Meta Just Sold 80 Percent of Its $14 Billion AI Data Center. Don't Call It a Sale.

Meta just sold 80 percent of its $14 billion El Paso AI data center to BlackRock. Let that land for a second. The same Meta that refuses to call itself a media company. The same Meta that spent half a decade trying to become the metaverse. The same Meta whose founder keeps saying "superintelligence" like it's a launch date. Sold. Four-fifths. Of its most important physical asset. Before a single rack is powered. The deal is structured as a joint venture. But don't call it that. This is a sale-leaseback in a trench coat, and it's the clearest signal yet that AI capex has outgrown the corporate balance sheet. Speed isn't the pulse of the market. Financial engineering is. Here's why now. Meta's Q2 2026 investor relations deck, the one that hit my screen at 4:42 AM, doesn't bury the lead. The company is guiding for another year of crippling AI infrastructure spend. The market has already punished hyperscalers for capex without corresponding revenue. Microsoft's Q4 FY26 numbers are the cautionary tale. Azure growth is fine, but the depreciation line is eating everything. So Meta needed an off-ramp that still lets it control capacity without owning the entire depreciation schedule. BlackRock, meanwhile, is sitting on one of the largest infrastructure funds ever assembled. It wants long-duration, inflation-linked cash flows. AI data centers with 20-year power contracts are exactly that. Put the two together and you get the El Paso structure. The headline facts are straightforward. The project is a 1GW AI data center in El Paso, Texas. Total development cost is around $14 billion. Meta contributes roughly $2.3 billion in initial equity. BlackRock drops in around $4.9 billion. Meta takes back a $1 billion distribution, leaving about $6.2 billion of initial equity in the vehicle. Debt covers the rest. That is, if you believe the numbers I was handed. There's a problem: another version of the same document says the total development cost is $1.4 billion and the remaining $12.5 billion is debt. You don't need an accounting degree to see the mismatch. I'll flag it and move on, because the structure matters more than the typo. From chaos to clarity: tracking the summer of AI capital might be the most important financial beat of the decade. This deal is the first real tell. Let's get into the mechanics. At its simplest, the El Paso deal is a joint venture between Meta and BlackRock. Meta holds a minority stake, somewhere around 20 percent post-closing. BlackRock holds the majority. Debt sits on top. Meta will be the anchor tenant, operational tenant, and probably the sole user. BlackRock will collect a preferred return. The land, the shell, the power interconnection, the chillers, the backup turbines, the switches—all of it becomes a yield-bearing asset for one of the largest money managers on Earth. This is the sale-leaseback variant. A straight sale-leaseback is straightforward: a company sells an asset it owns to a buyer, then leases it back for a term. The seller gets immediate capital. The buyer gets a steady rental stream. Meta is effectively doing that with an asset that doesn't exist yet. It's a build-to-suit, then a sale. The financial alchemy is the same: convert a massive upfront capex line item into a recurring operating expense. But there's a wrinkle. In a normal sale-leaseback, the asset is completed and operating. Here, the asset is a construction project. So the sequence matters. First, Meta and BlackRock create a project company. Second, that company signs construction contracts and buys the land. Third, Meta signs a long-term lease or capacity rights agreement. Fourth, BlackRock's equity and the debt facility fund the construction. Meta's initial equity contribution is probably a development-stage loan or a preferred interest that gets repaid out of the BlackRock contribution. Hence the weird $1 billion distribution back to Meta. Now, who controls what? Here's where I have to put on my old exchange hat. In crypto, we talk about custody. In real estate, it's control. The critical question for accountants is whether Meta has "control" over the special purpose vehicle that owns the data center. If Meta controls the entity, the whole thing gets consolidated on Meta's balance sheet and the sale is a financing, not a true sale. If BlackRock controls the entity and Meta doesn't have a variable interest that absorbs the majority of losses, then Meta can push the asset off its balance sheet. The whole game is in that distinction. What do the documents tell us? The leaked IR material says BlackRock owns 80 percent. If that's a true, substantive transfer of control, then Meta has a clean path to off-balance-sheet treatment. But the lease terms are everything. If Meta guarantees a minimum rent that effectively covers debt service and BlackRock's return, then the "sale" starts to look like a collateralized loan. If Meta also has a purchase option at a fixed price at the end of the term, it's even closer to an operating lease with a bargain purchase option. That's finance lease territory, which means the asset and liability go straight back onto Meta's books. The entire transaction could be structured to achieve a specific accounting outcome. I've seen this in DeFi, too—protocols structure incentive pools to call them "marketing expenses" instead of "token emissions." Same energy. Let's talk about the real asset. 