Here is the contradiction most readers skate past: the same afternoon that Meta's stock rose as much as 9% on the news it would sell cloud compute, according to RedMonk, the companies it would sell that compute to and against fell hard — per RedMonk, CoreWeave down 14%, Nebius down 17%, IREN down 7%.
Both moves can't be right. Either Meta's entry creates enormous value, or it destroys a slice of an existing market. The market on July 1 tried to price both at once, and in doing so revealed what it actually believes: that a fourth hyperscaler is being born, and the neoclouds are its first casualty.
That belief is the story. It is also, I'll argue, mostly wrong about why this business exists.
Last week we tracked the physical spine of the AI buildout — $68 billion in annualized data-center construction and bankrupt Bitcoin miners reborn as power landlords. This week the thread advances on two fronts: SoftBank is now leasing federal land in Ohio to build a sprawling data center, and Ro Khanna is introducing a "Data Center Bill of Rights" that would give communities the legal right to say no. Hold those two facts. They are where this ends.
Start with what Meta says. Zuckerberg has been unusually candid, calling a cloud business "a natural extension" and telling investors that "almost every week" outside companies approach Meta asking to buy compute or stand up API access. On the July 1 reporting he put a number-shaped frame on it: "The offers that you get for using the compute are so high that it may make sense, in some cases, to rent out or consider those kind of deals instead of your own internal uses."
Read that sentence again. The pitch isn't "we built a cloud." The pitch is "people keep offering us more for our chips than we paid, so we might sell some."
That is not a strategy. That is a landlord discovering his tenants will overpay.
Now the arithmetic Meta would rather you not linger on. On the July 29 earnings call, Zuckerberg said the quiet part: "We're getting a lot of offers for compute at a significant premium over what we paid for it." Same call, Meta reported free cash flow had dwindled 90% year over year on soaring capex, issued a weak revenue forecast, and watched the stock sink more than 7% after hours.
So the "premium offers" and the collapsed cash flow are two readings of one number: Meta bought too much compute, too fast, and is now looking for someone to help pay for it.
The 'premium offers' and the collapsed cash flow are two readings of one number: Meta bought too much compute, too fast, and now needs help paying for it.
This is the tell. AWS was born the same way — leftover capacity turned into a product — and Meta's boosters lean on that history hard. But AWS emerged from a retailer's spare servers. Meta Compute emerges from a $145 billion capex year that gutted the company's cash generation. The "business" and the problem are the same object viewed from different sides.
Zuckerberg himself frames it as a portfolio decision, not a market conquest: "A common trade-off that we need to make is around how much do you monetize something today versus develop future assets." And: "It would be foolish to basically just sell all of the compute and take a short-term profit." That is the language of a man managing an overhang, not launching a division.
The market's fourth-hyperscaler thesis also assumes Meta can do the hard part. It can't easily. RedMonk lays out the three things Meta needs — invest at scale, operate at a loss, and build a business orthogonal to its core. The first two are trivial for a company that, per RedMonk, lost roughly $84 billion on the metaverse over five years and kept going. The third is where hyperscalers are actually made: enterprise field sales, commercial-grade support, compliance, developer relations, B2B legal. Meta sells ads to CMOs. Cloud sells infrastructure to CIOs. Those are different muscles, and Zuckerberg admitted as much: "That's going to be somewhat a new muscle that we build as a company."
The one analyst quote that survives scrutiny is Tekonyx's Sid Nag: the market can support "a fourth hyperscaler, but not a fourth full stack cloud." Translation — Meta can rent bare metal to sophisticated buyers like Anthropic and Google. It probably can't become AWS. Which means the neocloud selloff may be the more accurate signal than the Meta pop: the thing Meta most plausibly builds is precisely the thing CoreWeave and Nebius already do.
But here's the pivot the whole debate misses. Whether Meta wins the software layer is almost beside the point, because the binding constraint on all of this isn't strategy. It's megawatts.
Follow the physical economy. Data-center construction is running at a $68 billion annualized pace, up 46% year over year, sustained since early 2024 — not a spike, a multi-year cycle. And the most valuable asset in that cycle isn't the GPU. It's the power infrastructure to run it.
That's why the Bitcoin miners are the tell within the tell. IREN, Cipher, CleanSpark, Hut 8, Core Scientific — companies built to run ASICs on cheap electricity — have realized they're sitting on the scarce thing. Core Scientific went bankrupt and re-emerged as an AI infrastructure landlord. The miners aren't pivoting to AI. They're pivoting to being the utility AI has to rent from.
The reporting is blunt about what that means for hyperscalers: securing large-scale power agreements in favorable jurisdictions is "the exact bottleneck" Microsoft and Meta face when trying to deploy fast. Meta's own capacity stockpile — the thing it now wants to resell at a premium — exists because it won that power race early. The premium offers are just the price of scarcity, and scarcity is the whole game.
Which is why SoftBank leasing federal land in Ohio matters more than any earnings-call quote. When you can't buy enough private power fast enough, you go to the one landlord with unlimited inventory: the federal government. (The Verge notes SoftBank donated $50 million to Trump's library months before the deal — draw your own line.)
And that is exactly where the wall goes up.
Ro Khanna — who represents Silicon Valley — is introducing a Data Center Bill of Rights that would let communities bar data centers within 2,500 feet of homes and schools, reject projects through a transparent process, and block utility-bill increases and water drainage. He traveled to Pennsylvania and reported the "biggest feedback" was opposition.
His words are the ones to keep: communities need the ability to say no so these facilities aren't "shoved down their throat," and the hyperscalers building them are "losing the country."
Several states, including New York, already have moratoriums. Data-center opponents are winning primaries. This is heading into the midterms as a live political liability.
So stack the whole thing. Meta overbuilt, cash flow fell 90%, and it's reselling the overhang as a fourth hyperscaler the market cheerfully priced in. The buildout it depends on is bottlenecked not on money — Meta has infinite money — but on megawatts and local consent.
You can outspend your rivals. You cannot outspend a community's right to say no. The capex budget is unlimited. The number of towns that want a substation next to the elementary school is not. That's the constraint no earnings call can override — and it's the one the 9% pop didn't price at all.