A company that sells chips has stepped in to act as a guarantor for someone else’s facility lease, promising to pay up to $105 billion if the tenant defaults. That liability cap is roughly equal to an entire year of NVIDIA’s own operating cash flow ($102.7 billion in fiscal 2026). On August 17, when NVIDIA, OpenAI, and SoftBank’s SB Energy announced their transaction in Ohio, this guarantee was the main event.
You might have seen these words in news headlines, but reading them together can be baffling—or even alarming if terms like “circular credit” or “subprime crisis” start getting thrown around. In reality, the questions on everyone’s mind boil down to just a few: Why is NVIDIA providing this guarantee? How does this money loop back into its own pockets? Does it truly resemble the subprime crisis of 2008? And if the industry cools down, who is in the greatest danger?
Two days ago, we discussed how the focal point of competition among AI companies has shifted from GPU counts to energization dates (After GPUs, AI Companies Are Fighting for Energization Dates). Today, energization dates must be fought for, and the capacity to pay for those energization dates relies on credit guarantees. The scarce resource itself is morphing into credit.
The easiest way to understand this deal is through a renting analogy. OpenAI plans to lease a mega compute facility at the PORTS-Pike site in Pike County, Ohio, for a 20-year term. According to The Wall Street Journal, the site’s total cost—including chips—exceeds $500 billion, making it the largest single data center project ever announced. The facility will be funded, owned, and operated by SoftBank’s SB Energy.
Here lies the problem: even though OpenAI’s business is growing rapidly, in the eyes of banks it looks like a tenant without proof of income. The company is still unprofitable overall and lacks a public credit rating. To build a facility of this magnitude, SB Energy needs to borrow heavily from banks. When lenders see an unrated tenant, they either demand exorbitant interest rates or refuse to lend altogether.
That is when NVIDIA stepped in. As disclosed in its 8-K filing with the SEC, NVIDIA signed multiple residual value guarantee agreements with SB Energy covering an initial phase of approximately 4.25 GW of IT compute capacity, capped at a maximum cumulative liability of $105 billion. This compensation ceiling is nearly identical to NVIDIA’s own operating cash flow for the entirety of fiscal 2026 ($102.7 billion).
This guarantee comes with strict prerequisites and triggers. The facility must first meet delivery standards, with phased execution expected to begin in 2028. Payouts are triggered only if OpenAI defaults due to bankruptcy or fails to pay rent, and even then, the guarantee covers only the shortfall: if re-leasing or selling the facility fails to recover the minimum guaranteed value, NVIDIA pays out the deficit. As data center space is delivered in phases, every year that OpenAI pays its rent on time reduces NVIDIA’s remaining guarantee obligations accordingly.
There is no such thing as a free guarantee, and NVIDIA clearly spelled out its terms of exchange in the same filing. Save for a few exceptions, the initial capacity of this site must exclusively run NVIDIA’s full-stack compute equipment. Meanwhile, NVIDIA also invested $1.5 billion for an equity stake in SB Energy, becoming a partner alongside SoftBank and OpenAI.
Examining this network of relationships carefully reveals that NVIDIA is simultaneously playing four roles: it is the chip supplier selling the hardware, an equity shareholder in the landlord, the guarantor backing the tenant, and ultimately the entity receiving the tenant’s money back through chip purchases. Meanwhile, OpenAI is both tenant and equity shareholder in the landlord, and SoftBank is both a major shareholder in OpenAI and the parent company of the landlord.
There is also an easily overlooked recourse clause in the document: any money NVIDIA pays to the landlord must be reimbursed in full by OpenAI after the fact. On paper, NVIDIA loses nothing, but that is precisely where the awkwardness lies. The premise for NVIDIA stepping in to front cash in the first place is that OpenAI cannot afford its rent; by the time OpenAI is expected to pay NVIDIA back, it will be at the exact moment when it has run out of money. How much value that recourse claim will actually yield during liquidation is anyone’s guess.
Jensen Huang published a byline blog post on the same day, explaining this operational mechanism quite candidly. He noted that the expansion pace of frontier model labs is “growing faster than their balance sheets and long-term credit profiles can support.” He continued that just as NVIDIA leveraged its massive scale and demand visibility in the past to secure critical chip resources, “we are now applying that same discipline to secure LPS capacity”—referring to land, power, and shell space.
In financial terms, NVIDIA is taking its visibility into future chip orders and translating it directly into credit, lending it to its customer for financing. Between existing and planned commitments through 2030, OpenAI has outlined roughly 12 GW—expandable to around 16 GW—of procurement, corresponding to $600 billion in total compute orders from NVIDIA. This long-term commitment does not require immediate cash layout today, but within the guarantee structure, it serves as tangible collateral when SB Energy seeks bank financing.
