Goldman: Skjult 1,8 billion USD stor off-balance sheet-tidsbombe skaber tvivl om AI-supercyklussen

Resume af Goldman Sachs analyse bearbejdet til dansk:

Goldman Sachs vurderer, at markedets forventninger til hyperscalernes AI-relaterede investeringer i 2027 er for lave. Konsensus peger på omkring 920 mia. dollar i capex, men Goldman mener, at tallet kan nå 1.100 mia. dollar — og i et mere aggressivt scenarie op til 1.400 mia. dollar. Pointen er, at AI-udbygningen kan blive en af de største kapital- og industricykler nogensinde.

Rapporten fremhæver, at finansieringen i stigende grad sker via gæld. Morgan Stanley anslår, at der allerede er udstedt 236 mia. dollar i AI-relateret gæld år til dato, og at global AI-relateret gældsudstedelse kan nå 570 mia. dollar i 2026. Samtidig er hyperscalernes gearing steget markant, og selskaber som Amazon, Meta, Alphabet, Microsoft og Oracle fylder nu mere i kreditmarkederne.

Den største risiko ligger dog ikke kun i den synlige capex, men i de enorme forpligtelser uden for balancen. Morgan Stanley peger på omkring 1.000 mia. dollar i købsaftaler og mere end 800 mia. dollar i ikke-igangsatte leasingforpligtelser. Disse poster fremstår ikke nødvendigvis som gæld på balancerne, men repræsenterer reelle økonomiske forpligtelser knyttet til AI-udbygningen.

Analysen beskriver også en voksende brug af SPV’er, private credit og “chip-backed” finansiering, hvor leverandører og finansielle aktører bruger hyperscalernes fremtidige forpligtelser som sikkerhed. Det skaber en mere uigennemsigtig finansieringskæde, hvor risikoen flyttes væk fra de store tech-selskabers balancer og over i leverandører, private credit-fonde og specialkonstruktioner.

Et centralt problem er, at investeringerne løber hurtigere end den dokumenterede indtjening fra AI. Capex-forventningerne er blevet kraftigt opjusteret, mens salgs- og cashflow-forventninger ikke er fulgt med i samme tempo. Samtidig udskydes en del af indtjeningspresset, fordi mange investeringer endnu ligger som “construction in progress” og derfor ikke fuldt ud rammer afskrivninger og marginer.

Konklusionen er, at den synlige AI-investeringsbølge kun er toppen af isbjerget. Den reelle økonomiske eksponering er langt større, når man medregner købsaftaler, leasingforpligtelser, SPV’er og leverandørfinansiering. Risikoen er ikke nødvendigvis en akut solvenskrise, men snarere en kombination af stigende gæld, skjult gearing, fremtidigt afskrivningspres og en afhængighed af, at AI-indtægterne vokser hurtigt nok til at retfærdiggøre investeringerne.

Uddrag fra Goldman Sachs:

Yesterday Goldman raised numerous eyebrows across Wall Street with a new report in which it stated that “consensus 2027 hyperscaler capex estimates are too conservative.” 

According to Goldman’s Ryan Hammond, analyst estimates imply hyperscaler capex will equal $920 billion in 2027, representing a sharp deceleration in growth from 84% in 2026 to 22% in 2027. Goldman disagrees, and estimates that if incremental investment reaches 2-3% of GDP, similar to the build-out of railroads and autos, hyperscaler capex would reach roughly $1.1 trillion in 2027 (45% growth). But that’s just the start: in a more extreme upside scenario, Goldman calculates that between hyperscaler cash flow generation and investment grade credit market capacity, there is room for potentially $1.4 trillion in capex (89% growth).

Yet as Goldman also pointed out in the same report, this massive spending increase – which is based on the premise that the LLM sponsors of the AI revolution such as Anthropic and ChatGPT will be able to keep rising token prices indefinitely even though they are already getting substantial pushback from corporations which are increasingly looking toward much cheaper offerings from China as the WSJ reported – will soak up virtually all of hyperscaler cash from operations by the end of 2026…

… at which point the only incremental source of capital will be even more debt (as a reminder, in late 2025 we discussed that AI debt issuance is already a bubble in itself, one which is growing dramatically via both public, private and SPV “project-financing” conduits).

