Resume af analyse fra Morgan Stanley, bearbejdet til dansk
Nvidias vækstmaskine kører fortsat på højtryk, men det seneste regnskab afdækker samtidig en markant ændring i selskabets rolle i AI-boomet. Nvidia er ikke længere kun leverandør af de chips, som driver investeringerne i kunstig intelligens. Selskabet er i stigende grad også blevet finansieringsmotor for kunder, datacentre og infrastrukturselskaber. Nye oplysninger viser fremadrettede garantier og forpligtelser på samlet 530,5 mia. dollar, mens Morgan Stanley advarer om, at en stadig større del af den reelle kreditrisiko ligger uden for den traditionelle balance.
Det er den måske vigtigste underliggende historie i Nvidias seneste regnskab.
På overfladen fortsætter den eksplosive vækst. Datacenteromsætningen nåede 89 mia. dollar mod markedets forventning på 86 mia. dollar, og Nvidia guidede for en omsætning på 108 mia. dollar i det kommende kvartal mod analytikernes cirka 104 mia. dollar. Mest opsigtsvækkende var dog, at CEO Jensen Huang og finansdirektør Colette Kress tog det usædvanlige skridt at give et billede af udviklingen helt frem mod regnskabsåret 2028. Her indikerede selskabet en vækst omkring 70 pct., klart over markedets hidtidige forventning på cirka 45 pct.
Ledelsens argument er, at investeringerne i AI-infrastruktur fortsat accelererer. Ifølge Kress ventes de fem største hyperscalers at øge deres kapitalinvesteringer til knap 800 mia. dollar i 2026 og 1.300 mia. dollar i 2027. Jensen Huang pegede samtidig på næste bølge: AI-agenter, der selv udfører opgaver, vil efter hans vurdering kræve dramatisk mere computerkraft end den hidtidige anvendelse af AI. Han vurderede, at en agent afhængigt af opgaven kan kræve 15 til 100 gange så meget computerkraft som en menneskestyret AI-anvendelse.
Men bag de stærke væksttal tegner der sig en anden udvikling.
Nvidia finansierer i stigende grad sit eget økosystem
Nvidia er begyndt at anvende sin enorme finansielle styrke aktivt til at understøtte opbygningen af den infrastruktur, som selskabets fremtidige salg afhænger af. Det sker gennem investeringer, købsgarantier, lange betalingsfrister, støtte til datacenterprojekter, lejeaftaler og garantier for finansiering.
Ifølge materialet voksede Nvidias forpligtelser til at købe hukommelseskapacitet hos leverandører på blot tre måneder fra 119 mia. dollar til 279 mia. dollar. Hertil kommer omkring 56 mia. dollar i forpligtelser, der skal hjælpe såkaldte neocloud-selskaber med at sikre jord, elektricitet og datacenterkapacitet samt 108,5 mia. dollar i lånegarantier og anden støtte til datacenterudviklere.
Samlet opgøres Nvidias fremadrettede off-balance-sheet-forpligtelser og garantier i materialet til 530,5 mia. dollar, og nogle løber længere end til 2032.
Det er dette, der har udløst diskussionen om såkaldt circular financing: Nvidia medvirker finansielt til at gøre det muligt for kunder og infrastrukturselskaber at købe eller anvende Nvidia-udstyr.
Kritikernes bekymring er principielt den samme som under dotcom-boomet, hvor leverandører af teleinfrastruktur finansierede deres egne kunders køb. Så længe efterspørgslen fortsatte med at stige, fungerede modellen. Problemerne opstod, da efterspørgslen ikke voksede hurtigt nok til at understøtte hele finansieringskæden.
Nvidia afviser imidlertid denne parallel.
Finansdirektør Colette Kress erkendte direkte under regnskabspræsentationen, at nogle investorer vil kalde konstruktionerne cirkulær finansiering. Men hendes argument er, at AI-industrien befinder sig midt i et historisk teknologisk platformsifte, og at de virksomheder, Nvidia understøtter, både har dokumenteret teknologisk styrke og kraftigt stigende brug og kundeefterspørgsel. Ledelsen forventer derfor, at flere af dem vil udvikle sig til nogle af verdens største teknologiselskaber.
Overskuddet vokser langt hurtigere end cash flow
En anden advarselslampe findes i Nvidias pengestrømme.
I kvartalet rapporterede Nvidia et nettoresultat på 56 mia. dollar, mens cash flow fra driften kun var 24 mia. dollar. Forskellen var dermed hele 32 mia. dollar. Et år tidligere var nettoresultatet 26,4 mia. dollar og det operationelle cash flow 15,4 mia. dollar, så forskellen er blevet betydeligt større både absolut og relativt.
En vigtig forklaring er kraftigt stigende tilgodehavender hos kunderne.
Accounts receivable er steget fra 38,5 mia. dollar ved udgangen af januar til 63,1 mia. dollar. Det betyder i praksis, at en voksende del af det salg, Nvidia har bogført som omsætning, endnu ikke er blevet betalt kontant.
