Elena VossOpenAI’s 10 GW Ohio data-center lease may come with a reported $250 billion backstop from Nvidia. If the parties agree, Nvidia would support the lease behind future GPU demand. The chip supplier would sit inside the project’s capital stack.
Key takeaways
- Nvidia is reportedly discussing a roughly $250 billion contingent backstop for OpenAI’s lease of a 10 GW Ohio data-center project, potentially placing the chipmaker inside the project’s financing structure rather than merely supplying GPUs.
- Nvidia is financing multiple links in the deployment chain: a planned $1 billion Naver data-center investment, a $1.5 billion Amkor packaging agreement with a prepayment, memory coordination with SK Hynix, and rent-back promises for idle cloud GPUs.
- Rent-back agreements with providers such as Firmus and Sharon AI create a utilization floor, shifting some risk from operators and lenders to Nvidia and making GPU-backed infrastructure easier to finance.
- Vendor support makes GPU sales a less reliable measure of independent customer demand because purchases may depend on Nvidia’s capital, supply commitments, or promised utilization.
- AMD’s agreement to invest up to $5 billion in Anthropic alongside potential purchases of up to 2 GW of MI450 systems shows that balance-sheet support is becoming part of accelerator competition.
Nvidia is moving from purchase orders to project finance
SoftBank’s energy subsidiary is building the Ohio project, OpenAI would lease the capacity, and Nvidia is discussing support large enough to affect how lenders finance that lease.
The talks remain unsettled. Reported terms do not establish an immediate $250 billion cash investment, and Nvidia’s exposure would depend on triggers and conditions that have not been reported. Even so, Nvidia is negotiating a project-finance role alongside its hardware sales.
Nvidia separately plans to invest $1 billion in Naver to help finance a South Korean AI data center.
Nvidia’s two commitments create different exposures: the Ohio proposal is a contingent backstop, while the Naver funding is a planned investment. Both would help developers secure financing before the first GPU arrives.
Nvidia is funding the bottlenecks around its GPUs
Customers can deploy Nvidia GPUs only after they secure packaging, memory, networking, power, cooling, permits, cloud capacity, credit, and sustained utilization. If power or packaging arrives late, completed accelerators can sit idle.
Nvidia announced a $1.5 billion agreement with Amkor that includes a prepayment supporting an Arizona packaging facility. Nvidia and SK Hynix also said they would secure next-generation memory supply and jointly develop high-bandwidth memory. Nvidia is putting capital and coordination behind the inputs its customers need.
By July 2026, hyperscalers had publicly announced 46 GW of planned AI data-center capacity. That total counts proposed capacity, not energized sites.
Lenders cannot finance power that utilities have not delivered. They must judge whether contractors can complete and energize a site and whether tenants will keep it busy long enough to repay the debt.
Nvidia is turning utilization into collateral
A smaller financing structure reveals the mechanism. Nvidia has promised to backstop young cloud providers such as Firmus and Sharon AI by renting unused GPUs in exchange for a share of revenue. By promising to use idle systems, Nvidia addresses the utilization risk facing operators and lenders.
Firmus and Sharon AI must finance servers before all future workloads are known. Nvidia’s rent-back inserts a demand floor between installation and customer revenue.
Nvidia’s promised rentals create vendor-supported offtake and shift some utilization risk from the cloud provider to the company whose systems fill the building.
As the contracted megawatt shows, lenders finance powered compute when a sufficiently creditworthy party commits to use it. Nvidia increasingly supplies that commitment.
AMD is pairing capital with chips
AMD and Anthropic agreed that Anthropic may buy up to 2 GW of MI450 systems beginning in the first half of 2027. AMD also committed to invest up to $5 billion in Anthropic. AMD is competing on customer capital as well as accelerator performance.
Nvidia would support sites in Ohio and South Korea; AMD would invest directly in Anthropic. Both chipmakers are putting their balance sheets beside future compute purchases.
Supplier support changes what demand measures
A GPU sale can reflect both customer demand and supplier support. When Nvidia invests in a customer, prepays a bottleneck, finances a site, or rents unused equipment, its revenue becomes a noisier measure of demand that customers can sustain without vendor help.
Buyers and lenders should separate end-customer commitments from Nvidia-supported utilization, then ask how much capital or offtake Nvidia must supply if demand falls short.
Bain estimated that AI companies would need $2 trillion in combined annual revenue by 2030 to fund projected compute demand and could fall short by $800 billion. That estimate puts a revenue requirement beneath today’s infrastructure commitments.
By taking utilization exposure, Nvidia can accelerate construction. Nvidia still depends on utilities delivering power, customers paying their leases, and deployed AI producing enough revenue to support the financing.
Even if the Ohio talks produce no transaction, Nvidia has already invested in sites, prepaid packaging capacity, coordinated memory supply, and promised to rent idle GPUs. The reported Ohio proposal shows how large that role could become.
The reported $250 billion first reads as extravagance. Against Nvidia’s smaller commitments, it is better read as the possible scale of support behind 10 GW of bankable GPU demand. Nvidia is pricing the accelerator and the distance between a chip order and a working AI factory.
Frequently asked questions
Is Nvidia committing $250 billion in cash to OpenAI’s Ohio data center?
Not necessarily. The reported proposal is an unsettled, contingent backstop for OpenAI’s lease, and Nvidia’s actual exposure would depend on triggers and conditions that have not been disclosed.
Why would Nvidia guarantee or support demand for its own GPUs?
The support can make large AI projects more financeable by reducing lenders’ concerns about tenant credit and future utilization. That can accelerate construction and create deployable demand for Nvidia systems.
How do Nvidia’s GPU rent-back agreements work?
Nvidia has promised to rent unused GPUs from young cloud providers such as Firmus and Sharon AI in exchange for a share of revenue. The commitment supplies a demand floor while those operators build their customer bases.
What infrastructure bottlenecks is Nvidia funding besides data centers?
Nvidia has supported advanced packaging through its Amkor agreement and is coordinating next-generation memory supply with SK Hynix. These efforts address constraints that can delay deployment even after GPUs have been ordered.
What risk does Nvidia take by underwriting AI infrastructure demand?
Nvidia becomes more exposed to weak utilization, missed lease payments, construction delays, and unavailable power. Its support works only if sites are energized and deployed AI services eventually generate enough revenue to sustain the financing.