Why the Market Cannot See It

It’s all about instruments. For twenty years, technology companies and their valuation were tied to an asset-light software infrastructure: gross margins usually near 90%, marginal costs near zero.⁴⁹ The toolkit to calibrate these companies was simple. Price-to-earnings, free cash flow yield, operating margin. That toolkit was correct, and using that toolkit would reflect that capital spending was a deduction from the company’s value. Apply this toolkit to a technology company deliberating becoming a utility company, and every dial flashes red.

Just to be clear, the instruments are not broken. At the end of the day, it really is all about free cash flow. The valuation skeleton of a business really boils down to present and future free cash flow. Compare a dollar on a buyback to a dollar on substation land, and one of them would be generating electricity in 2055. A dollar on a buyback can be spent whenever; a dollar spent to secure land, substations, and power capacity. And they might be unable to buy the power they need in 2055 at anything remotely close to today’s price. Negative free cash flow during a buildout isn’t a loss; it is a rate of construction, the rate at which a company is converting capital into future productive capacity. Why are we punishing that? The market always overestimates technology in the short-term but strongly underestimates technology in the long-term.⁵⁰

As I’ve said before, the unit is the watt, and in this phase, you can count what the accounts cannot yet see. Let’s look at this from a technical standpoint. A gigawatt is a billion watts, the output of one nuclear reactor.⁵¹ When a data center company says “1 GW”, it refers to the facility power at the meter. The GW doesn’t purely reach just the chips. Some of the GW goes towards cooling, conversion, and distribution. This is what engineers express as PUE, or power usage effectiveness. PUE usually eats up around 20-30% of a GW. Engineers capture PUE primarily through a ratio: the ratio of total facility power to the power actually delivered to IT equipment. So a PUE of 1.2, for example, means that 1.0 MW is consumed by actual compute: servers and GPUs. And 0.2 MW goes towards cooling, conversion, distribution, and other infrastructure needs. Hence, a PUE of 1.3 means that 1.0 MW of compute power is delivered for every 1.3 MW drawn by the facility. So you could say that a 1 GW facility delivers roughly 800 MW of actual IT load.⁵²

So fill it out. An H100-class draws about 700 W at the board;⁵³ wrap that in a server with CPUs, memory, networking, and storage, and you’ll be looking at around 1.2 KW per GPU of IT load, assuming the facility has a PUE of 1.25. 1.2 KW x 1.25 would give you around 1.5 kW at the meter. So each H100-class GPU represents roughly 1.5 kW of total facility power demand, after PUE. Divide that number out: 1 GW of facility power hosts on the order of 650,000 to 700,000 H100-class GPUs.⁵⁴ Blackwell-generation parts draw more per unit but deliver more per watt, making the basic order of magnitude the same.⁵⁵ So when Amazon says it added 3.9 GW in 2025, it’s describing about 2.5 million top-end accelerators, activated and energized in a singular year, all by one company.⁵⁶

Run the energy dimension. 1 GW running around the clock is exactly 8.76 terawatt-hours a year.⁵⁷ Amazon’s 3.9 GW works out to 34 TWh annually. To give you an example from where I sit: Singapore, an entire developed country of about six million people, consumes about 58 TWh a year against a peak demand of 7.7 GW.⁵⁸ A single company’s capacity addition equates to 60% of my country’s annual electrical consumption, and half of its peak load.⁵⁹ A training cluster wants power at 90%+ uptime, and solar runs at 25% capacity factor.⁶⁰ So if you wanted to serve a 1.0 GW campus, you’d need 4.0 GW of panels alongside storage.⁶¹ You can do this or simply sign a nuclear PPA.

Now, looking at the financials, all-in cost for AI capacity runs about US$50 billion per GW, of which perhaps US$10-15 billion is shell, land, and power infrastructure with a lifespan of thirty years.⁶² And the remaining US$35 billion is chips depreciating over the horizon of five to six years.⁶³ At that rate, the four hyperscalers get around 14 GW of capacity with their US$700 billion, which is consistent with what is being disclosed.⁶⁴ The dollars and the watts are telling the same story through different filings.

Rent a H100 at about US$2 an hour,⁶⁵ multiply that by the minimum of what 1 GW can host, which is 650,000 GPUs, times 8,760 hours. That’s around US$11 billion a year per gigawatt at full utilization. Let’s be conservative and call it US$8 billion at realistic loads.⁶⁶ Run that against US$50 billion of capex; that is a 6-year payback before operating costs, on chips with an accounting lifespan of 5-6 years.⁶⁷ Do the maths, and you’ll realize the spot prices are approximately breakeven with the depreciation schedule of the chips. But nothing in this industry is sold at spot. The take-or-pay contract is the primary business model; the chip is the melting ice cube, and the contract is the real asset that secures revenue. If I cross-check that number with a neocloud publishing both numbers: CoreWeave’s US$99.4 billion backlog coupled with 3.5 contracted GW is around US$28 billion per GW; divide that by the chip shelf life, and you’ll get around US$5.5-6 billion a year on a five-year term schedule.⁶⁸ Around the same vicinity as the number I calculated above. Hence, we have US$50 billion per GW in, US$25-30 billion per contracted GW back, and a chip that dies in six years.

