AI Meets Aggressive Accounting at Meta’s Gigantic New Data Center
- Eric Anders
- Nov 25
- 12 min read
Meta’s Hyperion Data Center: Off-Balance-Sheet Debt, On-Balance-Sheet Ecocide
By Eric Anders
Meta Platforms is building what it calls Project Hyperion – a 4-million-square-foot data center in rural Louisiana, designed to deliver over 2 gigawatts of computing capacity for artificial intelligence research (datacenters.atmeta.com). To put that in perspective, if run at full tilt, this data center could draw roughly as much electricity as 1½ million U.S. homes (reuters.com). It’s a critical piece of infrastructure for Meta’s AI ambitions, but it’s also a beast of a project – financially and energetically. And in a marvel of financial engineering, Meta found a way to build this $27 billion behemoth with “other people’s money” while keeping the hefty assets and debts off its own books (longbridge.comlongbridge.com). If that sounds too good to be true, that’s because it probably is. This situation provides a textbook case of structured finance colliding with economic reality – or as Wall Street Journal columnist Jonathan Weil quipped, “Artificial intelligence, meet artificial accounting” (https://www.wsj.com/tech/meta-ai-data-center-finances-d3a6b464).
Off-Balance-Sheet Magic, AI-Style
Here’s how Meta pulled off this bit of balance-sheet sorcery. Last month, Meta quietly moved the entire Hyperion data center project into a new joint venture with investment firm Blue Owl Capital (longbridge.com). Meta retains only 20% ownership, while Blue Owl (via a holding company blandly named Beignet Investor LLC) owns 80% (longbridge.com). Blue Owl’s fund put in about $7 billion cash and even paid Meta a one-time $3 billion reimbursement for costs Meta had already sunk into the project (reuters.com). The rest of the funding – a record $27.3 billion in bonds – was raised from outside investors, primarily Pimco (longbridge.com). In essence, Blue Owl and bond buyers fronted the money to build Meta’s AI super-center. Meta’s plan is to lease the finished data center back from this joint venture for up to 20 years (initially committing to just a 4-year lease starting in 2029, with the option to renew in 4-year increments) (reuters.com).
Why go through these convolutions? The goal is simple: get the benefits of a $27 billion facility without looking like a company that spent $27 billion. By using an external joint venture, Meta can claim that the data center – and the huge debt used to finance it – aren’t Meta’s problem. The company wants to preserve its near-pristine credit rating and avoid denting its financial metrics with mountains of new debt and assets on the balance sheet. In other words, it wants to have its cake and eat it: use other people’s capital to power its AI revolution, reassure those investors that Meta ultimately has their backs, and keep its own balance sheet squeaky clean (longbridge.com). This is quintessential off-balance-sheet financing – a technique companies often use to reduce the appearance of leverage and improve reported finances (illuminem.com). The catch is that such practices can obscure the true extent of a company’s obligations, masking risks from shareholders and regulators (illuminem.com). And in Meta’s case, the mask is starting to slip.
The Convenient Assumptions Behind the Curtain
For Meta to keep Project Hyperion off its books, accounting rules impose a high bar. U.S. accounting standards say if a company is the “primary beneficiary” of a Variable Interest Entity (a VIE – basically a special off-book entity), it must consolidate that entity’s finances on its own statements. The primary beneficiary is defined by two key tests (longbridge.com):
Power: Does Meta have the power to direct the activities that most significantly impact the joint venture’s economic performance?
Exposure: Does Meta have an obligation to absorb significant losses of the venture, or a right to receive significant benefits from it?
If both answers are “yes,” Meta would be forced to treat the joint venture as if it were part of Meta – nullifying the off-balance-sheet magic trick. So Meta asserts (in its disclosures) that the answer to at least one of those questions is “no” (longbridge.com). The favorable accounting treatment hinges on a trifecta of convenient assumptions (longbridge.com) (longbridge.com):
Meta supposedly lacks control over the venture’s most important decisions. In Meta’s view, Blue Owl – a financial investor – somehow holds the real operational power, not Meta (longbridge.com). Meta claims it “does not direct” the critical activities that drive the data center’s performance. This is implausible on its face. Meta is the hyperscaler with the expertise to build and run data centers; Blue Owl is just footing the bill. Who really determines whether this AI campus succeeds? Common sense says Meta’s technology, decisions, and skills will make or break the project’s economic performance, regardless of what the legal ownership chart says. Blue Owl may control the JV’s board on paper, but when it comes to the complex task of constructing and operating a 2 GW AI computing facility, Meta is unquestionably in the driver’s seat. The accounting rules care about substance over form, so claiming Meta has no “power to direct” here strains credulity.
