Analyst(s): Brendan Burke
Publication Date: August 11, 2026
NVIDIA signed memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to mobilize over $500 billion of third-party capital for AI infrastructure. The press release frames NVIDIA compute as an investable asset class. Futurum’s read is that the platforms exist to solve a collateral insurance problem and the DSX reference design is what makes the collateral fungible enough to finance.
What Is Covered in This Article:
- NVIDIA’s August 10 announcement of strategic partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to establish independent compute financing platforms mobilizing over $500 billion of third-party capital over time.
- Jensen Huang puts each gigawatt at $50 billion to $60 billion, Larry Fink sees over 70 gigawatts of demand in the United States alone, and third-party estimates project roughly $7 trillion of AI debt outstanding by 2029.
- GPU depreciation is unpredictable to every lender except the company that controls the roadmap, which is why Futurum expects residual value support to be the mechanism that moves marginal credits over the line.
- The role of the DSX reference design, which Futurum has covered since GTC 2026 as a revenue enabler and modeling resource for data center operators, in making deployed compute bankable.
- The trajectory toward rated securitization, with Larry Fink comparing this moment to the birth of the mortgage-backed securities market in the 1970s.
The News: NVIDIA (NASDAQ: NVDA) announced on August 10 memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to establish independent compute financing platforms intended to mobilize over $500 billion of third-party capital for AI infrastructure over time. The partnerships aim to create dedicated pools of capital at scale and at attractive rates for NVIDIA customers, including frontier AI labs, enterprises, and AI clouds. The agreements remain subject to execution of final documentation.
“We began by building chips; today, we are helping create a new class of productive, investable infrastructure: AI factories,” said Jensen Huang, Founder and CEO of NVIDIA. “These financing platforms will help customers access scarce compute at scale and build the DSX AI factories that will power every industry and country in the age of AI.” Goldman Sachs signaled the credit market ambition directly: “we’re excited for the new opportunity to create a market for credit backed by NVIDIA compute,” said David Solomon, Chairman and CEO of Goldman Sachs.
Is the NVIDIA DSX Reference Design the Real Collateral for $500B in Financing?
Analyst Take: NVIDIA compute financing arrived as a headline number, $500 billion across six of the largest capital allocators in the world, and the temptation is to read it as another entry in the ledger of circular AI deals. Futurum reads it differently. The product being launched here is a cost of capital. Every element of the announcement, from the choice of partners to Huang’s insistence on fungibility, serves a single underwriting argument: that a deployed NVIDIA system is a revenue-generating, transferable asset a lender can size, seize, and resell. That argument has a missing piece, because the resale value of a GPU depends on a roadmap only NVIDIA knows. The platforms are how NVIDIA fills the gap, and the DSX reference design, which Futurum has tracked since GTC 2026 as a revenue enabler and modeling resource for data center operators, is quietly becoming the underwriting document.
The $500 Billion Fills the Gap Between Hyperscaler Cash Flow and Venture Capital
Start with the arithmetic of the buildout. Huang told CNBC each gigawatt of AI factory capacity costs $50 billion to $60 billion, a figure consistent with the $50 billion to $60 billion per gigawatt range Futurum cited in its July analysis of DSX economics, with projections approaching $80 billion to $100 billion as facility costs escalate. Fink said the United States alone will need over 70 gigawatts of power for AI. Multiply those, and the domestic buildout exceeds $3.5 trillion before counting the rest of the world. Futurum estimates $7.9 trillion will be invested in data center capex globally from 2026 to 2030. Against that bill, $500 billion covers perhaps 6% of the projected debt stack. The number is a wedge, sized to prove the asset class rather than fund the buildout.
The wedge matters because of who sits in the middle of the funding stack. Hyperscalers finance capacity from operating cash flow and investment-grade balance sheets. Frontier labs raise equity, and Jensen estimates the world has put roughly $500 billion into AI startups in the past six months. Between those poles sit creditworthy cloud operators, the CoreWeaves, Nebiuses, sovereign clouds, and NVIDIA Cloud Partners who carry enterprise demand but pay venture-adjacent rates for debt.