Nvidia oplyser samtidig, at udviklingen vil fortsætte. Selskabet tilbyder visse investment-grade-kunder finansieringsordninger og forlængede betalingsfrister under store flerårige aftaler, hvilket også fremover vil påvirke timingen af de operationelle pengestrømme.
Risikoen er desuden koncentreret. Fem kunder tegner sig alene for 70 pct. af Nvidias samlede tilgodehavender. Der er formentlig tale om de helt store hyperscalers som Microsoft, Amazon, Alphabet, Meta og Oracle. På den ene side er det nogle af verdens finansielt stærkeste virksomheder. På den anden side understreger koncentrationen Nvidias meget store afhængighed af et relativt lille antal AI-storkunder.
Aktien og kreditmarkedet fortæller to forskellige historier
Materialets centrale pointe er derfor, at Nvidia i stigende grad skal vurderes forskelligt afhængigt af, om man ser på selskabet som aktieinvestor eller kreditor.
Aktieinvestoren kan fortsat se en virksomhed med eksplosiv vækst, høj indtjening og udsigt til fortsat massive AI-investeringer langt ind i 2028.
Kreditmarkedet fokuserer derimod mere på garantier, indirekte finansiering og potentielle fremtidige forpligtelser. Nvidias femårige credit default swaps, CDS, lå omkring 83 basispoint – mere end dobbelt så højt som i store dele af selskabets historik – efter en kraftig udvidelse af kreditspændene hen over sommeren.
Det er også baggrunden for, at Morgan Stanley nu har indledt kreditanalysedækning af Nvidia med anbefalingen Neutral.
Bankens centrale pointe er, at Nvidias ekstraordinære indtjening og balance er blevet et strategisk instrument til at finansiere AI-industrien. Morgan Stanley beskriver det som en form for “balance-sheet-as-a-service”: Nvidia stiller sin finansielle styrke til rådighed for andre dele af økosystemet. Det skaber muligheder, men samtidig nye risici, som endnu er tidlige, uigennemsigtige og potentielt meget store.
Morgan Stanley anbefaler derfor investorer at holde sig på sidelinjen både i Nvidias obligationer og CDS’er, selv om banken samtidig vurderer, at Nvidias fundamentale kreditkvalitet fortsat er særdeles stærk.
Potentiel kreditrisiko på omkring 170 mia. dollar
Morgan Stanley forsøger i analysen at gå længere end Nvidias traditionelle regnskabsførte gæld og beregne, hvor stor en del af selskabets forskellige støtteordninger der med tiden kan blive betragtet som gældslignende.
Bankens model frem mod kalenderåret 2028 indeholder cirka 170 mia. dollar i ekstra justeringer og eventualforpligtelser ud over den almindelige gæld.
Beløbet fordeles på tre hovedområder:
- cirka 38 mia. dollar relateret til leasingforpligtelser og garantier
- omkring 68 mia. dollar i potentiel residual value support i forbindelse med Nvidias nye finansieringsplatform på mere end 500 mia. dollar
- cirka 64 mia. dollar i illustrativ støtte fra revenue-sharing- og credit-support-modeller.
Det afgørende er dog, at Nvidia – selv med disse meget store potentielle forpligtelser – fortsat står ekstremt stærkt finansielt.
Morgan Stanley vurderer, at Nvidias kontantbeholdning fratrukket den traditionelle gæld kan nå omkring 520 mia. dollar i regnskabsåret 2029. Selv efter indregning af de cirka 170 mia. dollar ekstra forpligtelser vil netto-cash fortsat være omkring 350 mia. dollar. Den samlede gearing bliver i bankens model kun cirka 0,4 gange.
Det er netop paradokset i Nvidia-casen: Forpligtelserne er enorme, men Nvidias indtjening og forventede pengestrømme er endnu større.
Ny finansieringsplatform på over 500 mia. dollar
En central del af Nvidias nye finansielle rolle er et initiativ annonceret den 10. august sammen med blandt andre Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs og KKR.
Formålet er at etablere selvstændige finansieringsplatforme, som over tid skal mobilisere mere end 500 mia. dollar i tredjepartskapital til AI-infrastruktur.
Platformen befinder sig fortsat på memorandum-of-understanding-stadiet, så hverken timing, aktivsammensætning eller finansieringsstruktur er endeligt fastlagt.
Men Nvidia har åbnet for, at selskabet i visse tilfælde kan stille en såkaldt residual-value-support på op til 25 pct. af et konkret projekt. Det betyder grundlæggende, at Nvidia kan være med til at absorbere en del af tabet, hvis den underliggende AI-infrastruktur eller GPU-kapacitet senere viser sig at have mindre værdi end forventet.
Hvis 25 pct. blev anvendt på hele platformen, ville den teoretiske eksponering være omkring 125 mia. dollar. Morgan Stanley anvender mere nuancerede forudsætninger og når frem til, at en sådan eventualforpligtelse kan toppe omkring 90 mia. dollar kort efter udgangen af 2028.