Meta assumes it might walk away after a 4-year lease – so future use is uncertain. The joint venture structure is set up so that Meta’s lease on the data center starts with a short initial term (just 4 years, from 2029 to 2033) with optional renewals (reuters.com). By structuring it this way, Meta can treat it as an “operating lease” rather than a purchase or financing. Under current accounting standards, even operating leases do get recorded as liabilities, but Meta’s initial lease commitment would cover only four years of rent – a far smaller liability than the entire $27 billion cost. Meta is effectively telling auditors: “Who knows if we’ll still need this data center after 2033? It’s not reasonably certain we’ll renew, so we won’t book those later obligations.” This assumption is what keeps the bulk of the asset and debt off Meta’s balance sheet. But consider how unlikely this is: Meta and Blue Owl did not build a record-breaking AI facility just to use it for a few years and abandon it. The business logic strongly implies Meta will use Hyperion for the long haul (20 years or more), meaning renewals are likely. Pretending otherwise is a convenient fiction to minimize reported debt.
Meta downplays the chance its guarantee will ever be needed. To entice bond investors, Meta provided a “residual value guarantee” – essentially promising that if Meta doesn’t renew the lease at some point and no new tenant steps in, Meta will cover the entire remaining debt owed to bondholders (longbridge.com). In plain English: Meta has assured lenders that, one way or another, they’ll get their $27 billion back. This guarantee was crucial to getting Pimco and others to buy the bonds – it’s their safety net. However, this also means Meta is on the hook if the project goes sideways or if Meta ever wants out. Under accounting rules, if it’s probable that Meta will have to pay that guarantee, then that liability should be recognized. Meta’s stance, implicitly, is “Don’t worry, that scenario won’t happen,” which conveniently aligns with not counting the guarantee as a liability. But here’s the rub: This assumption conflicts with the idea that the lease renewal is uncertain. If Meta truly might walk away after four years, then the guarantee very likely would be triggered (forcing Meta to pony up billions). Conversely, if Meta wants to avoid ever paying the guarantee, that implies it will keep renewing the lease to make sure bondholders are paid through rents – undermining the claim that long-term use is in doubt. Meta wants to have it both ways: assure creditors it will backstop the deal while telling accountants those backstop payments will probably never come due.
Believing all these assertions at once requires some mental gymnastics. In order for Meta’s financials to remain pristine, we’d have to accept that Meta lacks true control over a project that is utterly dependent on Meta’s own capabilities; that Meta built a 20-year AI super-complex with a genuine possibility of using it only 4 years; and that Meta’s iron-clad guarantee is simultaneously so solid it makes $27 billion in bonds saleable, yet so irrelevant that it needn’t show up on the balance sheet. It’s a logical paradox, to put it mildly (longbridge.com). No wonder analysts and journalists are skeptical. As Weil dryly noted, “To believe Meta’s books, one must accept all these things — all at the same time.” In reality, Meta appears to be the primary beneficiary of this venture on both counts: it calls the shots and it bears the risks. By the letter of accounting law, that would normally bring the whole shebang back onto Meta’s balance sheet. The fact that Meta insists otherwise shows just how far companies will push complex financial engineering to achieve their desired optics (longbridge.com).
