What Nvidia's $105 Billion Guarantee for OpenAI's Ohio Data Center Means — August 19, 2026
· CompaniesAutomation
Radar Flash Edition: Nvidia guarantees up to $105 billion for OpenAI's Ohio campus, $145 billion less than reported in July, with the first 800 megawatts operational in 2028. What this means for your company's AI capacity and costs.
Flash Edition. On Saturday Nvidia was negotiating the guarantee for the OpenAI data center in Ohio and the figure was shrinking. It is now signed: a $105 billion guarantee, $145 billion less than reported in July, for capacity that won't exist until 2028. A supplement to the latest Radar.
What happened
Nvidia will guarantee up to $105 billion so that OpenAI can lease for twenty years the campus that SB Energy (SoftBank) will build in Pike County, Ohio. It is also investing $1.5 billion in SB Energy and will be the exclusive chip provider: 4.25 gigawatts expandable to eight, with the first 800 megawatts operational in 2028 (Source). What's new compared to what we reported on Saturday: the guarantee is no longer a $250 billion rumor, but a closed commitment 58% lower, covering rent, energy, and part of the asset's value if OpenAI were to default (Source). Huang denies this is circular financing: he estimates $200 billion in revenue for Nvidia from that site (Source).
Why it matters
The figure that affects you is not 105 billion: it is 2028. Not a single watt from this campus will serve your workload in two years: the capacity you will operate with until then is already built or under construction. Your request limits and your price per token will not improve because of Ohio, and any plan that assumes surplus and cheaper compute in a year is betting against the industry's timeline. The cut from 250 billion to 105 billion says the rest: it's not the ambition that's tight, but the risk the chipmaker is willing to assume. This is not investment advice: it's your supply chain hedging its bets.
For your company
Three things, none about price. One: ask in writing about capacity—requests per minute, priority level, what happens on a peak day, and how much notice they can give to change those limits; in 2027 the bottleneck will be availability, not the dollar per million tokens. Two: no new capacity is arriving before 2028, so savings must come from your side: cache repetitive requests, batch non-real-time tasks, and reserve the powerful model for where it changes the outcome. Three: test for degradation—what does your automation do in the face of a 429 error or triple the latency? If it stops, that's this week's task: queues, staggered retries, and a small backup model. If you prefer that resilience set up and measured, an AI consultancy can have it resolved in weeks.
Frequently Asked Questions
Is Nvidia giving $105 billion to OpenAI?
No: it's a guarantee, not cash. It enables lenders to finance the construction knowing someone is responsible for rent, energy, and part of the asset value. OpenAI doesn't receive that money; it receives a lease it can sign because someone is backing it.
Will this make the AI I use cheaper?
Not before 2028, and even then: more capacity coexists with more demand from agents, which consume in loops. The price drops over the next two years will come from more efficient models, not from this concrete.
My company isn't an OpenAI customer, does this affect me?
Yes, at the foundational level: Anthropic, Google, or the open model you run in the cloud share the same chain of chips, energy, and data centers. When things are tight, they are tight for everyone at once.