As Citi’s head of equity trading strategy said Friday in an interview with Bloomberg TV, “A lot of investors are looking at this as one of the largest construction and industrial production cycles the world has ever seen,” adding, in case it wasn’t obvious, that all “that needs to be financed.”

So where will the money come from to fund the $1.4 trillion (and much more in subsequent years) capex buildout? 

This report looks at two things: first, is the current state of the public debt capital markets, where AI is all the rage. Second, and more important, we analyze what may be a ticking time bomb in the entire capex and AI supercycle story: the $1.8 trillion in off-balance sheet commitments, circular deals and SPVs, which nobody is closely looking at yet which could prove to be the pivotal sticking point in the near future should the massive revenue expected from LLMs never materializes. A vivid example of this was the just announced deal in which Apollo and Blackstone raised $35 billion in a “chip-backed” SPV for Anthropic in one of the biggest ever private credit deals, which we profiled earlier this week, and which we summarized as follows:

Broadcom is backstopping a massive $36 billion private credit SPV with Apollo and Blackstone, which will help Anthropic buy Google chips… made by Broadcom. Oh, and yes: Google owns 14% of Anthropic. But wait, there’s more… because if that wasn’t enough, Morgan Stanley, which advised Broadcom and arranged the transaction, is also lending money to investors participating in the deal! And just because this is a “chip-backed” off-balance sheet SPV where nobody really knows who holds the debt, the monstrous circularity of all the deal aspects will be ignored until the AI credit bubble cracks. 

But more on that shortly. First…

AI in the Public Debt Markets

We start with a snapshot of where the AI debt capital markets are, which as Morgan Stanley writes in its latest “AI Debt Financing Tracker” (full note available to pro subs), markets are “warming up for a hot summer.”

Here is where AI stands in debt capital markets as of the end of May

April Ran Hot: April saw the highest supply YTD with >$74bn of AI-related issuance. Project finance structures to fund construction of data center shells accounted for 85% of the AI-related HY supply and 40% of the IG. “May” Be Catching Its Breath: We saw limited new issuance in the US markets in May, but hyperscalers have been broadening their investor base through non-USD issuance and collectively issued ~$24bn of debt in other currencies (EUR, CAD, CHF, and JPY). These issuers’ relatively smaller representation in EUR/GBP benchmarks (vs. USD) leaves room for more non-USD supply, as hyperscalers expand and diversify their investor base. Technicals Over Fundamentals: Fundamental backdrop remains strong, but for now the price action is being mostly driven by supply expectations. The market has broadly tightened across asset classes since end of 1Q despite elevated supply. Morgan Stanley forecasts ~$570bn of AI-Related Global Supply in 2026: We’re at ~$236bn YTD (until 5/31), >4x more than global AI-related issuance during same period in 2025. That’s just the start: issuance is expected to accelerate in 2H26 as the abovementioned $1+ trillion in capex needs to be financed mostly through debt . Chip Financing in Focus: We are starting to see more chip financing activity in public and private markets. Deals so far have come with shorter maturities vs. some data center construction transactions, and full amortization. We expect to see increasing investor demand for these structures

Here are the key charts, starting with YTD and projected debt issuance statistics: as noted above, YTD $236BN in AI-linked debt has been issued, a 357% increase from the same period last year. By year-end, MS expect this number to more than double to $570 billion.

There is a broad value dispersion in the issuance universe, with a broad distribution between structures and credit quality.

Here, Morgan Stanley shares its own capex forecast for 2027, which while not nearly as aggressive as Goldman’s upside case of $1.4 trillion, matches Goldman’s base case of $1.1 trillion. Note the dramatic increase from the previous (Nov 2025) forecast: the increase from $686 billion to $1.126 trillion is nearly double.

Incidentally, here is Morgan Stanley’s FCF forecast for the Hyperscalers: while almost all had positive cash flow in 2025 (ex. ORCL), in 2026 Amazon and Meta will both have flat or negative free cash flow, with just Alphabet and Microsoft solidly in the green.