Ratingbureauerne følger området tæt. Ifølge materialet betragter S&P potentiel residual-value-support på op mod 125 mia. dollar som gældslignende, når de underliggende kunder ikke har investment-grade-rating. Moody’s har ligeledes fremhævet størrelsen på de potentielle støtteforpligtelser som en faktor, der kan påvirke Nvidias kreditprofil.
Nvidia kan også garantere minimumsindtjening på AI-kapacitet
Endnu vanskeligere at vurdere er Nvidias nye modeller for revenue sharing, minimumspriser og garantier for usolgt computerkapacitet.
Nvidia introducerede i juli en model, som skal gøre AI-compute tilgængelig for mindre aktører end de største hyperscalers. En mulig konstruktion er, at Nvidia garanterer en minimumspris pr. GPU-time og samtidig får del i indtjeningen, hvis den realiserede pris bliver højere.
Den økonomiske funktion er vigtig: En minimumsgaranti fra Nvidia kan gøre det lettere for datacenterejeren at låne penge til projektet, fordi långiverne får større sikkerhed for, at anlægget genererer en minimumsindtjening.
Morgan Stanley modellerer som illustration fem gigawatt understøttet kapacitet og kommer frem til en potentiel eventualforpligtelse på omkring 81 mia. dollar ved udgangen af 2028, svarende til cirka 64 mia. dollar efter skat.
Der findes allerede et konkret fortilfælde. Nvidia indgik i september 2025 en aftale med CoreWeave om at købe op til 6,3 mia. dollar restkapacitet i cloud-infrastruktur. Dermed fungerer Nvidia i praksis som køber i sidste instans, hvis CoreWeave ikke selv kan afsætte hele kapaciteten.
Op til 105 mia. dollar garanti på ét datacenterprojekt
Et andet eksempel er PORTS-Pike-projektet i Ohio.
Her understøtter Nvidia jord, strømforsyning og datacenterbygninger til et projekt med en indledende kapacitet på omkring 4,25 GW samt mulighed for yderligere 3,75 GW.
Ifølge S&P’s beregninger kan Nvidia levere en garanteret minimumsværdi på helt op til 105 mia. dollar på de ni bygninger, der ventes sat i drift mellem 2028 og 2030. Ratingbureauet forventer, at en del af garantien gradvist vil blive indregnet som gældslignende eksponering, efterhånden som datacentrene færdiggøres.
Morgan Stanley: Nvidia kan absorbere risikoen – men gennemsigtigheden er for dårlig
Trods de voldsomme tal er Morgan Stanleys konklusion ikke, at Nvidia står foran en kreditkrise.
Tværtimod vurderer banken, at selskabets finansielle kapacitet er så stor, at selv meget betydelige forpligtelser kan absorberes. I bankens model kan Nvidia have mere end 1.000 mia. dollar i gældskapacitet, før selskabet rammer S&P’s tærskel for en nedjustering baseret på 1,5 gange gearing.
Problemet er i stedet usikkerheden.
Store dele af Nvidias potentielle støtteordninger er endnu ikke endeligt etableret. Flere befinder sig kun på MOU-stadiet, og investorerne har begrænset indsigt i den præcise størrelse, løbetid og risikofordeling. Samtidig kan mere af finansieringen i fremtiden blive placeret i special purpose vehicles og private finansieringskonstruktioner, hvor gennemsigtigheden er mindre end i et traditionelt børsnoteret selskabs balance.
Morgan Stanley vurderer samlet Nvidias “all-in credit exposure” til omkring 200 mia. dollar ved udgangen af 2028, hvoraf cirka 170 mia. dollar består af balancejusteringer og eventualforpligtelser fra de nye støtteordninger. Samtidig understreger banken, at kreditnøgletallene fortsat vil være solide: omkring 0,4 gange gearing og et frit cash flow efter aktionærudlodninger, som overstiger den samlede gæld.
Den afgørende risiko
Den overordnede konklusion er derfor todelt.
Nvidias seneste tal giver fortsat meget få tegn på, at selve AI-boomet er ved at bremse. Tværtimod peger både omsætningen, guidance og hyperscalernes investeringsplaner mod fortsat meget kraftig vækst.
Men regnskabet viser samtidig, at Nvidia gradvist er ved at ændre karakter.
Selskabet sælger ikke længere blot GPU’er til AI-industrien. Det investerer i kunderne, finansierer infrastrukturen, garanterer lejeaftaler, understøtter datacenterprojekter, tilbyder lange betalingsfrister, garanterer værdien af aktiver og kan i visse tilfælde love at købe kapacitet, som kunderne ikke selv kan sælge.
Det gør Nvidia til noget, der minder om AI-industriens centralbank.
Indtil videre er selskabets indtjening og balance så stærk, at de enorme forpligtelser ikke udgør en umiddelbar finansiel trussel. Men modellen øger sammenkoblingen mellem Nvidia og resten af AI-økosystemet dramatisk.