The circulation of commitments and funds thus forms a closed loop: NVIDIA provides the guarantee and takes an equity stake in SB Energy; SB Energy uses this credit enhancement to borrow bank funds to construct the facilities; OpenAI leases the facilities and purchases NVIDIA chips; OpenAI pays rent to SB Energy and chip procurement costs to NVIDIA. A substantial portion of the money OpenAI uses to pay rent and buy hardware comes from NVIDIA’s prior $30 billion equity investment in OpenAI, alongside capital injections from SoftBank.
Consider this set of financial metrics first: the operation currently retains a comfortable safety cushion. OpenAI’s annualized revenue just surpassed $40 billion, doubling year-over-year. NVIDIA’s own cash flow is even more extraordinary, generating $48.6 billion in free cash flow in a single quarter, backed by $50.3 billion in cash and marketable securities on its balance sheet—leaving it with formidable financial cushion.
Beyond this specific transaction, similar plays have spread across the industry. NVIDIA’s total equity investments across customers and its broader ecosystem stand at roughly $80 billion; adding up various guarantees and buyback commitments, the nominal footprint reaches $193 billion. Rival AMD is taking a similar playbook, issuing warrants for up to 10% equity stakes to OpenAI and Meta in exchange for massive chip procurement commitments. Hardware vendors lending their own credit to lock in long-term orders is becoming standard operating procedure across the industry.
The negotiation process behind this deal was equally telling. When The Wall Street Journal first broke the story in late July suggesting the guarantee could reach $250 billion, NVIDIA’s stock dropped roughly 5% in a single day. By mid-August, media reports indicated the size had been trimmed below $120 billion, before finally settling at $105 billion in the official release, covering only the initial 4.25 GW phase. Capital markets clearly have a psychological threshold for how much credit risk a single company should assume, and management promptly scaled back the exposure.
Seeing this guarantee model, many people’s knee-jerk reaction is to recall the 2008 subprime mortgage crisis. That intuition correctly captures the interconnected nature of the risk, but the specific mechanisms of the two scenarios differ significantly.
There is really only one similarity, but it carries immense weight: the exact moment the guarantor is forced to pay out happens to be when its own core business is hit hardest and cash flow is tightest. When would NVIDIA actually have to cough up that $105 billion? The day overall AI demand collapses and OpenAI suffers severe losses and can no longer pay rent. If AI demand falls off a cliff, NVIDIA’s own chip shipments will drop precipitously, and its $30 billion equity stake in OpenAI will suffer massive write-downs. It is like an insurance company whose sole client is itself.
That was precisely how AIG collapsed in 2008: its insurance payout triggers deteriorated in lockstep with its own solvency. In this deal, the asset concentration is even more extreme than AIG’s back then. AIG’s underlying guaranteed assets were at least spread across pools of thousands of individual residential mortgages, whereas NVIDIA is guaranteeing a data center tied to a single tenant, a single use case, and deeply bound to its own hardware.
However, the structural firewalls between the two are fundamentally different. AIG’s fatal flaw back then was that any downgrade in external credit ratings would instantly trigger collateral calls (margin calls), creating an immediate liquidity squeeze. NVIDIA’s guarantee agreement contains no collateral call mechanism whatsoever; $105 billion is a hard cap on cumulative liability, paying out only the net deficit following asset disposition, and all core terms were fully disclosed on the day of signing.
In terms of specific structure, this transaction differs from the subprime crisis in three distinct ways. First, the channel of risk contagion is different. The destructive force of the subprime crisis stemmed from bundling mortgages into layers of securities sold across the globe, where after multiple re-packaging steps even professional buyers could not see the underlying assets. In contrast, AI infrastructure debt currently resides primarily in private credit markets and standalone project entities, held to maturity by private capital rather than traded frequently in public markets; the total securitized data center debt across the entire industry stands at only around $61 billion.
Second, the guarantor’s own absorption capacity is vastly different. NVIDIA holds over $50 billion in cash and marketable securities, generates $48.6 billion in free cash flow in a single quarter, and maintains a low debt-to-asset ratio. Even if payouts are triggered down the road, they will be paid in installments covering net deficits post-disposition, meaning a single shock will not break the company’s operational liquidity.
Finally, there is a stark difference in transparency. The liability cap, trigger conditions, and payout formulas of this deal were published directly in public SEC filings on day one. On the eve of the 2000 telecom bubble, Lucent Technologies hid $8.1 billion in customer financing deep inside financial footnotes; Enron went even further, hiding off-balance-sheet residual value guarantees inside unconsolidated affiliate entities.