While debt issuance dipped modestly in May from April’s record, the ramp up in hyperscaler issuance is unmistakable (from $20BN in 2025 to $159BN for the 2026 period). Furthermore, just 5 companies (AMZN, META GOOGL, MSFT and ORCL) now account for 4% of the entire IG index.

Perhaps the most important chart in the presentation, is the one which shows the dramatic increase in hyperscaler gross leverage, which has surged from 0.9x in Q3 ’25 to 1.8x currently, doubling in just over two quarters, and surpassing the gross leverage of the entire energy sector. At this rate, hyperscaler debt is growing at about 0.3x turn per quarter.

Not surprisingly, as a result of the surge in combined leverage, hyperscalers are drifting wider, and after trading insider AA spreads for much of 2025, are now on top of A, and as MS warns, “may widen further on supply.” And it’s not just outlier Oracle: META is now trading wider to CDX IG.

Another surprising finding: as Amazon reminded us earlier this week when it issued the largest ever bond in Canada’s market, the hyperscalers are increasingly funded by non-USD issuance.

Away from investment grade, we see a frenzy of new issuance in the junk bond market, where there have been 12 data center construction deals YTD, vs just 4 in all of 2025. It is here that we observe the credit-support, primarily from Google, for the new issues such as Meridian Arc, Flash Compute, Cipher Digital and Terawulf, to lower the cost of capital.

Putting it all together, the AI debt bubble we first profiled in late 2025 has gotten much bigger, and with as much as $1.4 trillion in capex to be funded next year, it will only get much bigger. 

But while the public debt market – and disclosed capex – are arguably the two most important components of the AI supercycle, they are just the tip of the iceberg.

Which brings us to part 2.

SPVs And Off-Balance Sheet Commitments

In a separate, must-read report from Morgan Stanley’s Todd Castagno, titled “AI Ecosystem – Charting Recent Trends“, and looking at the more murky, less discussed, aspects of the AI ecosystem, he explains why merely looking at the projected capex spending – whether $1.1 or $1.4 trillion – is missing arguably the most important number: “as hyperscaler capital intensity reaches more than 40% of sales, incremental capex revisions are now accompanied by capital raises and/or innovative financing.” Which brings us to the elephant in the room: $1.8 trillion in off-Balance sheet operating leverage, to wit: “Long-term purchase commitments of ~$1tn and lease commitments >$800bn further support the AI buildout, but add off-BS operating leverage.” In other words, while the capex spending number is striking, what is even more shocking is how much more money is pledged for future commitments and is engineered through financial accounting, never appearing on anyone’s balance sheet.

This is how Castagno frames the problem:

We are entering a record-setting capital cycle: Hyperscalers are pouring billions into the AI build-out, with finance lease usage pushing capital intensity even higher, well past the dot-com record. Capex estimates continue to revise higher, while sales revisions lag.

As investment grows, depreciation is set to rise: A sizable increase to depreciation expense could pressure margins if nondepreciation costs don’t decline or if sales don’t revise higher. But before these investments become expenses, much of the capital will sit in Construction in Progress (CIP), thus delaying earnings and margin impact.

Billions in lease and purchase commitments are fueling the AI buildout: Hyperscalers are securing capacity amid surging demand through leases and purchase commitments, creating a timing gap between monetization and supplier payment, leading to a rise in DPO among hyperscalers

Let’s take a closer look.

The report published by Morgan Stanley’s Global Valuation, Accounting & Tax group, tracks the accounting mechanics of the AI spending boom (beyond just AI), with a focus on how much of the buildout is being financed through arrangements that don’t appear as debt on hyperscaler balance sheets.

The core picture

Hyperscaler capital intensity is set to exceed the dot-com peak. Morgan Stanley forecasts capex-to-sales ratios of 36%, 44%, and 42% across 2026–2028, well above the ~32% reached during the late-1990s fiber buildout.

Source: AI: Capital Intensity Enters a New Era

AI-related spend is expected to exceed 50% of Russell 1000 capex in 2026, with hyperscalers (AMZN, GOOGL, META, MSFT, ORCL) contributing roughly 90% of that. The scale is large enough that the financing structure, not just the spending level, becomes the analytical question.

