Meta Handed BlackRock 80 Percent of a $14 Billion El Paso Data Center. That's the Second 80/20 JV in Nine Months.

Meta and BlackRock filed a joint press release before the open on Tuesday, July 28, 2026, announcing a $14 billion, 1 gigawatt data center campus in El Paso, Texas. Site launch is scheduled for 2028. Meta will be the first and only tenant. Funds managed by BlackRock own 80 percent of the joint venture. Meta owns 20 and contributes the land and other assets at a value of roughly $2.3 billion. BlackRock puts in $4.9 billion of cash. Meta receives a $1 billion one-time payment on close. A $12.5 billion bond offering finances the rest of the capital stack.

Read the ownership split first. Eighty percent of a 1 GW AI campus is going to sit on BlackRock's books. Meta contributes the site, signs the offtake, and gets a check on the way in.

The reason this deal reads the way it does is that Meta has done it before. In October 2025, Meta and Blue Owl Capital formed an identical 80/20 joint venture for the Hyperion campus in Richland Parish, Louisiana, then valued at roughly $27 billion. Meta expanded Hyperion two weeks ago to 5 gigawatts and $50 billion of total build. Two JVs, two asset managers on the majority side, and one hyperscaler on the tenancy line. The template is now the story.

The Numbers

Sum the two campuses at their current sizing: 6 gigawatts, $64 billion of build cost, both structured so that the hyperscaler holds the smaller equity slice on the property and books the compute as an operating expense on the way out. Meta is the sole tenant on both. If you were sizing Meta's effective compute exposure by counting only the 20 percent equity share of these JVs, you would miss most of it.

What the Structure Actually Does

An 80/20 project-finance JV with an asset manager on the majority side is not a traditional data center lease and it is not a build-to-suit. It is a purpose-built special vehicle that owns the physical campus, borrows against the anchor tenant's contracted rent, and depreciates the building on the asset manager's books.

Three things happen at the same time when a hyperscaler runs this playbook. One, the gigawatt does not land on the hyperscaler's balance sheet at cost, only the 20 percent equity contribution and the lease commitments do. Two, the tenant obligation becomes an operating item that spreads over the term of the take-or-pay contract instead of a lump-sum capital line in the year the campus is built. Three, the asset manager gets a long-duration, credit-grade income stream backed by the strongest offtake in the US corporate market: a hyperscaler paying rent on compute that services its own revenue base.

The $1 billion one-time payment is the tell. Meta contributed roughly $2.3 billion of land and assets, took a $1 billion check out at close, and ended up with a 20 percent equity stake in a $14 billion project. In practical terms, Meta got compensated on the way in for the entitlements it had already spent to get the site permittable. That is how a hyperscaler turns a land bank into working capital while still guaranteeing itself the compute output.

The NAVER Rhyme

We walked through the NAVER, NVIDIA, and Brookfield deal yesterday: a $10 billion Korean sovereign AI factory with Brookfield as the exclusive project-finance partner on a non-binding term sheet, NVIDIA writing a $1 billion equity alignment check, and NAVER carrying the tenancy. The El Paso deal is the same shape with the country and the counterparties swapped. Asset manager on the majority position, hyperscaler-adjacent equity on top for alignment, contracted offtake underneath, and a gigawatt of physical asset in the middle that nobody wants sitting on the operating company's balance sheet.

The pattern is the same because the problem is the same. A gigawatt AI campus takes 24 to 36 months to build, needs $10 billion to $50 billion of capital, and has an anchor tenant that has to guarantee the offtake for the debt to price. Whether the tenant is a Korean incumbent (NAVER), a US hyperscaler (Meta), or a frontier lab (Anthropic on TPU, on Trainium, on Colossus, and now on MI450), the capital does not want to sit on the tenant's books. It sits on an asset manager's.

The financing loop we tracked through Q1 and Q2 (Google recycling $40 billion of equity into Anthropic TPU offtake, AMD writing $5 billion into Anthropic against 2 GW of MI450, NVIDIA sending $40 billion into OpenAI to anchor Vera Rubin) is the vendor version of the same trade. The El Paso deal is the landlord version. Both close the same loop. Both offload the balance sheet weight of the physical buildout to a party whose business model is holding long-duration assets.

What This Does to CapEx Analysis

Meta's 2026 CapEx guidance is in the $145 billion to $155 billion range. Google just raised full-year 2026 CapEx to $195 billion to $205 billion, per the July 22 earnings call. Microsoft, Amazon, and Meta together are running roughly $500 billion of stated 2026 CapEx across the hyperscaler complex. Those numbers get quoted every quarter in the AI bubble debate. They also get less informative every quarter, because more of the effective compute buildout is being financed off balance sheet through structures like this one.

