AINEWS
AI compute is moving from self build to an institutional asset class. BlackRock funds 80% of a 1 GW Texas campus, Meta leases the whole thing, and two same day deals point the same way.
Hyperscaler-scale AI data centers used to be funded on the cloud company's own balance sheet. On July 28, 2026, Meta, BlackRock, and two other counterparties drew a different line.
Meta and BlackRock announced a joint venture to develop a roughly 1 gigawatt data center campus in El Paso, Texas, with a total capital commitment of about $14 billion. BlackRock-managed funds hold 80 percent of the vehicle; Meta holds the remaining 20 percent. The campus is being built for Meta. Meta will not own it. The company will lease the entire facility from the joint venture under an initial four-year term, renewable up to four times, in a structure that could stretch the relationship to roughly two decades.
The size is calibrated to a new unit of demand. A 1 GW campus is not a marginal expansion. It is roughly the load of a mid-sized American city, and Telecompaper's follow-up reporting on the same announcement puts the site at the upper end of what a single AI cloud tenant can productively absorb. The 20-year optionality in the lease matches the depreciation curve of the underlying compute generations and lets Meta offload site, power, and water risk to a counterparty whose business is owning that risk.
The transaction unbundles two things that have lived on the same capex line. BlackRock's funds put up most of the equity for the building. The GPUs inside it stay on Meta's books. The result is a new asset shape: an institutional landlord holds the real estate, a hyperscaler rents it, and the chip vendor sells into a customer that has already locked in the rack space.
Two unrelated deals the same day point the same way. Nvidia and Hut 8, a North American data center operator that pivoted from crypto mining, signed a 15-year lease for a 1 GW Texas campus worth about $19.6 billion, with renewal options that can extend the contract to roughly 30 years and roughly $50 billion in total value. Tokyo-listed Sakura Internet ordered about ¥26 billion, or roughly $170 million at recent exchange rates, of Nvidia's next-generation Rubin HGX NVL8 GPU servers for delivery in January 2027. Three companies, three different roles, and the same direction of travel: patient capital pays for the building, the hyperscaler pays to use it, and the chip vendor sells into capacity that is already leased.
The structure matters because most public "AI bubble" arguments still assume the spending sits on Meta's, Microsoft's, or Alphabet's income statement. Increasingly it does not. When BlackRock's funds write the check for a 1 GW building, the capital cost is borne by the same investors who own the S&P 500 and the same pension funds that anchor retirement portfolios. The bubble question then becomes whether those institutions are being paid for the risk, not whether a single cloud company can outrun its own depreciation.
Four risks are worth tracking.
Single-tenant concentration. Both the Meta and the Nvidia-Hut 8 deals depend on one anchor tenant to make the underwriting work. If that tenant's model deployment schedule slips, the rent does.
Power and permitting. El Paso and the Texas Panhandle are unusually well-served by transmission, but 1 GW is a five-year construction program, and grid interconnection queues in West Texas are not empty. Delays show up as lease start-date slippage, not headlines.
Geographic clustering. The same day's two largest AI capex moves are both in Texas. Concentrating AI training capacity in one climate and one grid invites the correlated outage risk the industry spent the last decade trying to diversify away from.
"Paper deal" risk. The 20-year optionality in the Meta lease and the 30-year optionality in Hut 8 are not the same as 20- or 30-year revenue. Both depend on chip generations that have not yet shipped at scale, on training recipes still being written, and on an AI demand curve that is steeper than any prior infrastructure cycle.
The watch item is the next 12 to 24 months. The Meta-BlackRock vehicle is a template, not a one-off. The next test is whether a sovereign wealth fund, a pension plan, or a second private-markets giant closes a comparable structure with Microsoft, Amazon, or a regional cloud. If they do, the category stops being a transaction and becomes an asset class.