The code doesn't care about your press release. It cares about the term sheet. Meta just announced a $14 billion, one-gigawatt AI data center in El Paso. The headline says Meta is building it. The structure says Meta is selling 80 percent of it to BlackRock before the first GPU is racked.
This isn't a sale-leaseback. It's a sale-leaseback variant wrapped in a joint venture and layered with debt. Meta puts in $2.3 billion. BlackRock puts in $4.9 billion. Meta gets a $1 billion distribution. Then the remaining $12.5 billion of total development cost is debt-financed. Add it up. If the project is really $14 billion, the numbers don't close. Two billion plus four point nine minus one plus twelve point five is eighteen point seven, not fourteen. Something is off.
That gap is where the real story lives. You don't need to trust the narrative. You need to trust the math.
The Hook
The code doesn't care about your press release. It cares about the term sheet. Meta just announced a $14 billion, one-gigawatt AI data center in El Paso. The headline says Meta is building it. The structure says Meta is selling 80 percent of it to BlackRock before the first GPU is racked. This isn't a sale-leaseback. It's a sale-leaseback variant wrapped in a joint venture and layered with debt. Meta puts in $2.3 billion. BlackRock puts in $4.9 billion. Meta gets a $1 billion distribution. Then the remaining $12.5 billion of total development cost is debt-financed. Add it up. If the project is really $14 billion, the numbers don't close. Two billion plus four point nine minus one plus twelve point five is eighteen point seven, not fourteen. Something is off. That gap is where the real story lives. You don't need to trust the narrative. You need to trust the math.
The Context
Let me step back. Meta's AI strategy has shifted from model research to physical infrastructure. Superintelligence ambitions require compute, and compute requires gigawatts. El Paso is one of the largest AI data centers ever planned. A single gigawatt is roughly equivalent to a small nuclear reactor's output. This is not a pilot project. It's a bet on the entire AI buildout.
But Meta isn't paying for it alone. The structure works like this: Meta and BlackRock form a joint venture. Meta sells an 80 percent stake to BlackRock. The joint venture owns the asset. Meta becomes the operator and anchor tenant. BlackRock's infrastructure fund gets a long-term, dollar-denominated yield stream. Meta gets guaranteed access to 1GW of compute capacity without putting the full $14 billion on its balance sheet.
This is financial engineering, not AI engineering. The technical route is mature: power, cooling, networking, GPU clusters. Any competent team can build it. The interesting part is who absorbs the risk. And the answer is not BlackRock.
The broader context is a wall of AI capital expenditure. Microsoft's latest quarterly results show the same pattern. Every major hyperscaler is tripping over itself to announce bigger data centers. But the market is starting to ask a simple question: if AI revenue doesn't compound as fast as AI capex, whose balance sheet eats the difference? Meta's answer is, this time, not entirely mine.
That's the reveal. This deal is not a capex story. It's a risk-transfer story. And it tells you more about the AI trade than any earnings call.
The Core
Here's where I get skeptical. Based on my audit experience—I spent 2018 in Istanbul tearing through Compound and MakerDAO's early contracts, finding reentrancy bugs that would have drained the protocols—I learned that the real risk lives in the layers people skip. The same applies to a data center term sheet.
Let's walk through the layers.
First, the joint venture is an off-balance-sheet vehicle. Meta can avoid consolidating the full $14 billion on its balance sheet. It will report equity income or operating lease expense, not the debt. Wall Street sees a lower capex number, a cleaner cash flow statement, and Meta's stock doesn't tank. But the obligation is still there. Meta is the anchor tenant. If the data center doesn't fill with revenue, Meta pays the rent, not BlackRock.
Second, the debt layer. The remaining $12.5 billion is supposed to be debt-financed by third-party lenders. In normal project finance, lenders underwrite the cash flows. They look at the anchor tenant's credit rating, the power purchase agreements, the technology risk. Here, the anchor tenant is Meta. That means the debt is effectively Meta's credit risk dressed up as non-recourse project debt. If the project fails, lenders will go after the JV. The JV has one real asset: a 1GW data center that only Meta has the expertise to operate. If Meta walks away, the asset is a tombstone. So the debt is Meta's debt in everything but name.
Third, the distribution. Meta put in $2.3 billion and gets $1 billion back. That's a 43 percent return on day one, before any power is delivered. Why would BlackRock allow that? Because the asset is overfunded, or because Meta's contribution is being repaid for development services, guarantees, or the value of land and permits. But it also means Meta is getting paid to enter a deal that locks it into decades of capacity payments. That's the kind of leverage that feels great in a bull market and suffocates in a downturn.
