Nvidia's $3.5B Bet Has an Address

$3.5 billion bought no electricity. In February 2026, Nvidia disclosed that amount in guarantees for companies leasing land, power and data-center facilities. Its exposure was four times its Q3 level. The chip supplier had begun putting its credit behind places where its chips could run.

In the AI infrastructure buildout, AI cloud is not settling into the familiar commodity-cloud playbook. Scarce accelerators matter, but a fleet earns revenue only when a builder can secure power at a usable site, finance construction, obtain hardware and deliver capacity against a customer’s schedule. Compute has become a location-specific asset before it becomes a utility service.

Chips cannot outrun the grid queue

Conventional cloud let customers consume variable workloads without owning a data center. The provider absorbed the physical complexity. AI customers still want that convenience, but their largest commitments now expose what sits underneath it.

In October 2025, Nscale agreed to an up-to-$14 billion Microsoft deal involving about 104,000 Nvidia GB300 chips in Texas within 18 months and 12,600 GPUs in Portugal. The contract named hardware, quantities and places. In June 2026, Nebius announced a $10 billion, 310MW campus in Lappeenranta, Finland, with phased operations expected to begin in 2027. Neither an order for chips nor an investment plan makes a site operational.

Microsoft supplied the sharper test. It canceled an anchor-tenant agreement for 23,000 Nvidia chips at Nscale’s Essex site because the facility might not receive grid power until the 2030s. Texas, meanwhile, halted approvals for data-center grid connections while more than 474GW of projects sat in its queue—over five times the grid’s peak demand. A queue measures requests, not deliverable power. PJM Interconnection proposed that large data centers bring generation or curtail usage; Google contracted for 3,590MW from Constellation Energy. Both treated electricity procurement as an operating decision, not a background bill.

A powered site is still not sufficient. Crusoe paused a planned Wyoming data center after failing to win customers, according to reporting that cited concerns about cost and timing. Hardware, grid access and demand have to meet at the same address.

A reservation reduces demand risk, not delivery risk

Meta plans to spend up to $27 billion over five years for Nebius AI infrastructure, starting with $12 billion of capacity in early 2027. That gives Nebius a prospective buyer. It does not pay for every step required to deliver the first tranche. After the Meta deal, Nebius planned roughly $3.75 billion in convertible debt for data-center expansion and customized AI chips.

Long-term commitments have become a product, too. OpenAI offers discounted Guaranteed Capacity in return for one- to three-year customer commitments. Those reservations can make future utilization easier to estimate. Their value to a lender still depends on the customer, the terms and the builder’s ability to commission the capacity. The contracted megawatt is not the same unit as the operating megawatt.

Nscale’s chief executive said the company had more than $100 billion in contracted revenue. Nscale also faced data-center delays and investor doubts about its proposed IPO valuation. Separate reports described talks with Anthropic over roughly $45 billion for 460MW at a West Virginia campus; those talks were not a confirmed agreement. Adding a negotiation to a backlog would make the order book look larger without adding a watt.

Nebius appeared in 12 articles in the later comparison period, up from five in the earlier one; enterprise framing rose from 20% to 50%. Those are measures of coverage, not commissioned capacity. They show why customer commitments draw attention even while delivery remains unresolved.

Credit markets distinguish a backlog from a plant

AI infrastructure companies borrowed more than $100 billion in 2025, according to Refinitiv coverage. Goldman Sachs estimated that investors provided roughly $500 billion in financing to the broader category of AI-linked groups during 2026. The two figures cover different borrowers and forms of financing; dividing one by the other would manufacture a trend. Both show that building capacity now requires a capital market, not just an equipment budget.

Developers have sought credit ratings while their facilities remain under construction. Banks have marketed construction loans tied to Oracle’s future leases. Lenders must judge a customer contract, a construction timetable, power delivery and the eventual usefulness of the hardware before the asset produces cash. That is the interval examined more closely in AI cloud project finance.

Credit can reprice during that interval. In October, CoreWeave’s bond yields approached 15% amid borrowing concerns; its shares fell 15.5%, and Nebius shares fell 10% in the same selloff. Strong demand for AI services did not prevent investors from charging builders more for exposure to delayed cash flow. Crusoe’s $3.9 billion raise and its bet on factory-built data centers show the competing response: raise capital and try to shorten construction. Neither removes the need for a customer when the facility is ready.

Nvidia now backs deployment as well as supplying chips

Nvidia’s lease guarantees address one problem for young cloud providers: landlords and power suppliers may want a stronger credit counterparty than an unrated builder can provide. Nvidia has also promised to rent unused GPUs from providers including Firmus and Sharon AI in exchange for a share of revenue. That backstop for idle hardware transfers some utilization risk to the supplier whose chips the providers deploy.

Nvidia planned a $2 billion investment in Nebius as Nebius targeted more than 5GW of Nvidia systems by the end of 2030. Nebius then became the first customer for Nvidia’s Groq 3 LPX inference accelerator. Nvidia has reportedly sought to connect holders of its GPUs with Nordic data-center operators that have room to install them. These are different interventions—equity, access to new hardware and help finding a site—but each addresses a gap between a shipped accelerator and a billable service.

They also create exposure. A guarantee is not revenue, and a promise to rent idle GPUs does not produce an outside customer. Microsoft’s exit from Essex sets a limit on supplier support: Nvidia could help underwrite deployment, but it could not turn an uncertain grid connection into power on Microsoft’s timetable.

Efficiency can improve a working site, not commission a missing one

Nebius reported $575 million in Q2 cloud revenue, up 514% year over year. Reflection signed a deal worth more than $1 billion for Nebius computing capacity that includes access to Nvidia chips. Those commitments push against a purely pessimistic account of the buildout: customers pay substantial sums when providers deliver useful capacity.

Nebius reportedly moved to acquire Inferize for technology aimed at reducing idle GPU capacity and AI-request costs. Better utilization can spread a commissioned site’s fixed costs across more work. CME Group and Silicon Data also announced futures based on daily GPU-rental benchmarks, an early attempt to make compute prices easier to hedge. A benchmark for on-demand rentals cannot, by itself, hedge a particular campus’s construction delay or missing grid connection. It can become more useful as delivered capacity becomes more comparable across sites.

That leaves Nvidia’s opening figure with a precise meaning. The $3.5 billion was neither chip sales nor electricity purchased. It was credit placed behind the interval between a promised AI cloud and one that can switch on.