OpenAI had signed contracts for 10 gigawatts of U.S. AI compute capacity by April 2026. In May, it began offering customers discounted access to compute if they committed to spend over one to three years. One company was making enormous capacity commitments while asking its own customers to make smaller ones.

AI finance is reversing the venture model’s sequence. Equity remains suited to businesses that must discover whether customers will pay; gigawatt-scale AI infrastructure increasingly needs durable demand commitments and lenders willing to fund physical assets against them. The consequential financing decision is not just who supplies capital, but who signs for capacity before the building can earn its keep.

Equity was designed to fund the search for a customer

A software investor can fund a team before it has a repeatable product, then accept that most experiments will fail in exchange for a large return from the few that work. That arrangement gives a young company time to find demand. It is poorly matched to a data center that must obtain power, house chips and service financing costs whether the next model succeeds or not.

The warning was present early. Coverage in 2020 described AI-centric businesses facing high compute use and continuing human-support costs that could depress margins and complicate scaling. Those costs did not make venture capital useless. They meant that finding a customer and serving that customer could remain expensive long after a product’s first sale.

The mismatch becomes sharper when a company commits to a site measured in gigawatts. Investors can underwrite uncertainty about a warehouse robot or a construction-design tool with a venture round. A lender assessing a dedicated computing site needs a more specific answer: who intends to use the capacity, for how long, and what happens if that buyer needs less of it?

OpenAI is collecting commitments on both sides of compute

OpenAI said in April that it had signed contracts for 10GW of U.S. AI compute capacity, including more than 3GW added in the preceding 90 days. It reached a contracting target it had previously set for 2029. Signed capacity is not the same as energized, operating capacity; the announcement established the scale of OpenAI’s commitments, not that every contracted site could already deliver power and compute.

CoreWeave’s agreements show what a supplier can put against that demand. CoreWeave expanded its capacity agreements with OpenAI by $6.5 billion, bringing their potential total value to $22.4 billion. The potential value is not revenue already earned, and the announcement does not establish how CoreWeave financed each asset. It does give a supplier a named counterparty and a large prospective use for the capacity it is assembling.

OpenAI then introduced Guaranteed Capacity, which offers discounted compute access to customers making one- to three-year commitments based on spending levels. A customer gains a claim on future access and gives up some freedom to buy only when convenient. OpenAI gains a clearer view of future demand. OpenAI has not said in these announcements whether the customer reservations back particular upstream sites, but its contracts expose the same financing question at two points in the chain: how much uncertainty will another party agree to carry?

Debt can fund the site, not remove its operating risks

The New York Times cited Refinitiv data putting AI infrastructure borrowing above $100 billion in 2025. Smaller companies paid higher interest rates because debt investors remained wary of unproven AI businesses. Morgan Stanley estimated that hyperscalers would supply $1.4 trillion of $2.9 trillion in future AI infrastructure funding through 2028, leaving debt, private equity, venture capital and other sources to supply the rest. Those estimates describe a financing gap, not a guarantee that every proposed site will qualify for a loan.

Major lenders have had to work out how to finance and insure a novel data-center asset class while limiting their exposure. A long contract helps them assess a buyer’s intended use, but it cannot finish construction, secure electricity or ensure that a counterparty will still need the contracted capacity on the expected terms. The distinction matters most where a single customer’s demand is doing much of the work that diversified usage once did. The related project-finance shift makes contracts central precisely because the buildings are too large to treat as an ordinary software expense.

Power costs also have parties beyond the lender and tenant. When Anthropic, Macquarie and GIC formed Theseus Infrastructure to develop AI computing sites, Anthropic committed to cover consumer electricity price increases associated with the projects. Anthropic’s undertaking does not establish when those sites will operate. It does identify a cost that a capacity contract alone would otherwise leave outside the data-center balance sheet.

Venture firms still fund what cannot yet be contracted

a16z led Tutor Intelligence’s $75 million Series B to support deployment of warehouse robots. It also led Dream’s $17 million seed round for AI world models intended to automate construction design and engineering, and expanded the remit of its $2.2 billion Frontier fund to AI and robotics. Those decisions do not describe a venture retreat. They place equity where product performance, customer workflows and deployment economics still have to be proved before anyone can sensibly promise years of demand.

a16z appeared in 71 articles in the earlier comparison period and 166 in the later one, according to the coverage record. That count measures attention, not dollars invested. Sequoia’s roughly $10 billion in new funds offers a separate reason not to confuse new financing structures with the disappearance of growth capital. Venture firms still have capital; an uncertain first deployment and an operating data center simply ask different things of it.

Deployment has become its own financing decision

OpenAI committed more than $4 billion to its Deployment Company and acquired consulting firm Tomoro to help organizations build and deploy AI systems. AWS backed an AI-focused forward-deployed-engineer organization with $1 billion in resources. SAP launched software to help clients contextualize their data and put agents into business processes. None of those commitments fits neatly into a choice between funding a model and funding a data center. Each addresses the work required before a customer can use either one reliably.

Venture firms are adapting their instruments here too. General Catalyst’s Customer Value Fund made a $500 million commitment to Nooks in exchange for a share of returns generated by the capital it helps provide; the fund also supplied Otter.ai with $320 million in non-dilutive capital. Those arrangements differ from both a conventional equity round and a loan secured against a computing site. They connect financing to a company’s ability to turn customer spending into returns.

OpenAI can offer a customer a one-year compute commitment on a contract. The capacity behind it still needs a lender, a building and a substation.