1GW is not a typo. The largest data centers in the world a few years ago were a few hundred megawatts. 1GW is roughly the output of a small nuclear reactor or a large natural gas plant. Putting that much compute in one place creates a unique set of problems: power procurement, transmission interconnection, heat rejection, and the kind of electrical substation that cities fight over. El Paso is a smart choice because of Texas's deregulated power market and the potential for power purchase agreements with wind and solar. But a 1GW facility also means the grid has to guarantee reliability at extreme peak. If you're building AI infrastructure at that scale, you're not just building a building—you're building a small city. And you're putting a power utility inside it. This is why the financial structure matters more than the silicon. The marginal cost of an additional GPU is easy to calculate. The marginal cost of grid interconnection, water rights, and construction delays is not. Meta's core competency is software and social graphs, not civil engineering. By bringing in BlackRock, Meta is outsourcing the capital risk to someone whose entire business model is assuming long-term infrastructure risks. But it's not free risk transfer. The investor gets a yield. That yield becomes a fixed cost for Meta, whether or not the AI models generate revenue. Now let's get into the balance sheet math. Reconstruct the deal from the fragments. You have a $14 billion project. Meta puts in $2.3 billion. BlackRock puts in $4.9 billion. Meta receives $1 billion back—maybe for something like a development fee or tax equity allocation. Net equity in the project is $6.2 billion. That's about 44 percent of $14 billion. The rest has to be debt. Simple math says $7.8 billion. But another line in the same IR summary says the remaining $12.5 billion after initial equity is covered by debt. That puts total project cost at $18.7 billion. The discrepancy is massive. I can only guess: the $14 billion might be the asset basis, while the $18.7 billion includes interest during construction, financing fees, and capitalized costs. Or one of the figures is just wrong. Either way, the leverage is substantial. If the debt figure is $12.5 billion on a project that costs $18.7 billion, that's about 67 percent loan-to-cost, not 80 percent. If the project cost is $14 billion and debt is $12.5 billion, then the equity is only $1.5 billion, which is nowhere near the initial equity numbers. Something doesn't line up. I'm not going to pretend I have the full waterfall. What I can tell you is that the deal uses project-level non-recourse debt, which means lenders can't chase Meta's balance sheet if construction goes sideways. That's attractive for Meta. It's also why BlackRock has to bring real equity discipline. The lenders will have a construction consultant on site, a permitting checklist, and a power-curtailment review. That's the part people underestimate. We didn't believe the hype when people said AI would need its own capital markets. This is that happening in real time. The El Paso structure is not an isolated transaction. It's a template. Think about the "asset recycling" playbook that global infrastructure funds use for toll roads and airports. A pension fund buys a mature asset, funds the expansion, takes predictable cash flows. Meta is doing the same with AI data centers. The data center is the new toll road. Meta collects usage, BlackRock collects the toll. The only difference is the toll is paid in model inference and advertising revenue rather than coins. There's also a strategic component. By selling 80 percent to BlackRock, Meta is offloading not just capital but political risk. Local communities in Texas are suspicious of AI data centers because of water and power use. A BlackRock stake adds a layer of financial credibility, but it also shifts the conversation. The project becomes a "BlackRock-Meta JV" rather than a "Meta data center." That's a deliberate public affairs move. Run the same play across the country and Meta becomes a tenant in its own empire. Let's talk about the market signal. When two giants structure a deal this complex, they're trying to solve four problems. One: Meta wants to avoid diluting its own shareholders with more equity issuance. Two: Meta wants to keep its free cash flow for buybacks and product investments. Three: BlackRock wants a hard-asset return that isn't correlated with public equities. Four: both want to avoid the appearance of an arms race that burns their respective constituencies. This structure answers all four, but at the cost of complexity. The complexity risk is real. I've audited enough projects to know that every SPV layer, every intercreditor agreement, every construction milestone is a place where bad news hides. The El Paso project has at least three layers: operating company, property company, and financing vehicle. Each layer has its own tax treatment, its own insurance, its