CRIT_PATHCRIT_PATH · Illustration · Cushman & Wakefield; Linesight; Terrapin; Wood Mackenzie; Reuters

The Critical Path Runs Through the Grid

Months from financing for a large data centre, by task. Equipment from order to delivery.

Reported and Illustrative

The Constraint That Does Not Bind

There is a strong objection to all of this, and it deserves its full weight.

Cushman & Wakefield, whose data runs through this section, concludes that an AI infrastructure bubble is unlikely. Its reasoning is that demand still exceeds supply, colocation vacancy is extremely low (4.2% in the Americas, against 7.8% in Europe and 10.9% in Asia), and power shortages, permitting delays, congested queues and moratoriums make overbuilding physically impossible. Preleasing on capacity under construction in the Americas is 79%, and 89% including the hyperscalers' own builds.¹⁵ The constraints, in other words, are the industry's protection.

Cushman's own data shows where that protection stops. What is being built is let, as those figures show. What is planned largely is not. In West Texas, about 8 gigawatts of plans are little more than a third preleased; in Pennsylvania, 7.5 gigawatts are less than a fifth.¹⁵ The largest pools of planned capacity are the least committed. The paper gigawatts are precisely the ones without tenants.

PRELEASEPRELEASE · Planned capacity, top 30 markets · Cushman & Wakefield, Global Data Center Market Comparison (May 2026)

Built Is Let. Planned Is Not.

Planned capacity in Cushman's top 30 markets: size against share preleased.

Reported; rates approximate

Rates read from Cushman's chart; approximate. Markets with at least 50 MW planned.

So Cushman is right about the buildings going up. It is not right about the money.

The same constraint that stops the buildings being overbuilt does nothing to stop the financing being overcommitted. Leases, purchase commitments and loans are signed on a timetable. The supply chain delivers on a different one. A shortage of transformers does not reduce the interest on a construction loan; it extends the period over which the interest is paid without revenue. Scarcity protects the owner of an energised hall. It punishes the owner of an unfinished one.

Cushman draws the line itself. "It is important to distinguish," its report says, "between an AI data center infrastructure bubble and a broader AI investment bubble." The first concerns buildings; the second, valuations. Cushman rules out only the first. That is precisely the gap this series is about.

It even concedes a "non-zero chance that revenue growth plateaus or dips sooner than expected."¹⁵ That is the second leg of the Catch, in the industry's own words.

So this is where Part I leaves us. The plans call for far more capacity than the measured demand requires. Each gigawatt of it is a power station, a cooling plant, a building and a warehouse of the most sophisticated chips ever made, assembled from inputs that are scarce, concentrated and slow to expand. The physical clock runs in years. The financial clock started when the money was lent.

The build is too slow for its financing, even if every gigawatt of demand proves real.

Part II follows the bill: what happens to the cost when the timetable slips, what our model of a single 100-megawatt hall shows, and how three real projects, Jupiter among them, distribute the same shock to very different people.

The technology can be right. The financing can still be early.