Jon Gray, President and COO of Blackstone, made the trajectory explicit on CNBC: CoreWeave’s borrowing costs came down dramatically as it seasoned and went public. For a cloud operator, cheaper debt converts directly into deployed gigawatts, because every 100 basis points saved on a $10 billion facility is $100 million a year that buys compute instead of coupon. The partner selection tells the same story: Apollo, KKR, and Brookfield all manage insurance balance sheets with long-duration liabilities that match six-year compute paybacks, and together the six firms steward well over $5 trillion in assets.
GPU Depreciation Is Unpredictable Because NVIDIA Determines It
Collateral drives borrowing costs for data center operators. A lender can underwrite a building over 30 years and an aircraft over 25 because depreciation curves for those assets are actuarial. A GPU depreciates on a schedule set by its own manufacturer. Each new generation reprices the residual value of the last one, and NVIDIA now ships a new generation roughly every year. Banks discount what they cannot predict, and the entity best positioned to predict GPU obsolescence is NVIDIA itself. The rational response, and the one Futurum believes explains the architecture of these platforms, is for NVIDIA to sell depreciation insurance. When the party that controls obsolescence risk retains a slice of it, lenders can extend leverage on terms that were previously unavailable, and marginal deals that died in the credit committee get done. NVIDIA’s insurance of 25% of residual value means a lender at 75% loan-to-value stays whole as long as repossessed systems recover at least half their original value, creating favorable terms for risk-averse lenders.
NVIDIA’s appetite for exactly this risk is already visible. Its DSX revenue-sharing deals with Sharon AI and Firmus include take-or-pay commitments that function as revenue floors on GPU capacity. On CNBC, Becky Quick raised a Wall Street Journal report that NVIDIA would backstop financing for an OpenAI facility in Ohio at a scale of $250 billion; Huang declined to comment on it. Waldemar Szlezak, Global Head of Digital Infrastructure at KKR, supplied the supporting evidence for long collateral life, noting that A100 GPUs remain revenue-generating with high utilization six or seven years after launch. Depreciation has been proven, yet early customers need to prove that the revenue model will endure as well.
Reference Designs Do the Underwriting Work in NVIDIA Compute Financing
Turning compute into an asset class requires something more structural than depreciation insurance. The system architectures will be specified so that NVIDIA can continuously improve deployed systems, run any model on them, and, as Jensen says, “if anything were to happen, somebody else could take it over and operate it.” That is a collateral specification. A lender’s nightmare with a failed data center operator is a bespoke facility nobody else can run. A lender’s dream is an asset as standardized as an aircraft, where repossession leads to remarketing rather than salvage.
The DSX reference design is the instrument of that standardization. Futurum wrote in July that DSX reframes infrastructure competition from cores per dollar to tokens per megawatt, packaging validated rack, power, and cooling designs with the DSX Sim digital twin that lets operators model token output and facility economics before construction. The financing platforms extend that modeling resource to a new audience. A bank underwriting a DSX-certified facility can price projected token revenue per megawatt against a validated design, and an appraiser can mark the asset against a fleet of identical builds. DSX is young but already graduating from an operator playbook to underwriting infrastructure, and reference design compliance is becoming something close to a building code for bankable AI factories.
Fungibility built on a standard design is also what opens the door to structured credit. Fink told CNBC this moment resembles the start of the mortgage-backed securities market in the 1970s and framed data center financing as the next chapter of financial engineering. Szlezak was more direct, describing compute revenue streams that investors can securitize and divide by risk appetite. Standardized collateral plus long-duration contracted revenue is the recipe for asset-backed securities and collateralized loan structures, where tranching converts idiosyncratic project risk into rated paper. Investment-grade tranches unlock the deepest pools in finance, the pension funds and insurers, Fink says, will fund this through both private and public issuance, with $9 trillion sitting in US money market funds by Solomon’s count. Gray’s benchmark frames the ambition: US markets finance $700 billion a year of automobiles and a couple of trillion a year of housing. If compute debt reaches even automotive scale, the buildout stops depending on venture capital tolerance and starts pricing like real estate. That migration, out of the VC market and into fixed income, can unlock the $7.9 trillion we have forecast through 2030.
