A leaked number landed in my feed this week. Microsoft, the story goes, plans to scale its data center capacity from 12 gigawatts to 38 gigawatts by 2032. That's a 3.2x expansion. For context, 38GW exceeds the entire peak electricity demand of New York State.
I didn't flinch at the number. I flinched at the sourcing. "People familiar with the matter." No official confirmation. No capex figure. No return timeline. Just a gigawatt count and a dramatic comparison, dressed up as news.
That's not a data point. That's a narrative. And in my line of work, narratives are the product, not the truth. I've traded through enough of these to know that when a number is this clean and this shocking, the fine print is where the money hides. So let me pull the thread.
The Arms Race Nobody Is Pricing Correctly
Here's the context you need, stripped of the press release gloss.
Microsoft is one of exactly three entities on earth with genuine hyperscale self-build capability โ the other two are AWS and Google. That triad, plus Meta, Oracle, and xAI, is currently locked in a capital expenditure war that makes the 2017 ICO boom look like a bake sale. The premise is simple: whoever builds the most AI compute capacity first locks in the infrastructure rents of the next decade.
This is not a crypto story on its face. But it is a crypto story in every way that matters to my book. Because the scarce input to this entire race is not GPUs. It's electricity. And the only industry that has spent the last six years building out gigawatt-scale power contracts, substation interconnects, and liquid-cooling retrofits is Bitcoin mining.
That's the correlation the mainstream missed. When Microsoft talks about 38GW, it is bidding against the same grid capacity that Marathon, Riot, and Core Scientific spent years locking down. Every megawatt Microsoft signs is a megawatt that doesn't go to a miner โ unless the miner sells it back. Which, quietly, several of them are now doing.
The blockchain doesn't care about your narrative. It cares about joules.
The Gigawatt Is a Lie Until You Define It
Now the core of this. The part that determines whether 38GW is a fortress or a footnote.
"Gigawatt" is not a measurement. It's a family of measurements, and the gap between them is where fortunes get made and lost.
There are at least three things 38GW could mean:
Interconnection queue capacity. This is the loosest definition. It's the amount of power a developer has applied to draw from the grid โ not what's been approved, not what's been built, not what's energised. In the US, interconnection queues are notoriously backlogged, with multi-year wait times and high withdrawal rates. A queue position is an option, not a load. If Microsoft's 38GW is a queue figure, a meaningful chunk of it is vapor.
Facility load. This is total site power draw โ everything from servers to chillers to lighting. More honest, but still not the compute number.
IT load. This is the strictest definition: power delivered to actual computing hardware. IT load typically runs 60 to 70 percent of facility load, because cooling, power distribution losses, and overhead eat the rest. In a liquid-cooled AI hall pushing 100kW+ per rack, that overhead is real.
The spread between these three definitions can exceed 40 percent. So 38GW of interconnection queue could collapse to roughly 22GW of IT load once you strip the marketing. That's still enormous. But it is not the same number, and the article that dropped this leak never defined its terms.
Based on my audit experience, that omission is not an accident. When a figure is released without a denominator, it's because the denominator weakens the headline.
I didn't need a Bloomberg terminal to see this. I needed a calculator and a willingness to ask the boring question.
The Cost Nobody Put in the Deck
The leak gives us capacity. It gives us a timeline. It gives us a scary comparison to New York State.
It gives us zero capital expenditure numbers. Zero return-on-invested-capital figures. Zero demand assumptions.
That is a massive tell.
Building a data center is a three-to-five-year exercise. A hyperscale AI campus is closer to five to seven years once you factor in power procurement, permitting, and grid interconnection. Microsoft's own framing reportedly acknowledges that new technology could "overturn current assumptions" within that window โ a polite way of saying the hardware you pour concrete for today may be obsolete before the ribbon cutting.
Here's the operational risk nobody in the bull camp wants to discuss: AI accelerators are iterating on a roughly annual cadence. Power density per rack is climbing fast. Cooling architectures are shifting from air to direct-to-chip to immersion. If you design a 7-year facility around 2025 assumptions, you are building a museum, not a moat.
And the money. Hyperscaler capex is already running in the tens of billions per quarter across the sector. Microsoft alone is in the vicinity of $80 billion a year by most estimates. That capital is committed before a single token of inference revenue is billed. Cloud revenue is post hoc. You build first, you sell second. If demand shows up, you look like a genius. If it doesn't, idle capacity becomes depreciation, and depreciation becomes an earnings cliff.