own bankruptcy remoteness. If Meta ever finds itself in a liquidity crisis, it can't simply tap this "sale" asset. Its ability to access the data center depends on the lease, and the lease depends on BlackRock's patience. That is a big philosophical shift for a company that built its own server racks in the 2010s. Regulation doesn't stop deals; it just adds lawyers. I expect the SEC to look at this disclosure. The question will be whether Meta has told investors enough about its guarantees, its optionality, and its residual exposure. The IR deck may call it a joint venture, but if there's a guaranteed return for BlackRock and a put right in favor of BlackRock, then Meta is effectively the borrower of last resort. Shareholders need to know if this is an economic sale or a collateralized loan. The accounting distinction is not just pedantry. It changes every ratio in the credit file. Let's move to the broader context. Microsoft's Q4 FY26 investor relations deck is the shadow hanging over this structure. Microsoft has spent tens of billions on AI data centers only to see its depreciation and interest expense climb. The market reaction was brutal enough that every CFO in tech is searching for an alternative. A JV with an infrastructure fund lets a hyperscaler keep its revenue growth narrative while moving the depreciation off the income statement. If the El Paso structure works—if it's accepted by auditors and rating agencies—you will see Google, Amazon, and Oracle copy it within twelve months. Actually, they're already talking to the same funds. Exchange leads see the wave before it breaks. The wave here is asset recycling. We spent years watching retail money pour into GPU cloud starts. Now the institutional money is arriving with professional project finance. That's going to change who owns the compute. The high-water mark of the vertical integration era was when hyperscalers designed their own chips. Now they're going to rent the buildings, rent the chips, and rent the power. The only thing they'll own is the model and the user relationship. That's not a bad trade, but it inverts the old narrative. Let's build on the tax layer. This is where the deal gets even more interesting. The Inflation Reduction Act still provides valuable credits for solar, wind, and battery storage co-located with large loads. A 1GW AI data center in West Texas will almost certainly have a massive on-site solar farm and battery system to support its power resiliency. Those renewable energy credits are a real stream of income. BlackRock, as a taxable infrastructure investor, can monetize those credits efficiently. Meta, despite its enormous tax bill, might not be the optimal holder of those credits in a project this size. By giving BlackRock a majority stake, Meta is in effect selling both the asset and the tax appetite. The result is a lower overall cost of capital, but it also means the project's economics depend on the tax code. If Congress changes the credits mid-construction, the return stack shifts and someone has to eat the loss. There is also a water story. AI data centers are water hogs. Cooling a 1GW facility in El Paso, a desert city with strained water resources, is a political and environmental problem. The project may have to use air-cooled chillers or waste-heat recovery, both of which add cost. A BlackRock infrastructure fund doesn't care about the optics; it cares about the long-term permit risk. If the city water utility denies an expanded water permit, the project's construction timeline slips. That's a lender event of default in many project finance documents. I've seen projects die from worse. The flip side is that large infrastructure players are actually good at navigating this. BlackRock's global infrastructure group has spent years handling utilities, pipelines, and renewable assets. They know how to hire lobbyists, negotiate local abatements, and structure PPAs. Meta brings the AI demand; BlackRock brings the regulatory machinery. From a pure execution standpoint, the combination is smarter than Meta trying to go it alone. Let me give you my honest audit view. Based on my audit experience, the first thing I look for in any sale-leaseback is the residual value guarantee. If Meta guarantees the data center's residual value at the end of the term, it's not a sale. It's a loan with a put option. I haven't seen the full lease, but this deal smells like it has one. Meta will likely retain a right of first offer or a purchase option at fair market value. That gives Meta operational certainty, but it also means BlackRock has very little downside. BlackRock gets a contracted yield, a tax credit layer, and an institutional exit. That's not a partnership. That's a bond in disguise. I ran a rough model in my head. Say the facility costs $14 billion. If debt covers $7.8 billion at 6 percent interest, annual debt service is around $500 million. If BlackRock's equity of $4.9 billion requires a 12 percent preferred return, that's another $600 million per year. Total fixed cost before operating expenses is $1.1 billion. Meta's lease payment needs to cover that, plus BlackRock's fees. That's probably $1.3 to $1.5 billion per year. Meta's AI data