On the disclosed numbers, Meta's 20 percent equity stake in El Paso is worth roughly $460 million against a $14 billion build. Meta will book its 20 percent share of the JV under equity method accounting, and the lease payments will show up as operating expense once the site is producing tokens. The other 80 percent (roughly $11.2 billion of underlying asset cost, plus the $12.5 billion of bonds sitting above it) is somebody else's balance sheet problem. Repeat the exercise for Hyperion at $50 billion, and Meta is running two of the largest AI campuses in the western hemisphere with a low single-digit-billion equity footprint on its own books.

Two things follow. First, the CapEx number the market reads on a hyperscaler earnings call is now a lower bound on effective compute spend, not the ceiling. Analysts have to look through the JV commitments (usually disclosed in the 10-K commitments and contingencies note, sometimes only reconcilable a quarter or two after the fact) to size the full exposure. Second, the risk sitting behind the compute buildout is migrating from hyperscaler shareholders to asset manager LPs. If the AI revenue curve keeps up, the LPs get a long-duration credit-grade yield and everyone wins. If it does not, the loss sits on BlackRock and Blue Owl clients instead of Meta.

This is not new financial engineering. Telcos did it with cell towers. Real estate did it forever. What is new is the pace and the scale: an entire compute buildout for the largest hyperscaler complex in history is running through JV structures that did not exist at this size two years ago, financed by private credit funds that raised the capital specifically for AI infrastructure.

Where the Rest of the Complex Goes

Meta is out in front on this structure but is not going to be alone for long. Microsoft, Amazon, and Google all have the same problem: build $150 billion to $200 billion of physical infrastructure per year without breaking the operating margin line that funds the equity story. The obvious next steps to watch are Google finding an equivalent to Blue Owl or BlackRock on a US TPU campus, Microsoft splitting one of the Wisconsin or Georgia builds off through a private-credit JV, and Amazon doing the same on a Trainium 3 campus. OpenAI is a special case, because it is not public and has already told the market it plans to lean hard on third-party procurement, but the same structural logic applies: keep the tokens, ship the balance sheet weight.

Two categories of counterparty are best positioned. BlackRock, Blue Owl, Brookfield, and Apollo raised dedicated AI infrastructure funds in the last 18 months. Their mandates are long-duration, credit-linked, and specifically sized for gigawatt-scale builds. Ten years ago this deal shape did not fit any existing private credit fund because the assets did not exist at this scale. Now the assets, the funds, and the hyperscaler counterparties all exist at the same time, and the term sheets are getting traced.

Our Take

The El Paso deal is not the biggest number of the week. Hyperion at 5 gigawatts and $50 billion is. But the El Paso deal is the more important one, because it is the second instance of the same template, and templates are what turn one-off dealmaking into repeatable market structure. When you see a hyperscaler and an asset manager settle on an 80/20 split with the tenant retaining land contribution and taking a modest cash payment at close, you are looking at what the next $500 billion of AI infrastructure is going to be financed like.

The uncomfortable version of the same read: hyperscalers have found a way to keep building at gigawatt scale while limiting the balance sheet exposure the market can see. That is good for hyperscaler equity, good for asset manager fee income, and neutral to positive for the actual buildout timeline. It is worse for anyone trying to look at published CapEx numbers and infer how much AI compute is actually being contracted. The bull case for AI CapEx just became structurally harder to disprove, because much of the spend is now happening one step off the reported line.

Three signposts. First, whether Meta names a third asset manager on a third US JV before the end of 2026 (the pattern locks if it does). Second, whether Google, Microsoft, or Amazon files a comparable 80/20 project-finance JV on a named US site by Q1 2027 (if they do, this is the standard hyperscaler treatment for physical infrastructure, not a Meta idiosyncrasy). Third, whether SEC disclosure rules catch up before the pattern becomes universal: the current 10-K commitments and contingencies language was written for a world where the number in that footnote was measured in hundreds of millions, not tens of billions. When BlackRock, Blue Owl, and Brookfield collectively hold title on a hundred gigawatts of AI compute, the market is going to need a new line item, and it is going to want it on the face of the balance sheet, not in the notes.

We are tracking the campus buildouts on our AI buildout explainer and the vendor stack on the Meta provider page. The next data point that changes this thesis is the Q3 hyperscaler earnings cycle, starting with Meta on the same October window it always uses. If the JV commitments note grows by another $10 billion to $20 billion and analysts do not press on it, the template is going to run to its natural end.