The Real Costs: You Can’t Hide Reality (Forever)
Papering over $27 billion in spending doesn’t make it any less real in the physical world. Meta’s new data center isn’t a financial abstraction – it’s a massive construction project in Richland Parish, already well underway and reshaping the local landscape. By Meta’s own boasts, it will create 5,000 construction jobs and 500 long-term jobs (datacenters.atmeta.com). It will also guzzle unprecedented amounts of electricity to feed Meta’s voracious AI models. Meta has proudly announced deals to add about 1.5 gigawatts of new solar power in Louisiana and beyond to “match the total electricity use” of this facility with renewable energy (datacenters.atmeta.com). (The campus will even have its own on-site solar fields, per renderings.) But here’s the realist’s take: in practice, keeping a 2 GW data center running 24/7 will require a robust connection to the grid and standby power from conventional sources. In fact, the local utility (Entergy) is already fast-tracking new natural gas power plants in the region specifically to ensure reliable juice for Meta’s campus (neworleanscitybusiness.com). AI computing doesn’t just need “clean” power – it needs immense, always-on power, which currently means a big reliance on gas. That underscores an irony: even as tech giants tout renewable investments to hit climate goals, the AI gold rush is sparking a buildout of fossil-fueled generation to keep these machines humming (neworleanscitybusiness.com). The environmental implications are enormous. A federal report estimated U.S. data centers’ electricity use tripled in the past decade and could double or triple again by 2028, gobbling up as much as 12% of all U.S. power (neworleanscitybusiness.com). AI is a major driver of that surge (neworleanscitybusiness.com). The bottom line: Meta’s AI dreams carry very real costs – financial, physical, and environmental – that no amount of accounting alchemy can erase.
From a financial perspective, Meta’s maneuver might succeed for now in keeping investors happy with a clean balance sheet. The bonds found eager buyers, and Meta’s CFO is calling the deal “a bold step forward” (reuters.com). Indeed, by tapping outside capital, Meta frees up funds for other projects and limits its direct exposure if the AI bubble were to deflate (reuters.com). But make no mistake: Meta ultimately is on the hook if something goes wrong at Hyperion. The risk hasn’t vanished; it’s just been shifted around and disguised. Should AI demand fizzle or the project hit trouble, Meta would face hard choices – pay a fortune to bondholders for an unused data center, or double down and keep leasing a facility it might not need. Moreover, accounting gambits have a way of unraveling under scrutiny. If down the road auditors, investors, or regulators decide Meta in substance does control and absorb risk from this joint venture, the company could be forced to consolidate the whole thing onto its books, perhaps suddenly adding tens of billions in liabilities. That would be an unpleasant surprise for shareholders who thought Meta’s debt levels were low. In short, reality has a habit of reasserting itself. The energy will be consumed, the bills will come due, and transparency will eventually catch up with clever structuring.
A Reality Check from The Energy Realist
As the Executive Director of a clean energy nonprofit, I’m all too familiar with the tension between ambitious visions and ground truth. Meta’s case is a prime example. On one hand, they’re pushing the frontier of AI – a technology that just might help solve some of humanity’s toughest challenges in the long run (climate change, fossil fuels killing 8 to 9 million people each year while nuclear kills zero, and nuclear is the one that has to convince everyone that its clean energy is safe enough).
On the other hand, enabling that shiny AI future means confronting massive immediate demands: huge capital investments, huge energy consumption, and not-so-huge transparency. The Energy Realist in me believes we only solve problems by facing them honestly. If AI truly is going to help us navigate out of our current political, economic, and environmental crises, we can’t afford to indulge in fairy tales – be it fairy-tale accounting or wishful thinking about “100% clean” energy overnight. AI needs power – staggering amounts of it – and we need to supply that power in a sustainable, straightforward way. That means companies like Meta should be frank about the costs and risks, and take responsibility for mitigating them, rather than burying them in footnotes and off-book entities. It also means accelerating innovation in clean energy to meet AI’s growing appetite without cooking the planet. There’s an ironic symbiosis here: we have to help AI get the energy it requires (through massive infrastructure and grid upgrades) in order for AI to help us solve bigger problems like climate change.