But now, bring that forward because the finance story becomes a physical one. 14 GW of new AI load in a single year is about 123 TWh run continuously, which is more electricity than the Netherlands consumes for everything.⁶⁹ Added in twelve months, by four single companies, in a buildout that is only just beginning. American electricity demand grew at less than 0.6% a year across the first two decades of this century, and now that system is demanding 15-25% more electricity by the year 2030, with peak demand rising from about 760 GW towards 850-930 GW.⁷⁰ Furthermore, EPRI (Electric Power Research Institute) projects American data centers to consume 9-17% of national generated power by 2030, up from 4.4% in 2023, which is roughly a 60% increase from the number it published two years prior.⁷¹ It’s a mirror image of Part I, where every skeptic’s ceiling became the ground floor for the year ahead. And at the top of that range, one country’s data centers would draw more electricity than France consumes as a whole.⁷² Right now, the electricity for this does not even exist, and money cannot make it appear.

To make it clear, aggregate generation is not the problem. The United States will add a record 86 GW of utility-scale capacity in 2026, which on the surface looks like more than enough to cover the 14 GW of AI load.⁷³ But wait, solar is 51% of that number, and battery storage is 28%, whilst net new gas capacity comes under 4 GW, and new nuclear comes to zero.⁷⁴ I mentioned earlier in this article that a training cluster wants power at 90%+ uptime, and solar runs at 25% capacity factor. Hence, the relevant figure is the 4 GW of newly dispatchable capacity, against an AI buildout wanting about three times more than that. Essentially, the argument is that the added 86 GW of capacity additions cannot and shouldn’t be read as 86 GW of firm, 24/7 supply.

Which is why turbine order books are important to watch. GE Vernova, Siemens Energy, and Mitsubishi supply more than 70% of global production, and their order books now hold a figure higher than 170 GW.⁷⁵ Combined cycle lead times have extended from three and a half years to five, meaning power infrastructure needs to be planned further in advance, and a data center developer can’t simply say I need 1 GW of power next year.⁷⁶ Northern Virginia’s interconnection queue now runs to seven years.⁷⁷ Roughly one third of data centers scheduled to open this year may be delayed or never built.⁷⁸ And the binding constraint isn’t capital or demand; it’s firm power at a specific node, on a schedule that is measured in years.

I hope this illustrates what is genuinely scarce here, and it isn’t the GPU. Nvidia will sell that to anyone with the money. The scarce asset here is energized interconnection: a megawatt contracted, built, connected, flowing to the data center, and operationally running. When supply is constrained by the speed of turbine manufacturing and new capacity takes years to arrive, the value of existing energized interconnection and active capacity is determined by the replacement cost of the marginal new unit, not what it’s worth historically. That’s scarcity rent.

This brings me to a point that I’ve been meaning to express for a while. Notice which direction the risk runs. The worst-case scenario treats expensive electricity as a cost problem for hyperscalers. But Microsoft holds Three Mile Island on a twenty-year PPA. Amazon has Comanche Peak, and Google bought Intersect Power outright. When power does get dearer, which it will, the incumbent’s input cost barely moves while the marginal replacement cost of its asset base climbs.⁷⁹ So the point that I’m trying to make is that these companies, who are already energizing themselves, are creating a moat for themselves. And somehow no one can see it.

And then the buildouts end, which is the part that almost nobody is modelling. Only the chips recur. The shell, substation, interconnect, US$10-15 billion is not bought twice. Growth will flatten, capex will fall towards the refresh cycle whilst US$5-6 billion per contracted GW will continuously arrive towards an asset base already paid for. What was once a capital sink will begin to generate substantial free cash flow, making free cash flow invert. And the instrument that signals avoid today will print something entirely different. The metric was never the problem; the phase was.

“Ignore the P/E, look at this instead” is the signature move of every mania. In 1999, investors pointed to website eyeballs, an impressive number but one with no guaranteed cash flow behind it.⁸⁰ A contracted megawatt is different. If Meta signed a contract for that power, it represented a real customer and a claim on future revenue. In this case, capacity is an indicator of cash flow, but never a substitute. Buffett has described this market as “a church with a casino attached.”⁸¹ The casino is definitely real, but so is the church. Buffett also underwrote BNSF railway on enterprise value per mile of track.⁸² Valuing an asset based on contracted capacity is not a new trick. It’s an old trick in finance.

---