Og dermed opstår det spørgsmål, som både kreditinvestorer og ratingbureauer nu forsøger at besvare: Hvor meget af den enorme efterspørgsel efter Nvidia-chips er egentlig uafhængig markedsdrevet efterspørgsel – og hvor meget bliver muliggjort, understøttet eller indirekte finansieret af Nvidia selv?
Det er sandsynligvis den vigtigste nye risiko i Nvidia-casen efter det seneste regnskab.
——————-
Uddrag fra Morgan Stanley, WSJ og Zerohedge
While Nvidia’s earnings were impressive, the market did not think they were impressive enough, and it was not until Jensen Huang – who has become quite a consummate stock pumper – shockingly unveiled the company’s full-year guidance for 2028 (something the company had not done before) during the earnings call, forecasting a 70% increase in revenue, far above the 45% consensus, that Nvidia stocks broke the string of 5 consecutive drops on better than expected earnings numbers.
Still, that is not to detract from what clearly a stellar earnings report: take data center revenue, arguably the most important metric for the health of the AI boom, came in at $89 billion against expectations for $86 billion. Or look ahead to Q3, where Nvidia set guidance at $108 billion well ahead of the sellside (if not necessarily buyside) forecast of $104 billion. Or even further. ahead to the 2028 financial year, where as noted above, Huang and CFO Colette Kress took the surprise step of providing guidance more than a year out, which at 70% was sufficient to send the stock surging.
In short, the Nvidia growth story, certainly for the near future, remains intact. Kress says capital investment by the top five hyperscalers is expected to reach nearly $800 billion in 2026 and $1.3 trillion in 2027. Obviously, the company’s profit guidance for 2028 suggests it can see growth well beyond that, even if few others can at a time when Chinese open-weight models – most of which run on Chinese Ascend GPUs – are aggressively grabbing market share from US frontier models.
To justify his euphoric vision, Huang argued several times that AI is now moving from being directed by humans to being directed by AI agents, which would create a sort of snowball effect in demand for Nvidia’s products.
“The amount of compute necessary for an agent versus a human using it is probably 15 to 100 times, depending on the type of problem you’re trying to solve. And so the amount of compute necessary is just extraordinary,” Huang said.
In other words, to quote the Financial Review, “the headline message from Nvidia’s numbers is clear: the AI boom is running on, and Nvidia remains at the heart of it.”
Yet there was something else that was revealed on Wednesday night: digging a little deeper, it immediately became clear that Nvidia is not just the most important supplier to the AI infrastructure build-out. It is increasingly becoming, as the WSJ put it and as we quantified earlier this week, the Central Bank of AI, providing finance, guarantees and deals with customers, suppliers and other key parties.
To be sure, Nvidia has never hidden any of this – on the contrary, the GPU giant proudly announces deals where it invests in AI start-ups and data centres players, where it commits to buy computing power or underwrite leases, or the purchase of land, or guarantee financing, confirming that at its core, the AI ecosystem is one giant circular financing scheme.
But on Wednesday night, we got some actual numbers around those commitments. And they’re big.
As the FR calculates, there are commitments to buy memory capacity from suppliers, which increased from $119 billion to $279 billion in three months. There are $56 billion of commitments to help neo-clouds such as Firmus and Iren to secure the land, power and data centre capacity they need to grow. Then there are $108.5 billion of loan guarantees to help other data centers developers and players buy land, secure power and build data centre shells for future growth, and so on.
Add it all up, Nvidia revealed a staggering $530.5 billion of these off-balance sheet forward commitments and guarantees, which stretch out past 2032. It is an important disclosure and, while the number is very big, it may actually reassure the market; the sheer amount of profit this company is making every three months suggests Nvidia will have little trouble making good on its commitments during the next decade (unless of course, the abovementioned avalanche of Chinese open-weight models derails the unprecedented AI supercycle in the US).
To be sure, this disclosure will only add to fears that Nvidia is engaging in the sort of circular vendor financing that we saw in the dotcom boom, when fiber optic cable giants lent cash to their customers to buy more cable. only for the money-go-round to abruptly stop when demand for internet services didn’t come through as fast as everyone expected. We touched on this most recently in “The Off-Balance Sheet Time Bomb Inside AI Hits $3.1 Trillion: Up $1.3TN In Three Months.”
Nvidia’s CFO, embracing Nvidia’s new role as the Bank of AI, stared the doubters down with a simple defense: the demand only gets bigger from here.
“We recognize the scale of this support, and we know some will call this circular financing,” she said. “We see it differently. We’re going through a major computing platform shift, the creation of one of the most important technologies in human history, and these are once-in-a-generation companies.
“Technology leadership is proven, and their customer traction and usage are skyrocketing. We expect them to become the largest technology companies in history.”
In other words – to use the four most dangerous words in the finance – this time it’s different.