Finally, a confession – and a bit of irony – from the author: this blog post is itself an attempt to use AI to help us understand how AI’s own financing structures and power demands need to be monitored, stress-tested, and subjected to real due diligence. To write it by myself, I would have spent days rather than minutes. Instead, I leaned on AI to draft and refine, so that I could focus my limited time on the argument and the stakes. AI is here, and we all need to accept that, learn how it works, and start using it where we believe its power can be used for good. As a nonprofit director with limited time and resources, I’m leveraging AI to work smarter and get critical insights out to the public faster. Unlike Meta’s balance sheet, I’ll openly acknowledge my AI augmentation. Consider it a transparent partnership: the human (me) provides direction, judgment and a moral compass, and the AI provides raw power – much like Meta provides the direction and needs to supply the power (literally electricity) for its AI data center. In the end, whether it’s crunching numbers or writing articles, we have to keep it real about who’s doing what and who’s responsible. Meta’s giant off-book data center might be an “aggressive accounting” feat, but the energy – and truth – behind AI will eventually demand reckoning. As we charge forward into an AI-enabled future, let’s do so with eyes wide open, clear accounts, and a firm grip on reality.
Addendum: WSJ Article Summary For Those Who Can't Get Beyond the Paywall:
For readers without access to the Wall Street Journal piece “AI Meets Aggressive Accounting at Meta’s Gigantic New Data Center” by Jonathan Weil (Nov. 24, 2025), here’s a quick summary of its key points:
Meta’s Off-Balance-Sheet Data Center: The WSJ article reveals how Meta Platforms is financing a $27 billion AI super–data center in Richland Parish, Louisiana, using an off-balance-sheet structure. Meta moved the project into a joint venture (JV) with Blue Owl Capital, in which Meta holds only 20%. An entity for Blue Owl’s 80% stake issued $27.3 billion in bonds (mostly bought by Pimco), meaning Meta gets a massive data center built with debt financing that doesn’t appear on Meta’s own balance sheet. Meta will lease the facility for up to 20 years, starting with a 4-year term in 2029.
“Too Good to Be True” Accounting: The arrangement sounds almost magically beneficial for Meta – and the article suggests it “looks too good to be true, and probably is.” Under normal accounting rules, if Meta is the primary beneficiary of that JV (meaning it has control over it and significant financial exposure to it), Meta would have to consolidate the JV’s assets and debt onto its books. Meta argues it will not consolidate the venture, indicating it’s structuring things to avoid meeting the criteria of primary beneficiary.
Questionable Assumptions: The WSJ piece details the convenient assumptions Meta appears to be making to justify this accounting. Meta claims it “does not direct the activities” that most affect the JV’s economics – implying that Meta isn’t really in control of this project (even though Meta is the one with data center expertise, not Blue Owl). Meta also is using a short initial lease term (4 years) with optional renewals to classify the lease as an operating lease, which minimizes the recognized lease liability. They’re assuming it’s not “reasonably certain” that Meta will renew beyond the first term, so they don’t have to count decades of future rent as debt. Lastly, Meta provided a residual value guarantee to bondholders (covering the full debt if Meta doesn’t renew the lease or leaves early), but Meta would only keep that off its books by assuming it’s unlikely to ever pay out on that guarantee.
The Skepticism: These assumptions appear implausible or internally inconsistent. The WSJ points out the contradictions: If Meta isn’t in control, who is? (Blue Owl is a financier, not an operator, so it’s hard to see Meta as hands-off.) If Meta might leave after 4 years, why did it build such a huge facility? And if Meta does walk away, that would trigger the guarantee – meaning a major liability for Meta. On the flip side, if Meta is so confident it won’t ever have to honor the guarantee, that implies Meta expects to stay for the long term (making it reasonably certain the lease continues). Essentially, Meta needs to maintain three things at once – that it lacks control, that it might not stick around, and that its guarantee probably won’t cost anything – which the article suggests “strains credibility.”
“Artificial Intelligence, meet artificial accounting”: The WSJ article concludes with this cheeky line, underscoring the idea that Meta’s clever financing is a form of aggressive (or creative) accounting. It likens Meta’s maneuver to classic structured finance tricks used to keep debt off corporate balance sheets. The takeaway is that while the scheme may technically follow accounting rules (if one accepts Meta’s rosy assumptions), it raises eyebrows. Meta appears to be achieving “irreconcilable financial-reporting goals” – getting a massive AI facility funded and effectively guaranteed by Meta, yet not formally recorded as Meta’s asset or debt. The WSJ piece implies that regulators or investors may question this setup, and that Meta’s portrayal of the deal “looks too good to be true” for a reason.