Source: AI: Capital Intensity Enters a New Era

The off-balance-sheet mechanics

Here is the argument in a nutshell: it’s not just the (massive) committed capex; a growing share of committed AI spending sits off the balance sheet until goods or services are actually delivered. Three buckets matter:

1. Purchase commitments across the hyperscalers plus Nvidia have reached roughly $982bn. Critically, unless a company expects a loss on these contracts, the obligations stay off-balance sheet until the underlying goods are delivered and a payable is recognized. So nearly a trillion dollars of contracted future outflow doesn’t appear as a liability today.

Source: AI: Off Balance Sheet, On the Hook and AI: Capital Intensity Enters a New Era 

Nvidia’s own inventory and purchase obligations have risen to ~32% of consensus FY27 revenue, up from a 15–20% historical range, as  supply-chain commitment risk now extends to the chip vendor, not just the buyers.

Source: AI: Off Balance Sheet, On the Hook

2. Lease commitments not yet commenced have hit about $822bn. Over $800 billion in lease commitments highlight the growing strategy by Hyperscalers to quickly expand AI buildout, bypassing conventional channels.

Source: AI: Off Balance Sheet, On the Hook

Leases that are signed but not yet started are just one example of off-balance sheet commitments: variable lease payments, renewal options, residual value guarantees, and third-party lease backstops and guarantees are other common categories that escape the lease liability line entirely.

Source: AI: Off Balance Sheet, On the Hook and AI: Leasing the Future

For a sense of scale, variable lease payments exceed 10% of lease costs at AMZN, 25% at META, and 30% at GOOGL (MSFT and ORCL report none, likely due to materiality judgments).

Source: AI: Off Balance Sheet, On the Hook

Finance leases in particular push “true” capital intensity meaningfully higher than headline capex suggests: Morgan Stanley estimates that including finance leases, MSFT’s capex-to-sales could surge from 33%/50% (FY26/FY27) to 44%/64%, and ORCL’s from 76%/115% to as high as 101%/189%.

Source: AI: Off Balance Sheet, On the Hook and AI: Capital Intensity Enters a New Era 

3. Unpaid capex embedded in accounts payable and accrued expenses now totals around $110bn, with days-payable-outstanding rising sharply versus history (ORCL +370%, META +73%, MSFT +69%). This stretches working capital and signals a widening timing gap between when capacity is contracted and when suppliers are paid. It is done to buffer the risk from projected revenues not rising fast enough to offset off B/S commitments. It also means that the whole supply chain is effectively financing the AI build‑out.

Source: Hyperscaling Payables

The financing chain risk

The most systemically interesting point is structural: suppliers are using purchase and lease commitments from hyperscalers and Nvidia as collateral to access bank and private-credit financing (similar to the Broadcom/Google example discussed above). In effect, a hyperscaler’s off-balance-sheet commitment for “chip-backed” deals, becomes the basis for a supplier (data center, fab, power infrastructure) to take on debt.

Source: AI: Off Balance Sheet, On the Hook

This pushes leverage out of the visible hyperscaler balance sheets and into the supplier and private-credit ecosystem (Apollo, through its Athene insurance company has been especially active here, selling annuities to retirees and using the funds to capitalize SPVs such as the Broadcom “Big Sky” deal), where it is intentionally much harder to see and aggregate.

We noted earlier that there has been a surge in “chip-backed” Special Purpose Vehicle deals, which have explicit Residual Value support from hyperscalers. The reason for that is that whether GPU contracts count as leases or services is subjective, and companies use that flexibility to shift billions off the balance sheet. And since the lease-classification judgment is doing real work, below we show the indicative decision tree whether a compute-capacity deal contains a lease, noting the inconsistent disclosure (Nvidia disclosed $30bn of cloud service commitments, Oracle $10bn, while Meta declined to quantify the cloud-capacity portion of its $238bn). That inconsistency means cross-company comparisons of “true” leverage rest on differing materiality and classification choices.