The Bear Case Is AI Adoption Risk
The credibility of this structure depends on AI adoption. Securitization distributes risk without eliminating it, as 2008 demonstrated when standardized collateral and rated tranches made a correlated bet, the models called diversified. In compute, every facility’s revenue depends on token prices, which Szlezak notes have already fallen 99%, and on demand that Gray measured as a sevenfold increase in LLM usage across Blackstone portfolio companies in six months. If token economics deteriorate faster than efficiency gains offset them, revenue floors get tested across the sector simultaneously, exactly when NVIDIA’s guarantee obligations would come due. NVIDIA also carries an unresolved incentive conflict as the party that books hardware revenue at shipment and pays on residual value support years later. Solomon himself conceded on CNBC that spreads will widen at points, capital will be misallocated, and returns will not all be ample. And the paperwork caveat is real: these are MOUs, and the $500 billion is mobilized over an undefined period through structures that do not yet exist.
The competitive field now includes cost of capital, which changes the scoreboard for every rival. AMD’s Helios racks and its OpenAI warrant structure show it competing aggressively on silicon and on equity-linked deal terms, but AMD has no third-party financing platform, and an operator choosing between architectures will now compare financing spreads alongside tokens per megawatt. Custom ASIC programs are tapping the same capital pools, with Quick citing Apollo and Blackstone financing for Broadcom-linked infrastructure at $35 billion, while Google’s TPUs and AWS’s Trainium are funded from hyperscaler cash flow. The counter-move to watch is whether AMD or Broadcom partners assemble equivalent credit platforms, because if NVIDIA-backed paper prices 200 basis points inside everyone else’s, the financing advantage compounds every quarter it goes unanswered.
What to Watch:
- Whether NVIDIA collateral coverage applies to original cost or marked collateral value, whether the guarantee tenor matches the loan tenor, and what triggers payout.
- Whether DSX reference design compliance appears as an eligibility covenant for a rated ABS or CLO backed by NVIDIA compute leases
- New issues from CoreWeave, Nebius, and NVIDIA Cloud Partners against the roughly 10% unsecured benchmark will show whether the platforms move the market.
- NVIDIA’s guarantee and commitment balances in the upcoming 10-Q filings, and whether usage-linked revenue from DSX sharing arrangements begins appearing in results
- Whether sovereign AI programs route through these six platforms or build national alternatives.
The full announcement is available on the NVIDIA newsroom.
Disclosure: Futurum is a research and advisory firm that engages or has engaged in research, analysis, and advisory services with many technology companies, including those mentioned in this article. The author does not hold any equity positions with any company mentioned in this article.
Analysis and opinions expressed herein are specific to the analyst individually and data and other information that might have been provided for validation, not those of Futurum as a whole.
Other Insights From Futurum:
NVIDIA DSX Promises More Revenue per Gigawatt. Who Actually Captures It?
NVIDIA Engineers the Agentic Data Center with DSX Software Control and Vera CPU Acceleration
NVIDIA Vera Rubin Platform Dominates GTC 2026
Featured Image: CNBC
Author Information
Brendan is Research Director, Semiconductors, Supply Chain, and Emerging Tech. He advises clients on strategic initiatives and leads the Futurum Semiconductors Practice. He is an experienced tech industry analyst who has guided tech leaders in identifying market opportunities spanning edge processors, generative AI applications, and hyperscale data centers.
Before joining Futurum, Brendan consulted with global AI leaders and served as a Senior Analyst in Emerging Technology Research at PitchBook. At PitchBook, he developed market intelligence tools for AI, highlighted by one of the industry’s most comprehensive AI semiconductor market landscapes encompassing both public and private companies. He has advised Fortune 100 tech giants, growth-stage innovators, global investors, and leading market research firms. Before PitchBook, he led research teams in tech investment banking and market research.
Brendan is based in Seattle, Washington. He has a Bachelor of Arts Degree from Amherst College.
