The business of selling a GPU typically nears its end at shipment: the customer places an order, the supplier delivers, and revenue is recorded on the financial statements. But the latest reports surrounding OpenAI’s Ohio data center describe a different relationship. According to The Wall Street Journal and Bloomberg, NVIDIA is negotiating to provide guarantees for the financing of OpenAI’s compute leasing arrangements. Even after the chips are sold, the supplier may still be vouching for the customer’s ability to make long-term rental payments.
The report mentions a guarantee amount of approximately $250 billion and projects totaling 10 GW, but these figures have not been independently verified by NVIDIA, OpenAI, or the project’s financing parties through official disclosures, and therefore cannot be considered confirmed commitments. Yet, the most intriguing aspect of this rumor isn’t whether the numbers are large enough—guarantees suggest that chipmakers may now be sitting at the table alongside clients raising funds to build mining facilities, rather than merely selling shovels.
Why did this happen to NVIDIA? First, let’s look at where it already stands.
According to NVIDIA’s FY2025 annual report, data center revenue rose from $6.7 billion to $115.2 billion over five fiscal years. The slope on the chart is more interesting than the conclusion itself. FY2023 was still a relatively flat plateau, but the following two years suddenly became nearly vertical. For a company that originally sold computing hardware, this implies that its most important customer base switched its purchasing behavior in an extremely short time.
In the past, cloud providers updated servers by generation and could allocate budgets on a quarterly basis. Now, model companies require vast, continuous computing power that can run for years. Building a data center is no longer just about buying a few more racks of servers—it’s more like constructing a power plant first, then deciding what applications to run inside. Equipment procurement is only the first step; land, electricity, facilities, and debt financing must all be accounted for within the same project.
NVIDIA doesn’t need to build data centers itself to be pulled closer by this chain. If customers can’t secure financing, GPU orders remain nothing more than letters of intent. Only when customers obtain financing does the certainty of system deliveries and service revenues over the coming years emerge. Guarantees here are not charity—they are credit instruments that transform demand from “wanting to buy” into “able to buy.” But once these instruments are used, risk flows back along the same chain.
How is this different from previous large-scale AI procurement cycles? The difference lies in the change of the project’s narrative unit.
When OpenAI announced its "Stargate" project in 2025, it stated a potential investment of up to $500 billion over the next four years, with a capacity target of 10 GW. Meanwhile, according to a public announcement by GPU cloud provider CoreWeave, its cloud services contract with OpenAI could reach up to $11.9 billion. These figures cannot be added together—they represent different things: a planned investment, a service contract ceiling, and a power capacity target. Presenting them on the same chart is not about calculating a total, but about recognizing that transactional language has moved beyond single-server or single-year procurement.
It’s like building a railway. While train car orders are important, what ultimately determines whether the railway can operate is who first funds the tracks, who commits to consistently buying tickets, and who covers the shortfall when passenger demand is low. For AI data centers, GPUs are the most visible cars—but it’s power and financing that determine whether the train can even leave the station.
Therefore, when the word “guarantee” appears in the news, the market should not interpret it merely as NVIDIA securing another sale. It is more like the supplier signaling to the capital provider that they are willing to embed their assessment of downstream demand into a credit relationship. For model companies urgently expanding their computing power, this lowers the barrier to financing. For those leasing equipment or extending loans, it introduces an additional party willing to share the risk.
Where will the risks land? Many assume that as long as the model company continues to grow, the entire chain will remain intact. However, what large infrastructure projects fear most is not a slight drop in service sales in a single month, but rather when long-term leases, depreciation cycles, and debt maturities are already in motion, yet demand fails to materialize at the originally projected pace.
This diagram breaks down a computing power transaction into four roles: the model company needs computing power, data centers or cloud providers purchase equipment and deliver services, creditors or lessors provide financing, and chip suppliers deliver systems. In traditional supply relationships, the supplier’s primary risks are focused on delivery and payment collection. If credit support is incorporated into the contract, the relationship extends to lease performance, equipment residual value, and even project refinancing.
This is not an abstract finance class. CoreWeave’s S-1 filing reveals that Microsoft contributed 62% of its 2024 revenue. When customers are concentrated, a cloud provider’s ability to raise capital becomes tied to the payment reliability of a few major clients. For upstream suppliers, what they desire most is stable, long-term demand. But when a major customer provides credit backing for this stability, some of the uncertainty originally absorbed by downstream parties is pushed back up the supply chain.
So, what this rumor truly changes is not who will buy tens of thousands more GPUs, but that AI infrastructure is rewriting “orders” into long-term contracts spanning devices, leasing, and credit. Chip deliveries may end, but credit relationships may not conclude at that moment.
律动 BlockBeats