Over the past year (from autumn 2025 to October 2026), NVIDIA's relationship with its customers has systematically evolved from simple chip sales into a complex supplier financing system. As of October 2026, NVIDIA's total off-balance-sheet commitments and guarantees to the AI ecosystem amounted to approximately USD 530.5 billion, of which contracts and guarantees truly directed toward customer support totaled about USD 164.5 billion, accounting for roughly 72% of its shareholders' equity. During the same period, its unlisted equity investments expanded from approximately USD 3 billion a year ago to about USD 99 billion (on an actual holding basis).
This means that NVIDIA has substantively transformed from a chip supplier into a capital organizer and credit underwriter for its customers. Breaking down its various investment arrangements clearly reveals the investment and financing rules of top players in the AI ecosystem.
I. Equity Investment (Cash for Equity)
Equity investment binds computing power entry points with cash: USD 30 billion for OpenAI, a cap of USD 10 billion for Anthropic, over USD 10 billion for xAI, and approximately USD 4.7 billion for CoreWeave; the maximum loss is limited to the principal.
OpenAI
Finalized amount: USD 30 billion. In February 2026, NVIDIA participated in a funding round for OpenAI totaling approximately USD 110 billion, with NVIDIA contributing USD 30 billion (SoftBank USD 30 billion, Amazon USD 50 billion). This round valued OpenAI at approximately USD 730 billion to USD 840 billion. Previously, in September 2025, a plan to invest "up to USD 100 billion" was announced, which was later significantly scaled down to the USD 30 billion level.
Anthropic
Committed cap: USD 10 billion, reached in November 2025. This investment is tied to a commercial agreement: Anthropic committed to purchasing USD 30 billion worth of Microsoft Azure computing power driven by NVIDIA chips. In September 2026, NVIDIA further considered acting as a cornerstone investor in Anthropic's IPO, with the amount also pointing to a maximum level of USD 10 billion.
xAI
January 2026: Invested over USD 10 billion, participating in a funding round of approximately USD 20 billion for xAI. Later, xAI merged into SpaceX and went public, with NVIDIA's equity converted into approximately 123 million shares of SpaceX.
CoreWeave ("New Cloud" Customer and Investee)
NVIDIA is one of the largest single beneficiaries of CoreWeave. As of mid-2026, it held approximately 47.2 million shares of CoreWeave, with a holding market value of about USD 4.7 billion. In January 2026, it made an additional investment of USD 2 billion at USD 87.20 per share. NVIDIA's funding for CoreWeave is essentially supplier financing: the funds are used by CoreWeave to purchase NVIDIA GPUs, which are then repeatedly leveraged as collateral for borrowing to expand.
SB Energy (Data Center Project Developer)
On August 17, 2026, NVIDIA announced an equity investment of USD 1.5 billion in SB Energy as part of the consideration for the guarantee arrangement for OpenAI's Ohio data center, in exchange for "exclusive hosting of NVIDIA AI computing" at the campus.
Other Invested Customers/Ecosystem Companies
Through NVentures and direct balance sheet investments, NVIDIA has covered multiple purchasers of its chips or computing power operators. Representative targets include: Nebius, Lambda, Together AI ("New Cloud" customers), as well as emerging computing power operators such as Firmus Technologies and IREN. Among them, Lambda is both an investee of NVIDIA and a supplier to NVIDIA, as NVIDIA leases GPU servers from it, forming a two-way relationship. Xiangzhi Perspective: Equity investment is NVIDIA using capital to secure entry points for computing power demand and ecosystem control. The portfolio of OpenAI, Anthropic, xAI, CoreWeave, SB Energy, and NVentures is essentially binding future procurement priorities. Although the risk cap is the investment principal, these investees are simultaneously customers or computing power carriers, and their revenues exhibit obvious closed-loop characteristics of investing in their own ecosystem. When valuing, investment-driven demand should be excluded, and attention should be paid to whether the entry valuation is overextended.
II. Credit Guarantees and Residual Value Guarantees
Guarantees are contingent obligations rather than investments; the residual value guarantee cap for SB Energy/OpenAI Ohio is USD 105 billion, and loan guarantees for other customers are about USD 3.5 billion.