Ben Davies
Co-founder · Hinde Capital · London, 19 October 2026

Sources

1. 1Searchlight New Mexico, Oracle touts Project Jupiter benefits during tour as wrangling over power continues, 16 Oct 2026, https://searchlightnm.org/oracle-touts-project-jupiter-benefits-during-tour-as-wrangling-over-power-continues/.↩
2. 2Bloomberg, Oracle Asks Regulators to Expedite Gas Pipeline for Data Center, 13 Oct 2026, https://www.bloomberg.com/news/articles/2026-10-13/oracle-asks-regulators-to-expedite-gas-pipeline-for-data-center; Source New Mexico, Proposed gas pipeline for Project Jupiter data center delayed to 2027, 17 Aug 2026, https://sourcenm.com/2026/08/17/proposed-gas-pipeline-for-project-jupiter-data-center-delayed-to-2027-filings-show/.↩
3. 3Source NM, Oracle says it will cut Project Jupiter emissions after New Mexico org seeks to toss permit case, 15 Oct 2026, https://sourcenm.com/2026/10/15/oracle-says-it-will-cut-project-jupiter-emissions-after-new-mexico-org-seeks-to-toss-permit-case/ ; Albuquerque Journal, https://www.abqjournal.com/news/new-challenge-to-project-jupiter-air-permit/3143387 (projected permitted emissions from about 8.8 million to 5.7 million tons a year, per Oracle's 15 Oct announcement, https://www.oracle.com/news/announcement/project-jupiter-microgrid-cuts-permitted-emissions-by-35-percent-2026-10-15/ ; hearing on a motion to dismiss expected 26 Oct 2026, per press reports; NMED docket AQB 26-57(P), https://www.env.nm.gov/opf/docketed-matters/)↩
4. 4Oracle, Project Jupiter microgrid cuts permitted emissions by 35 percent, 15 Oct 2026, https://www.oracle.com/news/announcement/project-jupiter-microgrid-cuts-permitted-emissions-by-35-percent-2026-10-15/ ; El Paso Matters, Oracle could defer Project Jupiter rent payments over power supply delays, 24 Sep 2026.↩
5. 5Capacity, What Oracle's force majeure notice means for AI data centre risk, 2 Oct 2026, citing Bloomberg, https://capacityglobal.com/news/165bn-project-jupiter-what-oracles-force-majeure-notice-means-for-ai-data-centre-risk/.↩
6. 6Ares Management, In the Gaps, Alternative Credit Newsletter, Fall 2026; Reuters, Oracle's $18 billion data center debt under pressure, FT reports, 18 Sep 2026, https://kfgo.com/2026/09/18/oracles-18-billion-data-center-debt-under-pressure-ft-reports/.↩
7. 7Vercel, AI Gateway leaderboards, Sep 2026, https://vercel.com/ai-gateway/leaderboards.↩
8. 8CNBC, Intel, Broadcom CEOs warn capital crunch could become AI's biggest bottleneck, 12 Oct 2026, as reported by Benzinga, https://www.benzinga.com/markets/tech/26/10/62295019; Bloomberg and The Business Times, 12 Oct 2026.↩
9. 9US Census Bureau, QuickFacts, New Mexico (837,414 households, 2020-2024); US Energy Information Administration, average annual residential electricity consumption (863 kWh a month, or about 10,400 kWh a year, per residential customer in 2024, EIA table 5A, https://www.eia.gov/electricity/sales_revenue_price/pdf/table_5A.pdf ; rounded to about 10,500 in the author's conversions); Ofgem, typical domestic consumption values (2,700 kWh). Conversions are the author's: 1 MW for a year = 8,760 MWh.↩
10. 10NESO, Winter Outlook 2025/26, Oct 2025 (base-case de-rated margin 6.1 GW, 10% of average-cold-spell peak demand, implying a peak of about 60 GW), https://www.neso.energy/document/369621/download↩
11. 11International Energy Agency, Key Questions on Energy and AI, Apr 2026 (report text: about 950 TWh in 2030; underlying data 945 TWh). https://www.iea.org/reports/key-questions-on-energy-and-ai↩
12. 12International Energy Agency, Energy and AI, Apr 2025 (data centres about 1.5% of global electricity in 2024, rising to about 3% by 2030 in the base case), https://www.iea.org/reports/energy-and-ai↩
13. 13Cushman & Wakefield, Development Cost Guide 2026: Data Centers, North America, 3 Sep 2026, release: https://finviz.com/news/388803/cushman-wakefield-releases-2026-data-center-development-cost-guide-citing-21-rise-in-per-mw-construction-costs.↩
14. 14Stijn Van Nieuwerburgh, Financing the AI Buildout, Brookings Papers on Economic Activity, conference draft, Sep 2026, https://www.brookings.edu/wp-content/uploads/2026/09/4c_Van-Nieuwerburgh.pdf.↩
15. 15Cushman & Wakefield, 2026 Global Data Center Market Comparison, May 2026, https://www.cushmanwakefield.com/en/germany/news/2026/05/global-data-center-market-comparison (regional capacity, vacancy and multiples, p.3: Americas 43.4 GW operating, 25.3 GW under construction, 191.3 GW planned, 6.0x; EMEA 2.3x; APAC 2.4x; pipeline excludes early-stage projects); managed growth and approvals, p.4; infrastructure versus investment bubble, p.6; Texas, p.9; planned prelease rates by market, p.19 (read from chart, approximate); all rankings, p.51. The September Cost Guide (note 13) gives larger, later Americas figures.↩