Maybe it is, How much these sort of big statements calm investors’ fears that Nvidia is financing its own demand remains to be seen. And while it is mainly a long-term debate over the future of the stock, there are short-term questions to ponder, too.
Consider that the second-quarter results showed a $36 billion gap between net income at $56 billion, and cashflow from operations of $24 billion. Twelve months ago, that profit number was $26.4 billion and the cashflow number was $15.365 billion – the numbers were much smaller, clearly, but so was the relative gap.
The gap, as we pointed out in our Nvidia earnings post-mortem, reflects the sharp increase in receivables on the Nvidia balance sheet: the stuff Nvidia has sold but not been paid for. That number stands at $63.1 billion, up from $38.5 billion at the end of January…
And Nvidia’s filings on Wednesday night made clear that the buy-now-pay-later deals will continue: “Financing arrangements with certain investment-grade customers, including extended payment terms under large, multi-quarter agreements, will continue to affect the timing of our operating cash flows,” the company said.
While Nvidia clearly has the financial wherewithal to extend terms to key clients, it is also worth noting that the chip giant also said that just five companies – presumably the world’s five big hyperscalers in Microsoft, Amazon, Alphabet, Meta Platforms and Oracle – account for 70% of that accounts receivable balance. Perhaps that is a good thing, given these are among the most powerful tech companies in the world. But it is also another reminder of the concentration risk that exists inside Nvidia’s business.
Nick Ferres, chief investment officer at Vantage Point Asset Management, says what you take from Nvidia’s result depends on whether you’re an equity investor or a credit investor. While equity investors can go away happy, assured that the AI spending boom can run well into 2028, for credit investors, it’s a different matter entirely, and the nagging sense that Nvidia’s circular financing arrangements create another layer of risk simply isn’t going away.
One can see it most clearly in the company’s soaring credit default swaps, which closed just shy of an all time high, at 83bps, more than double where they traded for much of their history, before blowing out in July amid mounting concerns that not all is well with the circular financing scheme.
But don’t take our word for it: overnight none other than Morgan Stanley initiated coverage on Nvidia’s credit (full note available to pro subs), with a Neutral rating, warning that the company’s exceptional growth has turned balance-sheet strength into a strategic AI financing tool – a balance-sheet-as-a-service – and introduces new risks. As MS analyst Lindsay Taylor cautioned, “Our model and relative value are supportive, but we think the tail remains too early-stage, opaque, and sizable to step in.”
Let’s take a closer look at why Morgan Stanley is so cautious.
Introduction
Balance-sheet strength is becoming an ever greater part of the AI financing offering – since almost all hyperscalers are now cash flow negative and rely on debt to fund future growth – and conventional leverage alone no longer captures the full credit story, especially with trillions in debt hidden off-balance sheet. NVIDIA is one of the strongest credits in the AI ecosystem, supported by exceptional earnings growth, FCF generation, and a growing net cash position as the compute cycle broadens across hyperscalers, labs, and enterprises. That financial strength creates substantial capacity, increasingly turning the balance sheet into a strategic advantage – and part of the offering itself.
To be sure, Nvidia is not ashamed to aggressively use its balance to its advantage: recent developments show that capacity is being deployed more deliberately in backstop form. NVIDIA has introduced up to 25% project-level RVS tied to its >$500bn partnership, the ~4.25GW PORTS-Pike shell RVG, and a revenue-sharing / credit-support model. Targeted support can catalyze much larger pools of bank, private-credit, insurance, and public-market capital through asset- and contract-backed structures. Reasonable amortization, strong collateral, contractual offtake, and credit enhancement may make these financings attractive investments in their own right.
Still, instead of recommending the Nvidia credit to its clients, Morgan Stanley has initiated credit coverage of NVDA with a neutral view and the bank urges investors to stay sidelined in both cash bonds and CDS. That’s because, no surprise here, investor conversations continue to center on circularity and vendor-financing concerns, and current spreads arguably price the credit more like a BBB than an AA. To Morgan Stanley, the current ratings are justified and the bank remains constructive on the fundamentals: NVDA remains exceptionally well positioned even after incorporating a broader all-in credit exposure and allowing growth to plateau.
Yet much of the tail risk remains early-stage and difficult to track, with the >$500bn partnership still in MOUs, limited detail on revenue-sharing / credit-support, and potentially limited disclosure as financing migrates through SPVs and private structures.
Deep-Dive Analysis Framework
The bank’s analysis framework looks beyond reported leverage to capture that tail. Through CY28-end, MS combines growth forecasts from its Equity Research team with the credit team’s cash-flow and capital-structure assumptions to distinguish reported debt, minimum likely ratings-adjusted debt, and a broader all-in tail-risk debt measure – without suggesting every contingent commitment ultimately crystallizes. While scale matters, cadence matters too: when shells become ready, when financings are raised and drawn, and how amortization offsets that buildup over time.