OpenAI / SB Energy Ohio Data Center: Cap of USD 105 Billion
This is the largest single credit support arrangement in the past year. An 8-K filing disclosed on August 17, 2026, revealed that NVIDIA and SB Energy signed a residual value guarantee for the PORTS campus in Piketon, Ohio, providing credit support for a 20-year lease with an approximately 4.25 GW IT load for OpenAI affiliates, with a cumulative payment obligation cap of USD 105 billion, plus an additional option for about 3.8 GW. NVIDIA is neither the lessor nor the borrower, but a third-party guarantor: if OpenAI fails to pay rent, NVIDIA assumes the payment obligation up to the guarantee cap. This obligation is currently not on the balance sheet and belongs to off-balance-sheet contingent obligations. Previously, the market once rumored that NVIDIA might provide up to USD 250 billion in support to OpenAI, which ultimately settled on a structure of a USD 105 billion guarantee plus a USD 30 billion equity investment.
Loan Guarantees for Other Customers: Approximately USD 3.5 Billion (Earlier Time Point)
According to a mid-2026 report by The Wall Street Journal, NVIDIA has provided guarantees for loans of certain customers, with the most recent tally totaling USD 3.5 billion. This caliber may not cover the aforementioned USD 105 billion SB Energy residual value guarantee, belonging to an earlier or narrower statistical scope.
Total Off-Balance-Sheet Guarantee Scope
Combining 10-Q disclosures, NVIDIA's off-balance-sheet guarantee commitments are approximately USD 108.5 billion, which, together with AI cloud contracts and third-party data center leases, constitutes a total customer support exposure of about USD 164.5 billion. Xiangzhi Perspective: Guarantees are not investments, but rather NVIDIA underwriting customer credit. The maximum USD 105 billion residual value guarantee for the OpenAI/SB Energy Ohio project transforms NVIDIA from a chip supplier into the backstop for data center rent and residual value. The risk lies not in the equity investment principal, but in triggering actual compensation when OpenAI's cash flow deteriorates; it should be included in quasi-debt and contingent liability analysis.
III. Procurement Commitments, Lease Backstops, and "Take-or-Pay" Arrangements
Procurement and lease backstops serve as revenue backstops: USD 6.3 billion for CoreWeave until 2032, and USD 1.3 to 1.5 billion for Lambda; if unleased, they will convert into cash expenditures or asset occupation.
CoreWeave Computing Power Procurement Commitment: USD 6.3 Billion
On September 9, 2025, NVIDIA signed a procurement order with CoreWeave with an initial contract value of USD 6.3 billion, committing to purchase CoreWeave's unsold cloud computing capacity, with the obligation lasting until April 13, 2032. If CoreWeave's GPUs cannot be leased out, NVIDIA will take them over according to the contract.
"AI Computing Power Partner Program" Backstop Commitment
In July 2025, NVIDIA launched a program targeting small and medium-sized cloud service providers: if customers cannot find end-users to lease GPUs, NVIDIA commits to leasing them itself as a backstop, collecting a 50% share of revenues exceeding the threshold. The first batch of customers includes Sharon AI and Firmus Technologies. The essence of this program is to allow small and medium-sized cloud service providers to use NVIDIA's "guaranteed revenue" contracts to secure bank financing, thereby expanding GPU procurement. Less than two months after its launch, the program suspended some transactions due to customer resistance and control disputes, but NVIDIA stated that "the new business model remains valid."
Commitments to Third-Party Data Centers/Cloud Capacity
Off-balance-sheet commitments include USD 20 billion in uncommenced third-party data center leases and USD 36 billion in AI cloud contracts, which serve the continuity of customer computing power deployment to varying degrees.
Lambda GPU Leasing: USD 1.3 to 1.5 Billion
In the summer of 2025, NVIDIA signed a GPU server leasing agreement with Lambda, codenamed "Project Comet," to lease 10,000 GPU servers equipped with NVIDIA chips over 4 years, valued at approximately USD 1.3 billion; other reports indicated the total agreement scale reached USD 1.5 billion. NVIDIA transformed from an investor in Lambda to one of its largest customers, essentially providing a revenue backstop for Lambda's computing power business. Xiangzhi Perspective: The essence of such arrangements is "revenue backstop," which can lock in GPU demand and customer financing capabilities in the short term. However, once computing power cannot be leased out, the contracts will translate into real cash expenditures or asset occupation for NVIDIA. The USD 6.3 billion procurement commitment for CoreWeave, Lambda leasing, and the AI Computing Power Partner Program should all be regarded as off-balance-sheet operating obligations, rather than pure strategic synergy.