16. 16Loudoun County, Data Centers: The Loudoun Story (county data as of 1 Mar 2026), https://www.loudoun.gov/6408/Data-Centers-The-Loudoun-Story ; Loudoun County, Board Approves Process that Could Pause Legislative Data Center & Substation Applications, 17 Sep 2026, https://www.loudoun.gov/m/newsflash/home/detail/10874↩
17. 17Nscale Ltd., Form S-1 registration statement, 18 Sep 2026 (total contract value at 31 Aug 2026; facilities table, about 1.37 GW active and contracted IT capacity), https://www.sec.gov/Archives/edgar/data/2110365/000119312526395475/ck0002110365-20260918.htm ; Exhibit 10.25, GPU Services Agreement with Anthropic (sections 6(d), delay termination, and 9(a)-(d), qualifying financing and termination right), https://www.sec.gov/Archives/edgar/data/2110365/000119312526395475/ck0002110365-ex10_25.htm ; live capacity split as reported by Finansavisen.↩
18. 18McKinsey Quarterly, The cost of compute: A $7 trillion race to scale data centers, Apr 2025 ($5.2trn for AI workloads by 2030, within a range of about $3trn to $8trn; $6.7trn including traditional IT), https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-cost-of-compute-a-7-trillion-dollar-race-to-scale-data-centers↩
19. 19Gartner, Gartner Says Over $3.5 Trillion of AI Infrastructure Spending Could Be Exposed to Impairment by the End of 2030, press release, 14 Oct 2026, https://www.gartner.com/en/newsroom/press-releases/2026-10-14-gartner-says-over-us-dollars-three-and-a-half-trillion-of-artificial-intelligence-infrastructure-spending-could-be-exposed-to-impairment-by-the-end-of-2030.↩
20. 20Nvidia, results release for the second quarter of fiscal 2027, 26 Aug 2026, https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000073/q2fy27pr.htm.↩
21. 21I. J. Good, "Speculations Concerning the First Ultraintelligent Machine", Advances in Computers, vol. 6, 1965.↩
22. 22Deutsche Bank Research Institute (Adrian Cox, with Azeem Azhar of Exponential View), AI boom or doom: what will superintelligence feel like and when will we get there?, Thematic Research, 12 Oct 2026, pp.12 and 14.↩
23. 23Reuters, US power companies scramble to secure equipment as surging data center demand strains supply, 9 Jul 2026, https://www.kitco.com/news/off-the-wire/2026-07-09/us-power-companies-scramble-secure-equipment-surging-data-center.↩
24. 24Linesight, Construction Market Intelligence, Americas, Jun 2026, https://insights.linesight.com/cmi-2026-june/industry-trends-and-supply-chain-dynamics-2026-americas/.↩
25. 25Bloomberg, citing Wood Mackenzie, on gas turbine orders and prices, 1 Apr 2026, https://www.bloomberg.com/news/articles/2026-04-01/gas-turbine-prices-surge-crimping-efforts-to-power-data-centers ; Wood Mackenzie, Gas turbine prices soar 195% as market faces supply-demand crisis, press release, https://www.woodmac.com/press-releases/gas-turbine-prices-soar-195-as-market-faces-supply-demand-crisis/↩
26. 26Bloomberg, The Staggering Costs of a Single AI Data Center, Lynn Doan, 12 Oct 2026, https://news.bloomberglaw.com/international-trade/the-staggering-costs-of-a-single-data-center-from-land-to-chips↩
27. 27Thijs Van de Graaf, Inside the AI-Led Resource Race, IMF Finance & Development, Dec 2025, citing IEA, Energy and AI (2025), https://www.imf.org/en/publications/fandd/issues/2025/12/inside-the-ai-led-resource-race-thijs-van-de-graaf ; IMF eLibrary, https://www.elibrary.imf.org/abstract/journals/022/0062/004/article-A010-en.xml.↩
28. 28The Silver Institute, World Silver Survey 2026, https://silverinstitute.org/wp-content/uploads/2026/04/World-Silver-Survey-2026.pdf↩
29. 29Robert Friedland (@robert_ivanhoe) on X, Oct 2026, citing BMO's estimate of 35,000 tonnes of copper per gigawatt, as captured at https://digg.com/ai/xd3xwukb; Robert Friedland, keynote The AI Supply Chain Is a Copper Story, Mining Forum Americas, 28 Sep 2026 (organiser's video and automatic transcript, https://americas.miningforum.com/agenda/2026/speakers/robert-friedland); keynote, USC Marshall Energy Business Summit, 30 Sep 2025, https://www.youtube.com/watch?v=Q5hnwooyKVk ; BMO Global Metals, Mining & Critical Minerals Conference, Feb 2026.↩
30. 30Consejo Minero (Chilean Mining Council) figures, as reported by Reporte Minero, Menores leyes del cobre elevan presión por mejorar productividad minera, 26 Aug 2026, https://www.reporteminero.cl/noticia/noticias/2026/08/menores-leyes-cobre-presion-productividad-minera ; UPI, Declining ore grades pressure Chile's copper production, 2026.↩
31. 31International Copper Study Group, press release, Copper Market Forecast 2026-2027, 14 Oct 2026, https://icsg.org/download/2026-10-14-press-release-icsg-copper-market-forecast-2026-2027/ (mine production +1.3% in 2026, revised from 1.6%; about +3.1% in 2027; refined surplus about 104kt in 2026 and 368kt in 2027).↩