The resulting balance sheet remains exceptionally strong, all else equal, as shown in the next section. Cash net of balance-sheet debt could
reach ~$520bn by FY29/CY28, or still roughly ~$350bn after incorporating another ~ $170bn of adjustments and contingent obligations into all-in debt measure. Gross all-in leverage remains just ~0.4x, with FCF after shareholder returns at >100% of gross all in debt. And even if EBITDA and FCF plateau at FY27E levels, those metrics remain reasonable at ~0.7x and ~15% for FY29E, respectively.
That ~$170bn comprises three main buckets, all explored in sections later:
- $38bn of minimum ratings adjustments for lease liabilities, including uncommenced commitments, partner lease guarantees, PORTS-Pike RVG, the CoreWeave capacity arrangement, and other items
- $68bn of potential RVS tied to the >$500bn partnership, based on a new assumption-driven multi-year model incorporating multiple tranches, delayed draws, and amortization; and
- $64bn of illustrative support on 5GW under our new revenue-sharing / credit-support model.
The latter two represent the greatest source of variability and potential tail risk over time; uncertain ratings treatment and potentially limited future company disclosure only add to the analytical complexity and transparency challenge.
Relative Value & Trade
As noted above, Morgan Stanley recommends to “stay sidelined in both cash and CDS” – that’s because fundamentals and relative value are supportive, but the tail remains too early-stage and opaque to step in. The bank’s analysis is more constructive than it expected on NVDA’s ability to absorb support: even after all-in adjustments, the company remains deeply net cash with low leverage, while plateau growth stress still produces sub-1x leverage and ~15% FCF after shareholder returns / all-in debt. Financial capacity remains a powerful anchor, with a FY29/CY28 framework implying >$1tr of debt capacity versus S&P’s 1.5x downgrade threshold.
Relative value is also supportive. NVDA 5Y CDS at ~84/89bp trades ~25bp wide of similarly rated hyperscalers GOOGL and AMZN, while NVDA 2036s are only ~10-20bp inside lower-rated IBM and AT&T paper. As such, the current spreads aren’t seen as fundamentally rich.
Still, according to the MS credit team, it is too early to step in. The >$500bn partnership remains MOU-stage, the revenue-sharing / credit-support framework has limited disclosed detail, and the eventual scale, cadence, and ratings treatment of support remain uncertain. Transparency may also stay limited as exposures migrate through SPVs, contractual arrangements and potentially private financings. Against that backdrop – and with >$1tr of GPU/XPU-related ecosystem financing potentially still ahead – there is continued overhang risk. The bank would want more spread to get involved, with >100bp in CDS and roughly AT&T levels in cash looking more attractive.
If anything, MS warns that CDS could modestly underperform cash given potential lender hedging and thematic positioning. Cash benefits from NVDA’s lack of fundamental funding need, though recent issuance is a less favorable technical versus tender activity at AVGO and MU. The 10s/30s curve appears fair relative to semiconductors and hyperscalers.
NVDA Lease Liability & Contingent Obligation Exposure Forecasts
Shell Leases: Tenant Leases, Lease Backstops, Lease RVGs
Let’s begin with shell-related obligations, where NVIDIA can sit in several different positions – as tenant, future tenant, lease guarantor, or residual-value backstop provider. The structures differ, but the credit question is the same: what exposure ultimately sits with NVIDIA and, importantly, whether and when that exposure becomes debt-like for ratings purposes (as they are currently off-balance sheet).
1) NVIDIA’s own leases: reported and uncommenced. Operating lease liabilities were $2.9bn at FY26-end and $4.3bn LTM, and rating agencies generally include lease liabilities (operating and finance) in adjusted debt (sometimes scaled by 3-4% annually by S&P). NVIDIA also disclosed $32.4bn of uncommenced lease obligations, with 3- to 20-year terms expected to begin between the second quarter of FY27 and FY33. As these facilities are delivered, the associated present-value liabilities should migrate into ratings-adjusted debt. Assuming a 15-year term, 6% discount rate, and 3% annual escalator, the current commitments could ultimately add ~$20.8bn of lease liabilities, with commencements likely front-loaded. Known projects provide some visibility: Tract’s reported NVIDIA-leased Nevada facility, funded by SV RNO Property, involves ~200MW and a 16.4-year lease, with delivery across 2028-2029, while Hut 8 has announced two 352MW, 15-year Beacon Point leases, in May and July, with a high-IG tenant identified in press reports as NVIDIA.
2) Partner facility shell lease guarantees / backstops: NVIDIA disclosed $3.5bn of maximum gross exposure from partner facility-lease guarantees, received in exchange for warrants and declining over 5- to 7-year lease terms; partners also posted $712mn of escrow. An Ionic S-1 filing identified an NVIDIA guarantee of up to $860mn supporting an Nscale facility lease, which may form part of this exposure. Given limited visibility into the timing and future transactions, MS holds exposure – net of escrow and tax – roughly flat, assuming new guarantees offset amortization.
