Elena VossIn the week ending August 23, 2026, OpenAI signed a 20-year deal for 10 gigawatts of Ohio data-center capacity. Frontier-model rankings changed several times during those seven days; the power contract runs until 2046. Those two clocks made every other major deal look different.
OpenAI put decades of risk behind a temporary lead
The agreement with SoftBank’s SB Energy gives OpenAI capacity on a scale usually reserved for national power systems. Nvidia agreed to backstop part of the completed project’s value.
OpenAI takes long-duration demand exposure. SB Energy takes construction and operating exposure. Nvidia helps absorb financing risk. The model company, energy developer, and chip supplier are underwriting the same system.
Communities could still disrupt that system. Technology companies offered investments, guaranteed jobs, and open houses to secure support for data centers, while a Republican memo warned that local backlash could endanger an Ohio Senate seat. Electricity matters only where voters permit companies to consume it.
Stripe and Nvidia bought leverage without buying models
Stripe agreed to acquire OpenRouter for a reported $7.5 billion. Stripe processes payments; OpenRouter directs business spending across AI models. As customers spread workloads across several providers, Stripe can own the layer that selects, meters, and bills them.
Nvidia followed from the supply side. Its reported Poolside transaction paired a $6 billion non-exclusive license with a $1 billion investment at a $12 billion pre-money valuation. Nvidia also offered jobs to 109 Poolside staffers.
Nvidia committed six times as much to the license as to the equity. It could gain access, financial exposure, and talent without buying Poolside or demanding exclusivity.
Some large Nvidia customers were also reportedly told that prices for Vera Rubin and Grace Blackwell systems would rise by more than 15% in early 2027. Nvidia financed demand and repriced supply in the same week. Customers planning AI capacity must now hedge both model churn and supplier power.
Shorter model leads made the stop button more valuable
London-based Inherent, founded by DeepMind alumni and backed by $50 million in seed funding, said its smaller Faraday agent beat GPT-5.5 and larger Anthropic and OpenAI models at reproducing research-paper findings. Inherent has not yet proved that result beyond its chosen tests. But across the scaling, reasoning, and agentic eras, open models have taken half as long to catch the first closed model.
A laboratory can monetize a capability lead only while competitors cannot reproduce it. As that window narrows, control over production becomes more valuable.
OpenAI paused reinforcement-learning training for two weeks and changed its safety practices after the Hugging Face breach and evidence that Astra may have crossed a critical cyber threshold. The policy reached the training schedule rather than remaining a document beside it.
Two weeks is small beside a 20-year lease. It is large for investors and infrastructure partners accustomed to treating capability development as continuous production.
AI labs are preparing public markets to fund the lease
Anthropic reportedly expected to match or exceed SpaceX’s record-setting IPO and was preparing to file publicly as soon as the end of August. Its bankers had discussed raising more than $100 billion at a $2 trillion valuation. OpenAI’s CFO separately told employees that the company would be public in 2027, or sooner if the business continued to inflect.
Private investors funded the model race. Public shareholders are now being asked to finance obligations measured in decades, spreading the risk of power contracts, chip purchases, and uncertain capability leads across a much larger capital base.
OpenAI’s 10GW obligation makes a one-week benchmark lead look like a volatile input. The durable positions already sit with the companies that supply the power, route the spend, finance the obligation, or can halt the run when capability crosses a line.