The editorial explicitly rejects the 'Enron with better lawyers' framing as wrong in specifics but 'directionally instructive.' It argues hyperscalers are exploiting the same off-balance-sheet operating lease playbook airlines pioneered in the 1980s, applying it to GPUs and megawatts. The disclosures technically exist in 10-Q footnotes, but $1.65T of committed AI spending invisible to analyst screens represents a systemic transparency failure.
The original Nikkei analysis frames the aggregate obligations across Microsoft, Meta, Alphabet, Amazon and Oracle as 'hidden debts' that soar on 'opaque AI funding.' The framing implies these SPV, JV, and take-or-pay structures constitute real debt-like obligations that investors and analysts are systematically underestimating because they don't appear on standard debt screens.
By submitting the Nikkei piece to Hacker News, the poster signals concern about the opacity of AI infrastructure financing. The submission accumulated 182 points and 55 comments, indicating the developer community broadly views this as a significant and under-reported story worth surfacing.
The editorial highlights that 2026 is the inflection point when ~10 GW of contracted 2024 capacity actually starts drawing power, triggering take-or-pay obligations. Meanwhile developers are already raising for 2028 delivery, compounding the stack. This timing argument implies the abstract $1.65T becomes concrete cash outflows on a predictable schedule that markets haven't priced in.
A Nikkei Asia analysis published this week pegs the combined off-balance-sheet, AI-related obligations of the five largest US hyperscalers — Microsoft, Meta, Alphabet, Amazon and Oracle — at roughly $1.65 trillion. That number is not sitting in the debt line of any 10-K. It lives in special-purpose vehicles, joint ventures, sale-leaseback arrangements, and long-dated take-or-pay contracts with data center developers who themselves raised the money in the private credit markets.
The mechanics are worth spelling out. A hyperscaler wants 2 GW of new capacity in Virginia. Instead of putting $30B of capex on its own balance sheet, it signs a 15-year lease with a developer — Blue Owl, Stack, QTS, DigitalBridge — who then goes to Apollo, Blackstone, KKR, or a Middle Eastern sovereign for the financing. The debt lives at the developer. The obligation to pay it back, contractually, lives at the tenant. Under current GAAP, long-dated operating leases and unconsolidated JVs disclose as footnotes, not as debt — which is why a $1.65T stack of committed AI spending shows up on almost no analyst screen.
Meta's Hyperion campus in Louisiana, Microsoft's Stargate JV with OpenAI, Oracle's Abilene site, Amazon's Project Rainier for Anthropic — every one of them is structured this way. The number keeps climbing because 2026 is the year the ~10 GW of contracted 2024 capacity actually starts drawing power, and the developers are already raising the next round for 2028 delivery.
The cynical read is that this is Enron with better lawyers, and the cynical read is wrong in the specifics but directionally instructive. It is not fraud — the disclosures exist, footnoted, in the 10-Qs. It is regulatory arbitrage. Operating leases got kicked off-balance-sheet in the 1980s specifically so airlines could finance planes without blowing up their debt ratios; hyperscalers have taken the same playbook and applied it to GPUs and megawatts. The FASB's ASC 842 revisions in 2019 were supposed to end this. In practice, the classification games moved one layer out, into the JV and SPV structures.
The deeper problem is that the AI capex cycle has become a leveraged bet where the leverage is intentionally invisible to the equity markets that are funding the rest of it. Retail and index investors see a Microsoft balance sheet with $80B of debt and $75B of cash and conclude the company is bulletproof. They do not see the $400B+ of contracted lease obligations behind it. The rating agencies see it — Moody's flagged Meta's Hyperion structure in June — but the equity narrative has not caught up.
Compare this to how the same companies financed the last capex boom. In 2015-2019, Microsoft, Google and Amazon built out the first cloud regions largely with on-balance-sheet debt and retained earnings. Capex was $20-30B a year each, and it showed. This cycle, capex guidance for 2026 is $80-100B per hyperscaler, and only about half is landing on the balance sheet. The other half is the SPV stack.
The community reaction on Hacker News was predictably split. One camp treats this as a bubble marker — "when the financing gets creative, the top is in" was the top comment. The other camp argues it's rational: if you can lock in 15-year power and land at today's prices while your model roadmap is uncertain, of course you push the debt out to a counterparty who specializes in it. Both readings can be true. The 2000 telecom buildout was structurally identical — dark fiber financed through SPVs and vendor loans — and the fiber turned out to be useful, eventually, after the equity holders were wiped out.
There is also a concentration angle nobody is talking about enough. The private credit funds writing these checks — Apollo, Ares, Blue Owl — are themselves financed by insurance-company balance sheets that got restructured post-2008. A hiccup in AI demand doesn't just hit Nvidia's multiple; it propagates through data center REITs, private credit vehicles, and life insurance liabilities in ways that regulators have not stress-tested. The Fed's April 2026 financial stability report used the phrase "opaque interconnection" three times.
If you are a senior engineer making infrastructure decisions right now, this reshapes how you should think about vendor risk. The public story is that Azure, GCP and AWS have infinite capacity and infinite runway. The actual story is that their capacity is contractually locked into 15-year leases that assume revenue growth that has to materialize. If it doesn't, the response won't be a bankruptcy — none of these companies are going bankrupt. The response will be repricing. Reserved instance discounts will get less generous. Committed spend agreements will get shorter. GPU quotas will get tighter for anyone not on a multi-year commitment.
The practical move: if you're running a workload where a 2x price increase in inference compute would break the unit economics, you need a portability plan now, not in 2027. That means abstracting your model layer (LiteLLM, OpenRouter, or a homegrown router), keeping at least one workload live on a second provider, and treating any single-provider committed spend agreement as a hedge instrument, not a discount. The hyperscalers' financing structure has priced in demand growth you may not deliver — when the reconciliation happens, you don't want to be the marginal customer paying the marginal rate.
Also worth watching: the neoclouds — CoreWeave, Lambda, Crusoe — are financed the same way but with less balance sheet behind them. Their pricing is aggressive because their debt service is aggressive. That's fine while GPU demand exceeds supply. It becomes a counterparty risk the moment it doesn't. If your production inference is on a neocloud, know who holds the paper on the data center your GPUs are sitting in.
None of this predicts a crash. The AI buildout may well earn its cost of capital, the leases may get paid, and in 2035 we'll look back at $1.65T as a rounding error. What it does predict is that the next 18 months of hyperscaler behavior will be shaped less by product strategy and more by lease coverage ratios. When Microsoft or Meta start pushing customers hard toward multi-year commits, or when a neocloud quietly gets acquired by its landlord, that's the financing structure asserting itself. Build your architecture assuming that pressure is coming.
Archive link:https://archive.ph/20260720174223/https://asia.nikkei.com/bu...
"hidden" debts and "opaque" funding I seem to read about every other day.
On the one hand, these debts may be off the balance sheet, but institutional investors certainly know about them and can reason about the company's valuation. Retail investors may be caught out slightly more.But on the other hand, these companies are essentially paying for the service of taking
To quote the Big short: "I have five houses and a condo."
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
Well technically they don’t own the debt, the SPVs that own the data centers do. The giants just have long term commitments, but if shit hits the fan, it’s not the tech giants but the banks that lent the money to the SPVs that are at risk. This usually means all of us are on the hook.