I didn't need a Bloomberg terminal to see the inconsistency. I only needed a calculator. Eighteen point seven billion dollars of total funding against a purported fourteen billion dollar project. There are three possible explanations. One: the $14 billion number is stale, and the real project cost is closer to $19 billion. Two: the $12.5 billion debt figure includes mezzanine or seller notes that aren't counted in the headline cost. Three: there's an accounting offset, like construction-backlog revenue or contributed assets, that makes the numbers work. All three share a common thread: the headline number is not the whole number.
Meta's Q2 2026 investor relations deck dresses this up as capital-efficient infrastructure expansion. Microsoft's Q4 FY26 call uses the same language: satellite balance sheet structures. That is corporate finance translation for we don't want to consolidate the debt. I've spent fourteen years reading this language. Every time a company stops talking about ownership and starts talking about access, the balance sheet is doing something complicated.
Now add power. A 1GW facility cannot be plugged into the existing grid like a toaster. It needs a dedicated transmission corridor, substations, and a power purchase agreement that guarantees capacity for decades. Who signs that PPA? If the joint venture signs it, lenders will require Meta's credit support because the JV has no operating history. If Meta signs it, Meta holds a direct liability that nobody is calling out in the press release. Either way, the off-balance-sheet structure is not as clean as it looks.
This is exactly the kind of gap that creates alpha. Alpha isn't in the 10-K, it's in the footnotes. The footnotes say Meta is no longer a pure buyer of AI infrastructure. It's a tenant and a risk operator. The code doesn't lie, but the press release does.
The Contrarian View
The market will read this as Meta accelerating its AI spend. I read it as Meta de-risking its balance sheet. Both are true. But the second one matters more.
Retail sees a famous tech company building the biggest data center in America. Smart money sees an asset manager collecting a six to eight percent infrastructure yield, backed by a covenant-lite tenant with a trillion-dollar valuation. The trade is not about AI adoption. It's about who holds the bag if the AI buildout runs ahead of demand.
Here's the counterintuitive part: BlackRock might be the better position. BlackRock gets a stable, asset-backed yield and can sell its stake to LPs at a premium. Meta gets compute capacity and a clean balance sheet. But Meta is also signing a contract that turns its own future into a fixed cost. If AI revenue hits a speedbump, Meta cannot simply cancel the gigawatt. The lease is the leverage.
In a bull market, anyone can be a genius. Data center announcements are up, credit spreads are tight, and infrastructure funds are desperate for yield. That's exactly when financial engineering gets sloppy. The debt may look non-recourse, but the reality is recourse to the operating business. The lenders are not betting on El Paso weather. They are betting on Meta's share price, ad revenue, and patience with a capital-intensive business that hasn't yet paid for itself.
I've seen this movie before. In 2022, Terra's oracle mechanics looked like a stablecoin feature. It was a leverage bomb. The collapse wasn't random. It was the inevitable unwinding of over-leveraged structures. I shorted LUNA on the way down and made $120,000 in 72 hours, but I didn't do it because I was lucky. I did it because I could see that the ecosystem was using one layer of debt to secure another. This deal has a similar shape. Not identical, but similar: a headline asset, a complicated vehicle, and a hidden claim on the parent company's future cash flows.
There's another blind spot: demand. Everyone assumes AI compute demand is infinite. It isn't. It's a function of model adoption, token prices, enterprise budgets, and energy policy. A 1GW data center is a 20-year asset. Even if AI becomes the most valuable industry on earth, the timing gap between building 1GW and actually earning a return on 1GW is the entire ballgame. Meta is betting that the gap can be bridged with BlackRock's balance sheet. That may be true. But the financing cost is embedded in Meta's future margins. That cost is invisible in today's press release. We don't get to see the final term sheet, but we don't need to. We can see the shape of the risk.
The Takeaway
So what do you do with this? If you are long AI infrastructure, stop looking at data center announcements as unalloyed bullish signals. Start looking at the credit structures underneath them. The specific level to watch is the spread between Meta's corporate bonds and U.S. Treasuries. If that spread widens beyond 150 basis points, the market is telling you that Meta's fixed commitments are getting heavy. That is your price level. That is the trade.
If you are an operator, the lesson is simpler: when a hyperscaler offers you a lease, read the termination clauses. Meta's El Paso deal is a product, not a plan. BlackRock is selling an 80 percent stake in a promise. Meta is buying the right to rent its own future.
Trust the math, fear the hype, ignore the noise. The code doesn't care about your press release. Neither does the term sheet. And the term sheet is the only code that matters here.