Source: AI: Off Balance Sheet, On the Hook

Here, Morgan Stanley’s well-known circular financing interconnection map (OpenAI, Oracle, Nvidia, Microsoft, CoreWeave, AMD, Amazon, with circular customer / investor / vendor-financing / repurchase relationships) underscores the concern that the same dollars and the same counterparties recur throughout the chain, with SPV facilitating the circularity of the relationships.

Source: Mapping The AI Ecosystem

Risk factors

Surging capital spending, especially when it is being used to purchase a rapidly depreciating asset such as a computer chip (which serves as the basis for all “chip-backed” SPVs), is a problem. That’s why, the Morgan Stanley team flags the hyperscaler sector’s depreciation overhang as “next margin watch item.” Cumulative depreciation for MSFT, ORCL, META, and GOOGL is expected to exceed $520bn over three years.

Source: AI: Appreciating Depreciation Flows and AI: Depreciation Flows Update

As depreciation rises as a share of revenue (ORCL potentially from 7% to 28% by FY28, META from 9% to 19%), every other cost line has to fall just to hold margins flat – and that only works if sales revise up to match, which so far they haven’t. And since depreciation will explode higher in coming years to account for both rising capex and aging chip inventory (a core pillar of Michael Burry’s bearish AI thesis), hyperscalers will find it ever more difficult to maintain the stratospheric margins Wall Street is accustomed to.

Source: AI: Appreciating Depreciation Flows and AI: Depreciation Flows Update

Worse, as noted above, Capex growth is outpacing the revenue and free-cash-flow growth that’s supposed to justify it. Capex revisions have climbed far faster than sales and FCF revisions, meaning the spending is being committed ahead of demonstrated monetization, which is perhaps the most important conclusion from this note.

GOOGL consensus capex revisions for 2026 have increased 139% vs one year ago (June 2025), with META and AMZN up 85% and 81%, respectively. ORCL had largest jump with revisions up 175% from one year ago.

Meanwhile, the earnings hit is being deferred, not avoided. Rising Construction-in-Progress balances (ORCL ~+200%, META ~+90%, GOOGL ~+55% YoY) mean capital is sitting in CIP and not yet depreciating, so reported profit margins are artificially inflated and dramatically understate the future expense load already locked in.

Source: AI: Appreciating Depreciation Flows and AI: Depreciation Flows Update

There’s more: the AI ecosystem has found yet another way to shuffle the same dollar multiple times among the handful of circular actors: customer concentration sits beneath a massive RPO number. Remaining Performance Obligations across major AI players exceed $2tn (ORCL up >300% YoY, GOOGL +406%, MSFT +97%), but that backlog is concentrated in a handful of large, long-duration contracts, which concentrates counterparty risk; in other words, if just one of these circular actors stumbles, it will instantly drag down the entire ecosystem with it. 

Source: AI RPO Revealed 

Bottom line for the off-balance-sheet question

We started this long article by highlighting the “big” $1.4 trillion 2027 capex projection by Goldman. As it turns out, that’s just the tip of the iceberg as headline capex figures materially understate the economic commitment to the AI buildout. Only when you add ~$1tn of purchase commitments, ~$822bn of not-yet-commenced leases, finance-lease effects, and supplier debt collateralized by hyperscaler commitments, does one start to get a sense of the full size of AI balance sheets, and how much is truly at stake if the AI revolution fails to materlize in the form of record AI revenues to cover the massive on and off-balance sheet commitments.

The good news is that, for now, the risks aren’t an imminent solvency problem, but a set of timing and disclosure mismatches – a deferred depreciation wall, capex running ahead of monetization, leverage migrating into the supplier/private-credit layer, and classification judgments that make true capital intensity hard to compare across companies. For their part, hyperscalers have taken advantage of the current moment of market euphoria to raise as much capital as they can, aware that the window will eventually shut. The question is when sentiment turns, and when all these funding conduits are shut, will there be enough funding to sustain the AI revolution for the foreseeable future. Of course, this question becomes moot if demand shifts, and instead of buying the latest and greatest offering from Anthropic or OpenAI for a stratospheric number of tokens, consumers and enterprises turn to dirt-cheap alternatives from China which offer 90% of the performance of frontier models for a fraction of the cost, then all bets are off.