Putting these facts together, this looks much more like an equipment vendor offering credit enhancement for project financing to lock in mega orders. The capital originates primarily from private credit, with NVIDIA acting as a centralized credit hub for the industry. It is not currently at the stage of triggering a systemic crisis; the real cause for concern is whether this backing model will expand unchecked in the future.
If the AI industry eventually faces a cyclical downturn, tracing down the risk chain reveals that the entity in the most perilous position is actually not NVIDIA with its deep pockets. Risk-bearing capacity sorts into a clear hierarchy from weakest to strongest.
The first tier consists of creditors behind compute leasing platforms like CoreWeave, including numerous pension funds and private credit investors seeking steady yields. Their business model is pledging purchased GPUs as collateral to borrow money, using those funds to buy more GPUs, and renting them out to frontier model clients. Debt in the CoreWeave ecosystem totals around $249 billion. While the company’s corporate credit rating is speculative-grade, its project loans secured Moody’s A3 investment-grade ratings on the back of leases with major clients. However, GPUs depreciate rapidly—making this operation akin to borrowing long-term against used cars. If second-hand hardware prices plummet, these entities could plunge directly into insolvency, while the lack of public trading in private credit will keep paper losses invisible until investors ultimately foot the bill.
The second tier is OpenAI itself. Its run-rate annual revenue has indeed surged to $40 billion, but its burn rate is even more staggering—projections suggest its compute spending alone could reach $121 billion in 2028, with losses that year climbing to $74 billion. Its capability to sustain huge rent payments over the long haul rests entirely on an endless stream of primary market funding and a forthcoming IPO. S&P previously downgraded Oracle’s credit rating to BBB- over OpenAI’s payment commitments, primarily because OpenAI accounts for over half of Oracle’s $600 billion-plus backlogged orders. Rating agencies are already pricing in this customer concentration.
The third tier is SoftBank Group. In March 2026, SoftBank took out a $40 billion one-year bridge loan (at roughly 6.14% interest) specifically to increase its stake in OpenAI, maturing all at once in March 2027. Equity in frontier model labs typically comes with strict lockup periods and cannot be liquidated on demand. Borrowing short-term one-year debt to purchase illiquid long-term equity represents a textbook maturity mismatch, betting heavily that capital market refinancing windows remain wide open.
At the bottom in the fourth tier sits NVIDIA. When an industry downturn strikes, it will indeed take hits on multiple fronts: declining core revenue due to slow chip sales, massive write-downs on its $30 billion equity investment in OpenAI, shrinking valuations across portfolio ecosystem companies, and covering data center shortfalls in installments under the $105 billion cap. Yet because NVIDIA holds abundant net cash, operates without collateral call provisions, and settles guarantee claims in batches based on post-disposition deficits, its multi-year cash flow remains sturdy enough to weather one or two cyclical shocks.
To track how this credit architecture evolves, there are several clear observation windows on the calendar. The nearest is in late August, when NVIDIA’s quarterly report will attach the full text of the guarantee agreement—with key details to watch in the residual value formula and off-balance-sheet accounting treatment. SB Energy is expected to launch its IPO in September, and its prospectus will reveal whether its project debt can be sold independently without NVIDIA’s guarantee. If its debt finds no buyers without NVIDIA’s backing, the thesis of NVIDIA acting as the central credit hub for the industry will be firmly validated.
The medium-term window requires close tracking of OpenAI’s IPO timeline and the interest rate and conditions of SoftBank’s $40 billion bridge loan refinancing in March 2027. Beyond these, two slow variables demand long-term monitoring: the gap between second-hand GPU market transaction prices and book depreciation assumptions, and the pipeline schedule of compute-backed debt issuance. On the eve of the telecom bubble collapse, it was precisely the sudden overnight freeze in the bond market that ignited a full-blown crisis.
In our previous article, we discussed four execution details for engineering implementation: whether final regulatory approval for power expansion has been secured, whether key hardware delivery lead times are locked in as scheduled, whether environmental impact review approvals are fully complete, and who bears the extra costs of schedule delays. Now that hardware vendors are stepping onto the field themselves to act as guarantors, we need to add three business and financial dimensions on top of those.
First, does the guarantor’s own primary operating revenue originate from the exact same source of demand as the underlying data centers covered by the guarantee? Second, among the parties providing fresh capital, who ultimately bears the risk—commercial banks, long-term private credit investors, or retail investors and stock buyers in secondary markets? Third, is the gap between actual secondary market transaction prices of compute hardware used as collateral and the projected residual values on financial balance sheets widening or narrowing?
Examining the flood of upcoming gigawatt-scale data center partnerships through the lens of these seven specific questions will prove far more constructive than dwelling on short-term market sentiment. Understanding where the real risk falls matters immensely more than rushing to a grand conclusion.