IV. Assisting Customers with Debt Financing and Asset Packaging
GPU-collateralized debt platforms, procurement credit enhancement, and extended payment terms reduce customer financing costs, indirectly amplifying NVIDIA's chip demand and credit exposure.
Debt Platform with GPU as Underlying Assets
NVIDIA collaborated with six institutions including BlackRock, Apollo, and Goldman Sachs to build a hardware financing system with a scale exceeding USD 500 billion. Using NVIDIA AI hardware as underlying assets, it issues public and private debt to pension funds, insurance companies, sovereign wealth funds, etc., raising funds for chip procurement or leasing financing for AI enterprises.
Implicit Endorsement of GPU-Collateralized Debt
"New Cloud" customers (CoreWeave, Lambda, Fluidstack, etc.) use GPUs as collateral for borrowing. NVIDIA's procurement commitments and lease backstops make the future cash flows of these GPUs more predictable, thereby indirectly reducing customers' debt financing costs. As of June 2026, CoreWeave's consolidated debt and lease obligations were approximately USD 52 billion, with financing costs approaching 10% to 15%, while NVIDIA's own borrowing cost is only about 4.6%. This credit spread is precisely the risk portion "absorbed" by NVIDIA's credit output.
Extended Payment Arrangements for Investment-Grade Customers
NVIDIA acknowledged in its financial report that "financing arrangements with certain investment-grade customers, including extended payment terms under large multi-quarter agreements, will continue to affect the timing of operating cash flows." This means that some customers are purchasing chips with longer payment terms, and NVIDIA is substantively providing trade credit. In Q2 of fiscal year 2025, accounts receivable expanded from USD 38.5 billion to USD 63.1 billion, an increase of approximately USD 24.6 billion in five months, which is directly related to such arrangements.
Xiangzhi Perspective: Through the GPU-collateralized debt platform, procurement commitment credit enhancement, and extended payment terms, NVIDIA has effectively reduced customers' financing costs and tied customers' capital expenditure capabilities to its own credit. The spread between CoreWeave's approximately USD 52 billion debt/lease obligations and NVIDIA's approximately 4.6% financing cost indicates that NVIDIA is absorbing customer credit risk, and the quality of chip revenue is partially diluted.
Risk Nature Analysis
Exposures cannot be simply aggregated: equity loses principal, guarantees may trigger compensation, procurement commitments lead to cash expenditures, and debt assistance indirectly amplifies; pay attention to whether the off-balance-sheet growth rate is faster than profit.
The amounts in the above list are not investments of the same nature. Among them, equity investments (USD 30 billion for OpenAI, USD 10 billion for Anthropic, USD 10 billion for xAI, approximately USD 4.7 billion holding for CoreWeave, etc.) are real cash considerations paid or committed, with the maximum loss limited to the investment principal;
Credit guarantees and residual value guarantees (USD 105 billion cap for SB Energy/OpenAI, USD 3.5 billion for other customers) are contingent obligations: they do not use cash currently, but may form actual compensation when triggered, upgrading the risk nature from "equity investment loss" to "customer solvency risk";
Procurement commitments and backstop arrangements (USD 6.3 billion for CoreWeave, AI Computing Power Partner Program, Lambda leasing) are off-balance-sheet contractual obligations, which will translate into real cash expenditures or asset occupation upon fulfillment; debt financing assistance (GPU debt platform, extended payment terms) does not directly appear in NVIDIA's "investment" accounts, but indirectly amplifies NVIDIA's chip demand and credit exposure through customers' balance sheets and cash flows.
S&P Global has already included some guarantees in NVIDIA's debt adjustment items in its adjustment model, expecting its debt adjustment amount to climb from approximately USD 4.2 billion in 2028 to about USD 37.7 billion in 2031-2032. This trend indicates that off-balance-sheet credit exposure is expanding at a rate faster than profit growth, and its risk nature is fundamentally different from traditional equity investment.
Xiangzhi believes that, from the perspective of risk nature, these exposures cannot be simply aggregated by investment amount, but should be layered by risk nature: the maximum loss for equity investment is the principal, guarantees and take-or-pay will turn into compensation or cash expenditures when triggered, and debt assistance indirectly amplifies demand and risk through customers' balance sheets. A core judgment is that if the growth rate of off-balance-sheet credit exposure continues to be faster than profit, NVIDIA's risk attribute will quantitatively change from a growth manufacturer to an AI infrastructure credit underwriter.