The basic structure is illustrated in the chart below . S&P indicated third-party lease guarantees/backstops are included in adjusted debt when the ultimate beneficiary is non-IG, with the adjustment tax-affected.
3) RVG related to PORTS-Pike Ohio site: On August 17, NVIDIA announced support for land, power, and shell capacity at the PORTS-Pike Technology Campus, with residual value guarantees (RVGs) tied to leases covering an initial ~4.25GW of IT load and an option to extend to another ~3.75GW. Support becomes effective as sites enter service between 2028 and 2030 and declines as lease payments are made. If the unrated tenant defaults, NVIDIA can assume the lease, pursue reletting or sale, or cover the shortfall versus a guaranteed minimum value, subject to reimbursement. Support terminates upon the earliest of the lease’s 20-year anniversary, permitted tenant termination, the tenant achieving a satisfactory credit rating, or other customary events.
This is similar to a partner shell-lease guarantee but technically distinct. The other major data-center shell RVG precedent is Beignet/Hyperion, which is also technically different. There, Meta is both JV investor and tenant under an initial four-year lease with four additional four-year renewal options, while the RVG supports minimum data center value for the first ~16 years of operation.
MS had adopted S&P’s PORTS-Pike debt-adjustment forecast. The agency expects nine separately leased buildings to come online between 2028 and 2030, with NVIDIA providing up to $105bn of guaranteed minimum value (GMV), subject to a pre-agreed schedule. S&P adjusts debt, net of tax, for the gap between GMV and assessed property recovery value, under its commercial real estate methodology, as each building reaches ready-for-service, with the adjustment rising from $4.2bn in CY28/FY29 to ~$37.7bn in CY31/FY32 before declining. Consistent with its hyperscaler webinar guidance, the bank assumes the adjustment will apply as long as the ultimate beneficiary is non-IG. Moody’s is less mechanical, expecting exposure to peak around 2031 without publishing a debt-adjustment schedule, while making maintenance of exceptional financial strength after guarantee exposure an explicit condition for a further ratings upgrade.
Residual Value Support Related to $500bn Partnership
On August 10, NVIDIA announced partnerships to establish independent compute financing platforms with Apollo, BlackRock, Blackstone, Brookfield, GS, and KKR to mobilize >$500bn of third-party capital for AI infrastructure over time (or, as we dubbed it, the launch of collateralized compute obligations). The initiative was launched in response to investor concerns circular around funding capacity, although management has emphasized circular-financing concerns. The concept is not entirely new: many partners were already active in AI infrastructure and NVIDIA-linked financings, including Apollo in the xAI/Valor chip-leasing platform, Blackstone in CoreWeave financing, and BlackRock/GIP in the AI Infrastructure Partnership.
The $500bn remains in the MOU stage, leaving the timing, asset mix, and support structure uncertain. At roughly $50bn/GW, it could correspond to ~10GW of full-stack infrastructure. NVIDIA has described the objective as creating “repeatable financing platforms,” while unconfirmed Bloomberg reports on August 11 and 14 suggest debt may be in the form of private offerings and bonds out of special-purpose entities, issued in tens of billions of dollars at a time, with initial deals potentially coming within months. The reports also indicate the deals will vary according to the customer and shell owner, with collateral expected to be some combination of the underlying chips and offtake agreements. We expect a mix of amortizing, asset-backed, and contract-backed financing across private and public markets, potentially spanning both chip leasing and GPUaaS/compute leasing inclusive of shell.
Management’s emphasis in a corresponding blog that “access to capital is uneven” suggests beneficiaries could skew toward unrated and high-yield counterparties. That blog stated that in “some cases”, it may provide “a residual-value support mechanism for up to 25% of an opportunity, assessed carefully on a project-by-project basis”. NVIDIA-tied financings may require less RVS given broader GPU adoption (vs. other chips), redeployability, and potentially stronger customer diversification and/or quality.
It is unclear at this point how many transactions will ultimately occur, over what time-frame, or how closely future RVS structures will resemble the limited precedent set in large chip and construction financings. To illustrate hypothetical timing and scale of NVDA exposure, MS presents its new assumption-driven RVS model:
First tranche: The bank uses the AVGO-tied TPU financing disclosed in June as a basic template of an NVDA partnership financing, with ~$35bn financing supporting ~1GW of GPU purchases by a chip lease SPV. Assume RVS is limited to a shorter-dated, bank-held A1 tranche, keeping support within the “up to 25%” guidance. Incorporating delayed draws and amortization, one can model that NVDA’s RV contingent obligation valuation would peak very early and then decline with assumed amortization.
Broader platform: MS scales the framework across 15 ~$30bn+ tranches to ~$500bn of financing by CY28-end, in line with the scale of the announced partnership.
While a simple 25% application would imply ~$125bn of tail-risk exposure, MS assumes ~85% of the platform carries 25% RVS, resulting in ~$105bn of cumulative A1 financing. Given the incorporated draws and amortization, illustrative tail-risk exposure peaks near ~ $90bn shortly after CY28-end, reinforcing that peak exposure largely tracks financing raises before declining. Exposure could ultimately be lower if RVS needs decline with improving customer credit quality or if financing ramps more slowly.
