Kevin O'Leary is building the world's biggest data center. He just said 60-65% of them will never get built.

On the MarketBeat YouTube channel on October 16, 2026, host Bridget Bennett aired an interview titled \"Mr. Wonderful Warns Most AI Data Centers Will Shut Down (Here's Who Wins).\" The guest was not an enterprise-AI startup founder — it was Kevin O'Leary, the Shark Tank investor who chairs O'Leary Digital and is personally trying to build what he calls \"Wonder Valley,\" a planned 40,000-acre AI campus in Box Elder County, Utah with 9 gigawatts of on-site gas generation. He has skin in the game at unprecedented scale, and he used the interview to say something that contradicts his own commercial interest: 60-65% of announced North American data center projects will never get built.

A clarification is necessary because the assignment that surfaced this story conflated two unrelated \"Wonderful\" entities. Wonderful AI — the Amsterdam-headquartered enterprise-AI-agent startup that closed a $550M Series C at a $5B valuation on September 2, 2026 with Salesforce as a new investor — has not, as far as the public record shows, issued any bear thesis on data center shutdowns. The speaker here is the unrelated Mr. Wonderful, Kevin O'Leary, and that distinction actually strengthens the call: the warning comes from a developer who is mid-bid on the biggest project in the country, not from a pundit with no exposure.

His math is hard to dismiss as outsider noise. He frames the test as three filters: do you have power, do you have a permit, do you have an anchor tenant? Most announced projects fail at least one. A single hyperscale project now runs ~$16B from planning to operations, with planning alone costing $20-30M per gigawatt, and tenants that used to ask for 250-500 MW are now asking for at least 1.4 GW — a jump that can only be served by sites that already have utility-scale power agreements.

On paper, the bull case has gotten stronger since the summer, not weaker. The four largest hyperscalers — Amazon, Microsoft, Alphabet and Meta — are guiding to roughly $720-745B of capital expenditure in 2026, up about 77% from a year earlier. Amazon raised its 2026 capex estimate to roughly $220B in its Q2 release; Microsoft spent $41B in fiscal Q4 alone (July 29, 2026) and brought 88 new data centers online across FY2026; Meta bumped its 2026 capex range to $125-145B; and Oracle, a fifth hyperscaler, told investors its FY2026 capex reached $55.7B and guided to $70B of net capex in FY2027 after signing $67B of new AI infrastructure contracts in Q4 alone. Goldman Sachs Research raised its US data center power demand forecast on October 5, 2026, now expecting +12 GW in 2026 and +17 GW in 2027.

That is a one-sided way to look at the same number. The other side is Kevin O'Leary's 60-65%, applied to the much larger universe of speculative projects that hyperscaler announcements have spawned — REITs, developers, and shell companies that announced campuses but have neither power agreements nor anchor tenants. Multiplied through the pipeline, the practical implication is that even a $740B capex year from the four biggest hyperscalers will result in less physical capacity than the headline announcements implied.

Listed data center REITs are the cleanest way to play the bear thesis because their assets are long-duration, debt-heavy, and dependent on contractually locked power — and most of their announcements over the last 18 months have been speculative. Equinix trades at 12.6x EV/sales and 27.7x EV/EBITDA on its TTM through June 2026, with $95.9B of market cap and $9.83B of revenue. The company raised its full-year 2026 revenue growth guidance to 11-12% and AFFO/share to 10-12% on July 29, 2026 — strong, but the multiple assumes the buildout completes. Digital Realty Trust, the more levered of the two at 4.8x net debt/EBITDA and 85x trailing P/E, has 3.0 GW of IT capacity in service and 6.3 GW buildable. Both stocks gave back gains on September 15, 2026 after a weaker outlook from Digital Realty spooked the sector, and the pair has underperformed since.

The contrast with the power sellers is the point. Vistra, the Texas-based independent power producer with 38,700 MW of generation capacity, sold 5-10x more reliably than the REITs throughout the October selloff — its customer is whichever data center actually energizes, not whichever one gets announced. Constellation Energy, which runs the largest US nuclear fleet at 32,400 MW, has already locked in 20-year power contracts with Microsoft to restart Three Mile Island's Unit 1 and operates behind a regulated moat that doesn't depend on any one project breaking ground. These two utilities trade at 10-15x EV/EBITDA, roughly half the multiple on the data center REITs, and that gap is the entire market mispricing the \"useful vs. provisioned\" divergence.

NVIDIA is the clearest beneficiary of the divergence and the easiest case to underwrite. Q2 FY2027 (reported August 26, 2026) delivered $96.2B of revenue, up 106% year-over-year, with data center revenue of $89.0B, up 117%. The customer base is the four hyperscalers that are spending the $740B, and NVIDIA does not care which announced campus gets canceled — its chips go into the surviving 35-40% of the pipeline, not the announced 100%. Oracle sits in a more interesting middle: it is signing $67B of new AI infrastructure contracts in a single quarter and committing $70B of FY2027 capex to back them, but its stock sold off after its June 10, 2026 release because the contracts require it to fund data centers before customers pay, putting pressure on free cash flow. Oracle is the hyperscaler most exposed to the \"useful vs. provisioned\" gap if its anchor tenants scale slower than the buildout.

Days to quarters: the test of Kevin O'Leary's call comes from the next two earnings cycles of Equinix (October 29 guidance) and Digital Realty Trust (October 29, 2026), and from Q4 hyperscaler capex updates. Any guided slowdown in new project signings, or any meaningful reduction in \"backlog conversion\" to operating capacity, will be the first confirmation. The current setup — REITs at 26-28x EV/EBITDA, utilities at 10-15x — is what the gap looks like before the market accepts the divergence.

One to three years: the survivors of the 60-65% cull will be the listed projects with permits, power, and anchor tenants today. Power infrastructure (gas turbines, nuclear restarts, grid interconnection) takes 18-36 months to deliver and is the binding constraint on which 35-40% survives. Any company whose revenue model depends on the announced 100% — speculative REITs, certain private developers, Crusoe competitors, and any merchant data center without a hyperscaler anchor — is in line for the cull. The risk to Kevin O'Leary's thesis is the opposite: if power bottlenecks ease faster than expected — through grid reform, behind-the-meter gas, or siting reform — the cancellation rate falls and the REITs catch a bid. Meta's October 21, 2026 earnings will be a leading indicator of how aggressively the hyperscalers expect to self-supply power going forward.