32. 32International Energy Agency, Global Critical Minerals Outlook 2025, https://www.iea.org/reports/global-critical-minerals-outlook-2025; IEA commentary, 2 Mar 2026.↩
33. 33S&P Global Market Intelligence, Major copper discoveries and From 6 years to 18 years: the increasing trend of mine lead times (average lead time 17.8 years for mines starting 2020–24), https://www.spglobal.com/market-intelligence/en/news-insights/research/from-6years-to-18years-the-increasing-trend-of-mine-lead-times ; https://www.spglobal.com/market-intelligence/en/news-insights/research/major-copper-discoveries↩
34. 34S&P Global Market Intelligence, New major copper discoveries remain scarce; late-stage assets drive growth, Dec 2025, https://www.spglobal.com/market-intelligence/en/news-insights/research/2025/12/new-major-copper-discoveries-remain-scarce-late-stage-assets-drive-growth↩
35. 35Freeport-McMoRan, Third-Quarter 2026 Operational Update, 2 Oct 2026, https://investors.fcx.com/investors/news-releases/news-release-details/2026/Freeport-Provides-Third-Quarter-2026-Operational-Update/default.aspx ; Ivanhoe Mines, Q3 2026 production release (76,401 t), 8 Oct 2026, https://www.ivanhoemines.com/news-stories/news-release/ivanhoe-mines-reports-76401-tonnes-of-copper-produced-by-kamoa-kakula-for-q3-2026/ ; First Quantum Minerals, Government of Panama approves processing of stockpiled ore at Cobre Panamá, 7 Apr 2026, https://www.first-quantum.com/news/government-of-panama-approves-processing-of-stockpiled-ore-at-cobre-panama/↩
36. 36Sifted, Nscale's £2bn UK data centre hit by grid delays, 13 Jul 2026, https://sifted.eu/articles/nscale-delays ; IET, UK's largest AI supercomputer delayed to 2030s due to grid issues, https://engx.theiet.org/b/articles/posts/uk-s-largest-ai-supercomputer-delayed-to-2030s-due-to-grid-issues ; TechRadar, on UK Power Networks' early-2030s estimate.↩
37. 37The Telegraph, Microsoft axes AI deal in setback for UK data centre pioneer, 18 Oct 2026, https://www.telegraph.co.uk/business/2026/10/18/microsoft-axes-ai-deal-setback-uk-data-centre-pioneer/↩
38. 38Morgan Stanley Research (Nicholas Lentini, Michael Wilson et al.), Following AI's $1.6 trillion Money Trail: Mapping the Revenue Generated by the Capex Boom, US Equity Strategy, North America Insight, 15 Oct 2026, p.13 (Brian Nowak, hyperscalers).↩
39. 39DatacenterDynamics, Virginia senator proposes three-year state-wide data center moratorium, 15 Oct 2026, https://datacenterdynamics.com/en/news/virginia-senator-proposes-three-year-state-wide-data-center-moratorium ; Virginia's Loudoun County board votes to pause new data center applications, 18 Sep 2026, https://datacenterdynamics.com/en/news/virginias-loudoun-county-board-votes-to-pause-new-data-center-applications (also NBC4 Washington, https://www.nbcwashington.com/news/local/northern-virginia/loudoun-county-board-pauses-data-center-applications/4155005/) ; Cronkite News, 15 Oct 2026.↩
40. 40Bloomberg, New Data Centers Worth $68 Billion Disrupted in US, Data Show, 21 Sep 2026 (Newley Purnell, reporting Data Center Watch), as carried by Insurance Journal, https://www.insurancejournal.com/news/national/2026/09/21/886238.htm ; also on Yahoo Finance, https://finance.yahoo.com/technology/ai/articles/data-centers-worth-68-billion-053233556.html↩
41. 41Heatmap News, Heatmap Pro data on cancelled data-centre investment, 6 Oct 2026, https://heatmap.news/daily/260-billion-data-center-opposition↩
42. 42American Buildout tracker, Stargate Abilene, 14 Oct 2026, https://americanbuildout.com/stargate-abilene/ ; Distilled (Michael Thomas), OpenAI's Stargate Data Centers Are Taking Longer and Costing More Than Its Competitors', 17 Jun 2026, https://www.distilled.earth/p/openais-stargate-data-centers-are↩
43. 43Bloomberg, Nscale's $35 Billion IPO Plans Muddied by AI Data Center Delays, 15 Oct 2026, https://www.bloomberg.com/news/features/2026-10-15/nscale-s-35-billion-ipo-plans-muddied-by-ai-data-center-delays↩
44. 44Applied Digital, Applied Digital Brings an Additional 75 MW of AI Infrastructure Online at Polaris Forge 1, 2 Oct 2026, https://ir.applieddigital.com/news-events/press-releases/detail/161/applied-digital-brings-an-additional-75-mw-of-ai ; TradingKey, Applied Digital Stock Forecast: 250 MW Goes Live as $36 Billion AI Backlog Faces Its Earnings Test, 7 Oct 2026, https://www.tradingkey.com/analysis/stocks/us-stocks/262201410-applied-digital-apld-stock-forecast-october-6-2026↩
45. 45Galaxy Digital, Galaxy Completes Phase I of Its Helios Data Center Campus, 6 Jul 2026, https://www.galaxy.com/newsroom/galaxy-completes-phase-i-of-its-helios-data-center-campus ; investor release, https://investor.galaxy.com/news-releases/news-release-details/galaxy-completes-phase-i-its-helios-data-center-campus↩