Expect company disclosure in quarterly filings to be limited. S&P’s August 18 NVIDIA note characterized the potential $125bn RVS as debt-like, with exposure assessed against underlying collateral value over time. More broadly, on a hyperscaler webinar that day, S&P indicated that chip RVS can result in debt adjustments based on the gap between amortized loan balances and depreciated chip values, if the ultimate beneficiary is non-IG, with the adjustment reflected net of the 21% US corporate tax rate. Moody’s also highlighted the potential ~$125bn support commitment in its August 18 NVIDIA note. While it did not specify a debt-adjustment methodology, it indicated that contingent obligations of this scale could weigh on the credit profile and would be assessed alongside operating performance, leverage, liquidity, and the likelihood that guarantees are exercised.
Revenue-Sharing, Credit-Support, Unsold Compute Model
While credit investors have focused on residual value support given the scale of the $500bn partnership and NVIDIA’s quantified 25% framework, the revenue-sharing, price-floor, and unsold-capacity arrangements could ultimately prove more consequential – and harder to track.
NVIDIA’s July 1 blog introduced a “revenue-sharing and credit-support model” aimed at expanding compute access beyond the largest hyperscalers. Morgan Stanley equity analysts recently highlighted a potential structure in which NVIDIA guarantees a minimum GPU-hour price while sharing in upside revenue above that floor, in addition to standard product revenue.
Economically, these arrangements can create a minimum revenue floor that supports GPU – and potentially shell – financing, helping bridge shorter-duration customer contracts with longer-duration infrastructure debt. The concept, illustrated in the chart below, builds on existing compute-backed financing structures.
Next, Morgan Stanley outlines scenarios where the backstop is and is not triggered. It also models 5GW of supported capacity over the medium term, producing a contingent-obligation estimate of ~$81bn at CY28/FY29-end, or ~$64bn net of tax.
The analysts then assume two-year customer contracts and five-year committed floors, although actual terms could vary. The bank’s Tech Equity team forecasts ~25GW of hyperscaler compute deployment in 2027 ex-TPUs, making another 2-5GW across neocloud/GPUaaS providers plausible as a starting point for our analysis.
The assumption may prove high, given the limited scale of the existing CoreWeave arrangement (est. <500MW), the dependence of support needs on evolving customer credit quality and contract duration, and possible overlap with the $500bn partnership exposure already reflected elsewhere in our model. Conversely, the assumption may prove low if NVIDIA increasingly uses capacity floors to accelerate the $500bn platform, particularly where relatively limited RVS alone is insufficient to unlock financing. The clearest precedent is NVIDIA’s $6.3bn CoreWeave commitment to purchase residual cloud capacity, announced in September 2025, which is modeled as beginning in CY26 and declining over five years. It does not appear to be a revenue-sharing arrangement and may more closely resemble the NBIS/META construct, but think broader unsold-compute backstops belong in the same contingent-support category.
Ratings treatment may be more conservative than initially assumed, specifically that unsold-compute-capacity arrangements might avoid debt treatment given compute contracts are typically considered an operating expense. It appears to that the CRWV backstop is captured in S&P’s current other debt adjustment (tax affected). Plus, on its August 18 hyperscaler webinar, S&P indicated that backstop guarantees – including support for “unpurchased cloud capacity of a counterparty” – can be treated similar to lease guarantees and included in debt-like adjustments when the beneficiary is non-IG, with the adjustment reflected net of the 21% U.S. corporate tax rate.
To summarize, here are the takeaways from the Morgan Stanley analysis:
- NVDA is deploying support more deliberately through RVS tied to the >$500bn partnership, the PORTS-Pike shell RVG, and a revenue-sharing / credit-support model.
- Conventional leverage increasingly understates the credit story as ecosystem support sits in contingent, contractual, and potentially off-balance-sheet forms.
- Morgan Stanley forecasts $200bn of all-in credit exposure by CY28-end, including ~$170bn of adjustments and contingent obligations tied to these support mechanisms.
- Metrics remain robust vs. that all-in gross debt: ~0.4x leverage / >100% FCF after shareholder returns-to-debt, and ~0.7x / ~15% in FY29/CY28 under plateaued growth.
- Spread compensation has improved after the widening, but patience is recommended. Relatively, CDS may underperform bonds given less supportive technicals.
- The key overhang is the unusual scale and limited transparency of >$1tr of guided GPU/XPU-related financing through creative structures in the ecosystem
Pr. mdr.
349 kr.
Få straks adgang til denne artikel og alle andre artikler på ugebrev.dk
- Alle artikler på ugebrev.dk
- Om investering, finans, ledelse, samfundsansvar, life science og Bestyrelsesguiden.dk
- Daglige nyhedsmails med nyheder og analyser
