The article argues that hyperscalers are deliberately routing hundreds of billions in AI capex through joint ventures and SPVs to keep debt off their 10-Ks while retaining operational upside. It cites Meta's $27B Hyperion JV with Blue Owl, Morgan Stanley's estimate that ~$800B of AI infrastructure will be SPV-financed over three years, and CoreWeave's GPU-collateralized loans as echoes of pre-2008 asset-backed structures.
By submitting the Futurism piece and driving it to 318 points, the submitter amplifies the thesis that the true leverage behind the AI buildout is being hidden from investors. The framing 'trying to hide a staggering amount of debt' treats the SPV structures as intentional opacity rather than routine finance.
The piece singles out OpenAI's roughly $1.4 trillion in future compute purchase commitments as a category of exposure that dwarfs the SPV debt itself, yet appears only as footnote 'purchase commitments' rather than balance-sheet liabilities. The argument is that the counterparty risk to Oracle, Microsoft and the Stargate financiers depends entirely on OpenAI's ability to grow into obligations larger than any software company has ever underwritten.
The article acknowledges the industry defense that JV leases and SPV structures are legitimate corporate finance tools that have existed for decades, and that GAAP already requires disclosure of material lease obligations. It presents this position specifically to rebut it, arguing the technical legality is 'materially misleading' because the whole point of the structure is to shift economic exposure off-book while keeping operational control.
A Futurism piece pulled together the accounting mechanics behind the AI capex boom and the picture is uglier than the quarterly earnings suggest. Meta, Microsoft, Oracle, Amazon and Google are on pace to spend north of $400 billion on AI infrastructure in 2025 alone, and a growing share of that spend is being routed through structures that keep the debt off the parent company's balance sheet.
The headline example is Meta's Hyperion data center in Louisiana. Rather than issue corporate debt to build it, Meta set up a joint venture with Blue Owl Capital. Blue Owl owns 80%, Meta owns 20%, and the JV borrowed roughly $27 billion from a syndicate led by PIMCO to fund construction. Meta signs a long-term lease to use the facility. The debt sits on the JV's books, not Meta's. Morgan Stanley estimates that special-purpose vehicles like this will finance about $800 billion of AI infrastructure over the next three years, and roughly half of that will never appear on a hyperscaler 10-K.
The pattern repeats across the sector. Microsoft and BlackRock launched a $30 billion AI infrastructure fund (targeting $100B with leverage) that buys data centers Microsoft then rents. Oracle raised $18 billion in investment-grade bonds in September to help fund its share of the Stargate build with OpenAI. CoreWeave has borrowed against its Nvidia GPUs as collateral in deals that look a lot like the asset-backed structures that seized up in 2008. And OpenAI itself has committed to roughly $1.4 trillion in future compute purchases — obligations that show up as "purchase commitments" in footnotes, not as liabilities.
The standard defense is that these are lease obligations, not debt, and lease accounting has been a legitimate corporate finance tool for decades. That's technically true and materially misleading. The point of an SPV is to move economic exposure off the balance sheet while keeping the operational upside — which is exactly what Enron did with the Raptor vehicles, and what the big banks did with structured investment vehicles before 2008. The instrument is legal. The systemic question is who eats the loss if the underlying asset — a GPU cluster with a 3-to-5-year useful life — depreciates faster than the revenue it generates.
And that's the number nobody wants to sit with. Nvidia's own guidance treats H100/H200-class GPUs as having a useful life of roughly four years. Meta and Microsoft depreciate their AI hardware over six. Google recently extended its server useful life to six years. That gap — call it the "depreciation arbitrage" — flatters near-term earnings and pushes the writedown risk into the outer years. If AI workloads consolidate onto Blackwell and its successors faster than the CFOs modeled, the earlier fleet becomes a stranded asset. The JV takes the impairment. The pension fund LP in the private credit vehicle takes the loss. The hyperscaler walks away from the lease and signs a new one on the next-gen facility.
The private credit angle is where this gets genuinely dangerous. Private credit AUM has roughly tripled since 2019 to about $1.7 trillion, and AI infrastructure is now one of the fastest-growing allocations. PIMCO, Blue Owl, Apollo, and Blackstone are underwriting these deals against future hyperscaler lease payments — the same hyperscalers whose own AI monetization stories are, generously, still developing. Anthropic is on pace for roughly $7B in 2025 revenue against burn that requires continuous fundraising; OpenAI is projecting $13B against compute commitments an order of magnitude larger. The revenue side of the equation is not yet load-bearing.
Community reaction on HN was unusually blunt. The top comment thread compared the JV structures to the pre-2008 conduits banks used to warehouse mortgages, with one commenter pointing out that "the whole point of a special-purpose vehicle is to make it someone else's problem when the special purpose stops working." Others pushed back — reasonably — that Meta and Microsoft have the cash flow to eat a bad AI bet in a way Bear Stearns never had. That's the real distinction. The Mag 7 balance sheets can absorb a lot. The private credit funds and their pension LPs, structurally, cannot.
For working developers this is less about picking sides in a macro debate and more about reading the tea leaves on where compute pricing goes. If you are building on top of a foundation model API or a managed GPU service, your unit economics are downstream of whether these financing structures hold. A few concrete implications:
First, expect compute pricing to stay artificially soft through 2026 while the buildout is still absorbing capital. The financing is committed, the data centers are getting built, and the marginal cost of a token is falling. That's good for anyone shipping AI features. It also means the "just call the API" architecture will keep beating self-hosting for most workloads.
Second, watch the capex-to-revenue ratio at your foundation-model vendor. If OpenAI or Anthropic quietly starts raising API prices, or introducing hard rate limits on cheaper tiers, that's the leading indicator that the private credit side of the equation is asking for its money back. Lock in annual contracts if your usage is predictable. Build abstraction layers (LiteLLM, OpenRouter, your own gateway) so you can swap providers when — not if — pricing rationalizes.
Third, if you work at a company running its own AI infra, the depreciation-schedule question is not academic. Ask your CFO how the finance team is modeling GPU useful life. Six years is aggressive. Four is realistic. The difference between those two assumptions is the difference between a fundable capex plan and a writedown in 2028.
The honest read is that none of this is fraud and all of it is fragile. Off-balance-sheet financing works until the cash flows it's secured against wobble, and then it works in reverse very quickly. The tell to watch is not the hyperscaler earnings — it's whether a private credit fund holding AI infrastructure paper takes a mark-to-market hit that leaks into the press. That's the 2007 signal, not the 2008 one. When it happens, the story will not be about AI. It will be about a Blue Owl or Apollo vehicle nobody outside finance had heard of. Bookmark this piece for that day.
Do they? Is a company with $200 billion annual revenue and earnings (EBITDA) of $100 billion having $420 billion of off-balance-sheet debt really staggering?In many other industries that would be a perfectly normal amount of debt to have. It's only unusual because we are used to tech companies
Are they really "trying to hide" this debt? I think it's pretty common knowledge that a lot of these companies are using debt/bonds for funding. The debt not showing up where the author wants is a reporting formality not an attempt to hide it.
If you're talking about dodgy accounting at hyperscalers, a larger worry might be that they are overstating profits by depreciating their assets (such as datacenters and CPUs/GPUs) too slowly.Estimates are that this could overstate profits by tens of percent. (However, this only allows ear
> Meta alone has amassed around $420 billion in off-balance-sheet debt, according to Nikkei,Isn't this an existential type of bet?
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
As long as this debt does not make it into life insurance and pension funds, we are fine. The trouble is that private credit is taking control of some life insurance companies and off-loads this debt to these. When these fail, it will become everyone's problem.> Risks to financial stability