This is the same structural error I watched play out in crypto lending in 2021. Everyone assumed the demand curve was permanent. Nobody stress-tested the downside until the downside arrived.
The CoreWeave Detail Is the Whole Story
The most important line in the entire leak isn't the 38GW. It's the exclusion.
The reporting distinguishes between Microsoft's owned and leased capacity and compute rented from neoclouds like CoreWeave. That distinction is not a footnote. It's the strategy.
Microsoft is running a light-and-heavy asset hedge. Heavy assets โ self-built and long-leased campuses โ lock in long-term cost advantages on power and silicon. Light assets โ rented neocloud capacity โ preserve flexibility if demand undershoots.
If the 38GW is mostly heavy, the capital commitment is staggering and the downside is brutal. If a large slice is light-asset leasing dressed up as owned capacity, then the number is inflated by exactly the mechanism I described above.
Either way, the headline number tells you almost nothing about actual risk exposure. You have to read the asset composition to know what you're looking at.
This is where I get cynical. The entire leak reads like a document built to impress, not to inform. Scale worship. A shock comparison. No capex. No ROIC. No demand evidence. It's a press release wearing a news article's clothes, and the source is "people familiar," which in practice means someone with an interest in the AI infrastructure narrative staying hot.
Airdrops aren't free money, and leaked capacity numbers aren't free information. Both have a sponsor.
Contrarian: Retail Buys the Boom, Smart Money Watches the Grid
Here's where I part ways with the crowd.
The retail read is simple: AI is booming, Microsoft is doubling down, therefore everything AI-adjacent rips. Buy the GPU names. Buy the neoclouds. Buy the miners pivoting to AI hosting. Buy the token with "decentralised compute" in its whitepaper. Pure hopium, dressed as conviction.
Smart money is watching a different variable entirely. Not the GPU. The grid.
The binding constraint on this entire build-out is not capital, not chips, not engineering talent. It's electrons. Interconnection queues are years deep. Local communities are pushing back on power prices and water usage. Microsoft's own 2030 carbon-negative pledge sits in direct tension with a 38GW power appetite that rivals a mid-sized country. Nuclear power purchase agreements โ including restarts of plants like Three Mile Island โ are the tell that the industry knows the grid is the bottleneck.
So the real trade is not "who builds the most AI." It's "who controls the cheapest, fastest, most reliable power."
And that reframes everything for crypto. The Bitcoin miners who survived the 2022 hash-price collapse did it by owning power contracts and land banks, not by owning ASICs. Those same assets are now the most valuable thing in the market. A miner sitting on a 500MW interconnect with a substation already built is worth more to a hyperscaler than a miner with the newest fleet of rigs.
Front-running isn't just an MEV concept. It's an infrastructure concept. The players who positioned early on power are front-running the entire AI compute curve.
Takeaway: What I'm Actually Tracking
I don't trade headlines. I trade the confirmable signals underneath them. So here's what I'll be watching over the next two to four quarters, and what I'd suggest you watch too.
Capex versus PP&E. Watch Microsoft's quarterly filings for the gap between announced capital expenditure and actual property, plant, and equipment growth. If capex keeps climbing but PP&E lags, the money is going into options, not iron. Two consecutive quarters of slowing capex growth would be the first crack in the narrative.
Grid interconnect versus claimed capacity. The single most important disclosure would be actual energised megawatts versus queue positions. If an official number lands materially below 38GW, the leak was inflated. That's your signal.
Neocloud rental volumes. If Microsoft increases its rented capacity from CoreWeave-style providers, it's quietly pivoting to light assets. That would mean the capex commitment behind the 38GW headline is far smaller than advertised.
Miner power deals. Every announced hosting contract between a hyperscaler and a Bitcoin miner is a data point on where the real bottleneck sits. Track the megawatt prices. They're the new hash price.
Azure AI revenue growth. If the cloud growth rate slips below the low twenties, the demand assumption behind the entire build-out is wobbling. Idle capacity does not announce itself. It shows up in margins, three quarters later.
The bull case for AI infrastructure is real. The physics are real. The demand is plausibly real. But the 38GW number as reported is a marketing artefact, not a financial disclosure, and the difference between those two things is where I make my living.
So the question isn't whether Microsoft can build it. It's whether the demand shows up before the depreciation does.
And nobody โ not the leakers, not the analysts, not the people writing the headlines โ can answer that yet. Which is exactly why I'm not buying the story. I'm buying the confirmation, when it comes. Not before.
The grid doesn't leak. It just delivers, or it doesn't.