center operating margin will have to be extremely high to make that rent productive. If AI inference revenue disappoints, the rent remains. Meta doesn't get to fire the building. That's the real vulnerability. The AI narrative has been built around scaling capacity in the hope that revenue follows. A structure like this locks in a fixed cost before the revenue is proven. In the old days, Meta would build a data center and carry it on the balance sheet, with the flexibility to slow depreciation or write it down if the use case withered. Now, the write-down risk moves to BlackRock, which is actually a point in Meta's favor. But the rent risk stays with Meta. In a downturn, Meta will still pay BlackRock, even if it means cutting metaverse losses or slowing hiring. That's a senior claim on Meta's cash flows disguised as an operating lease. The market doesn't know how to price this yet. On the surface, selling 80 percent of a $14 billion asset for roughly $11 billion in proceeds is a huge win. But proceeds don't hit the income statement. They come in as a financing cash flow, and the lease expense hits operating cash flow. Net cash flow may not improve at all. The only real improvement is in GAAP operating income, because rent is tax-deductible and depreciation is replaced by a lease expense that could be lower or higher depending on the terms. It's a cosmetic change that flatters EPS while increasing the effective leverage of the company. Let me be clear: I'm not saying this is fraud. Every major company uses off-balance-sheet structures. This is how airlines finance planes and how cruise lines finance ships. The problem is that AI data centers are not planes and ships. They are the core production asset of a technology company. The more Meta pushes its production assets off-balance-sheet, the more it becomes a tenant in its own business. The market's perception of Meta as a high-margin software company is partly an illusion created by financial engineering. Now the contrarian angle. Most financial commentary will say Meta is being clever. I think it's something more radical. Meta is admitting that the AI buildout is too big for its balance sheet even after years of supernormal cash flows. A company that is genuinely confident in the long-term ROI of AI shouldn't want to share the upside with BlackRock. But Meta is sharing it. That means the internal return assumptions are either too low to justify owning the asset, or too uncertain for the board to stomach. Either way, the "superintelligence" mission is now financed by a fixed-income fund. There's a certain poetry in that, and a profound shift in who controls the future. The contrarian take isn't that Meta is scared. It's that Meta is turning itself into a tenant. A hyperscaler that leases its core infrastructure from a Wall Street asset manager is no longer a true owner of the AI means of production. That's a massive cultural shift. When the next bear cycle hits, Meta's margin structure will look different. BlackRock will still collect rent. Meta's shareholders will absorb the volume risk. That transfer of risk and reward deserves more scrutiny. Look at the valuation angle. Wall Street loves asset-light models. Meta's multiple would expand if it can show strong revenue growth with lower capex intensity. This deal helps. But asset-light is another way of saying "we don't own the thing that generates the product." The cloud business is ultimately a real estate and power business. If Meta doesn't own the real estate and power, it has less control over downtime, upgrades, and security. The lease renegotiation at the end of the term will be existential. BlackRock can hold the asset hostage if Meta's revenue per inference doesn't grow as fast as the rent escalation. There are smaller technical risks, too. Construction cost inflation is one. A 1GW data center in West Texas will face concrete, steel, and transformer supply shocks. The debt facility may have covenants tied to completion milestones. If the project is delayed by twelve months, the interest reserves will be exhausted and the equity partners will have to write additional checks. That's where BlackRock and Meta start fighting. In my experience, the first casualty of any cost overrun is the relationship. The second is the off-balance-sheet treatment. Environmental and power risk is another. Texas's grid, ERCOT, has a reputation. A 1GW load in a region that sometimes asks people not to charge their EVs during summer heat waves is fragile. Power purchase agreements can hedge price, but they can't hedge grid failure. If ERCOT imposes rolling blackouts, the data center's uptime guarantee goes out the window. The lenders will underwrite the power availability, but they'll do it by requiring redundant substations and a microgrid with batteries. That adds cost, and cost is what the JV structure tries to hide. I want to get more specific about the "lease" because this is where I can add real value. Let's imagine Meta signs a 20-year lease with the JV for the entire facility. Annual rent would be somewhere between 10 and 12 percent of the asset's cost, plus pass-through operating expenses. A 10 percent rent on $14 billion is $1.4 billion per