Talk Your Book

The small print. Read it anyway.

This is not investment advice. It is one person thinking out loud about markets, usually with a position on, so assume I may be talking my own book. That is not a confession, merely useful information. Views can change quickly, occasionally before the ink is dry, and I am not going to defend a trade simply because I once wrote about it with great conviction. Nothing here is an offer, solicitation or recommendation to buy or sell anything.

I write HindeSight primarily for me. The process forces me to test the thesis, join a few dots, throw some away and, with luck, understand the problem rather better than when I started. If it sparks a thought in someone else, even better. There is no publication schedule. I write when I have something to say or when the urge becomes sufficiently annoying that writing it down is easier. I can be tactically bearish and medium-term bullish, long one week and short the next. Markets rarely reward consistency for consistency’s sake. My objective is simpler: find the best risk/reward and try to extract as much upside as possible while giving as little back as possible.

Hinde Capital no longer manages external capital, so there is nothing to sell you. Past performance is no guide to future returns and, frankly, neither is this letter. Do your own work, size risk so you can still sleep, and seek regulated advice where appropriate. No liability is accepted for losses arising from reliance on this material, including those incurred after forwarding it to your sister- or brother-in-law. © Hinde Capital 2026.

Hinde Capital · hindecapital.com · Not investment advice. Utterly talking my own book.