Nvidia guarantees up to $105 billion of OpenAI's lease

Nvidia agreed to guarantee up to $105 billion of OpenAI's lease on an Ohio data centre being developed by SoftBank-owned SB Energy, one of the chipmaker's largest infrastructure commitments and a further case of it financing demand for its own chips, The Economic Times reported. It is the largest number in today's record, and it is not a purchase, a round or a contract. It is a promise to pay someone else's rent.

The names on the guarantee are already on the cap table

OpenAI closed $122B in committed capital at an $852B post-money valuation in March 2026, anchored by Amazon, NVIDIA and SoftBank, per OpenAI. The chipmaker now guaranteeing the lease and the owner of the company developing the building are both named in that round. Annualised revenue on the ledger's record at the same date is $25B, per The Information, and the company filed confidentially for an IPO in June 2026, per CNBC.

A guarantee is a contingent liability, not cash: Nvidia pays only if the tenant does not. What it records is how few balance sheets appear at each layer of a single building. The chips inside it, the capital behind the tenant, the owner of the developer, and now the rent, all trace back to the same short list of names.

The buildout leaves a physical trace

Mexico's computer equipment exports rose 172% in the first half of 2026 to $82.9 billion, passing its automotive and auto-parts exports as AI demand grew, per Mexico Business News. A country whose export identity was cars now ships more computing hardware than vehicles. That figure is a half-year total, not a single month, and it is the closest thing on today's record to a measurement of the buildout rather than a commitment to fund one.

One tier down, the same demand is being sold as product:

• DigitalOcean shares rose 5% after its Cloudways service made Managed AI Agents generally available, a product line starting with the open-source agents OpenClaw and Hermes, per 24/7 Wall St.
• Cisco and Quali launched a stack automation platform for enterprise AI deployment, with full-stack Cisco AI PODs due to be generally available in October 2026, per Technology.org.

And on the same day, a model built for a laptop

Alibaba released Qwen3.8-27B, a 27-billion-parameter open-weight model built to run on laptops and other consumer hardware, and published the weights of its most powerful model, Qwen3.8 Max, alongside it, per CNBC.

These are not the same workload. Training a frontier model and running a 27B one on a laptop sit at opposite ends of the compute question, and nothing in today's record says the second reduces demand for the first. But they landed on the same day, and they point in opposite directions: one commitment concentrates a hundred billion dollars of obligation into a single site in Ohio, while the other puts usable weights on hardware that is already bought and paid for.

What today's record does not say

No item today carries the term of the Ohio guarantee, what triggers it, or what the site will cost to build. None gives OpenAI's current revenue: the $25B figure above is dated March 2026 and should be read as that. Those numbers are the ones that would turn a headline figure into an assessment, and they are not on the record yet. When they appear, they belong here.