year. Meta's annual GAAP capex relief is not $14 billion; it's the avoided depreciation and the cash outflow for construction, which is replaced by a $1.4 billion operating rent. The difference is a cash flow game. Shareholders will see Meta's reported capex fall, but they won't see the rent commitment unless the footnote is well-drawn. That's a red flag. The most dangerous part of this arrangement is the hidden recourse. A careful auditor will look at the SPV's ability to pay debt. Lenders will want a completion guarantee from Meta during construction. They will also want a liquidity support agreement after the lease starts. These are off-balance-sheet contingent liabilities. They may not appear as debt, but they are credit exposure. In a severe downturn, Meta could face a margin squeeze while simultaneously being called on to support a failed SPV. That is how off-balance-sheet structures create 2008-style surprises. I've lived through enough market cycles to see the pattern. In the DeFi summer of 2020, we saw yield-chasing capital flow into unaudited protocols. In the NFT boom, we saw floor prices detached from intrinsic value. In the ETF approval sprint, we saw institutions pile into a regulated product with clean settlement. Now we're seeing a new pattern: hyperscalers treating AI data centers as exotic collateral for infrastructure funds. The question is not whether the demand is real. It is. The question is whether the financial structure hides risk in a way that will come back to bite equity holders. Let's return to the missing numbers. The difference between $14 billion and $18.7 billion is not small. It could be the difference between a straightforward project and a leveraged construction budget that includes several years of interest reserves. If I were a credit analyst, I would demand a full sources-and-uses table. I would want to know the equity commitment from each partner, the interest rate locked on the debt, and the lease escalation schedule. Without that, no one can say whether Meta's effective cost of capital is 8 percent or 14 percent. The market is currently treating this as a neutral event. I think it's a risk reclassification event. What happens next? Watch Meta's next 10-Q. Look for the footnote that describes the El Paso variable interest entity. If the VIE is consolidated on Meta's balance sheet, this whole sale is a wash and the press release was marketing. If the VIE is unconsolidated, look for the guarantee language. If Meta guarantees the project company's debt, then the entire project should be treated as liability on a risk-adjusted basis. I suspect the guarantee is there, because no lender will lend $7.8 to $12.5 billion to an 80 percent BlackRock-owned project without a user front-lettter from Meta. That guarantee is the Achilles heel. I also expect a copycat wave. In the next twelve months, every hyperscaler with a $50 billion capex plan will invite BlackRock, KKR, Apollo, or Brookfield into a data center JV. That will compress yields and increase competition for infrastructure capital. The early movers will get better terms. Meta got in before the wave peaked. That's smart. But it also means the market will eventually price AI infrastructure as lower-risk to justify the rent, and the cap rates will compress. At that point, the financial engineering advantage disappears and only operational efficiency matters. The deeper lesson is about balance sheet discipline. During the AI hype cycle, everyone wanted to own everything. Sam Altman said the AI industry needs $7 trillion. Microsoft signed huge leases. Google built global fiber. Meta kept buying GPUs. But the natural state of a mature industry is specialization. Real estate owners own real estate. Chips are owned by chip designers. Models are owned by AI companies. The ownership layers are now separating, and this BlackRock deal is the first major divorce. It will not be the last. Let's close with a question. If Meta, a company with $40 billion in annual free cash flow and a near-monopoly on social ads, can't comfortably own a $14 billion data center outright, what does that say about the economics of AI infrastructure? It says the Build-It-Yourself phase is over. The age of asset-light AI is beginning. That's not a doom sentence. It's an opportunity for everyone who can build, operate, and finance infrastructure as a service. But it's a fundamental shift in how tech earnings will look. Meta's net income will be flattered by lower capex, but its cash flow will be clamped by rent. Investors will have to think like credit investors, not just growth investors. I'll be tracking the footnotes. The first sign of trouble will be a waiver or an amendment to the project loan. The second sign will be BlackRock's fund pulling back the yield target. The third will be a Meta earnings call where the CFO says "one-time project finance charge." That's when the market will realize that the $14 billion sale was never free. The El Paso deal is a beautiful structure. It's also a warning. Speed isn't the pulse of the market. Financial engineering is. And the engineers are now in charge of the AI arms race. We should all be reading the footnotes before we tweet the headlines.