Who Pays for Nebius’s $10B Campus Before GPUs Earn?

Nebius plans a $10 billion, 310MW Finnish data center, with phased operations scheduled to begin in 2027. It must fund the buildings and power before a rack there can earn revenue.

AI cloud is becoming an infrastructure-finance business before it has demonstrated durable compute margins. A builder’s advantage increasingly depends on whether customer contracts, collateral and credit support can pay for land, power, construction and accelerators until utilization revenue arrives. Faster chips improve the eventual product; they cannot finance the interval on their own.

GPU-backed borrowing turned cloud expansion into an asset-finance problem

The financing model appeared before today’s campus plans. In 2023, CoreWeave raised $2.3 billion in debt collateralized by Nvidia chips. By 2024, Meta was describing two AI clusters containing 24,576 H100 GPUs apiece. The physical unit of competition had grown well beyond a server order: operators needed assembled systems, buildings and reliable power at cluster scale. CoreWeave’s borrowing supplied an early answer to who would pay before customers used that capacity.

CoreWeave still had to earn enough from those chips to repay its lenders. Its GPU-collateralized strategy helped it expand rapidly, but reporting in 2025 put its debt at $11.2 billion. Buildings can serve successive generations of equipment; accelerators face a much shorter commercial life. Lenders and operators therefore have to match debt terms not just to a campus lease, but to uncertain utilization and the value of the chips securing a loan. Moving borrowing into a special-purpose vehicle changes where the obligation sits, not whether the underlying capacity must earn enough to support it. The related project-finance buildout extends that structure across other operators.

Nebius must build the asset before its largest contract starts

Nebius makes the timing unusually plain. After its Meta agreement, the company said it planned to raise about $3.75 billion in convertible debt for data-center expansion and customized AI chips. It also announced a $10 billion, 310MW data center in Lappeenranta, Finland, to be built by Polarnode, with phased operations scheduled to begin in 2027. Those are separate figures—not a claim that the convertible raise funds the entire Finnish campus—but together they show the scale of spending required before the planned capacity is operating.

Nvidia planned a $2 billion investment in Nebius as Nebius targeted deployments of more than 5GW of Nvidia systems by the end of 2030. Neither the investment plan nor the deployment target establishes completed capacity. Nebius still needs construction, power, equipment and customers to line up on a timetable that its financing can bear. That is the capacity lag hidden inside a cloud-access agreement.

Customer contracts can support construction without proving utilization

Meta said it could spend up to $27 billion over five years on Nebius infrastructure, with $12 billion of capacity beginning in early 2027. Reflection separately signed a deal worth more than $1 billion for Nebius capacity, including access to Nvidia chips. These agreements give financiers named prospective buyers and substantial contract values to assess. “Up to” spending and capacity scheduled to begin next year are not the same as revenue already earned, however; lenders still have to evaluate delivery terms, customer credit and construction risk.

Other builders are making the same move. Meta was reportedly under contract to buy roughly 1.6GW of capacity from Crusoe across sites in Texas and Missouri. Oracle signed about $150 billion in data-center leases in one quarter, bringing its reported data-center and cloud-capacity commitments to $248 billion. Banks subsequently marketed more than $56 billion in investment-grade construction loans tied to Oracle’s future leases. Oracle’s leases are obligations on its side of the transaction, rather than cloud revenue booked by Oracle; to construction lenders, a creditworthy future tenant can nevertheless make a project easier to underwrite. The distinction between contracted megawatts and operating ones matters throughout the capacity market.

Nebius has operating evidence too. It reported $575 million in second-quarter cloud revenue, up 514% year over year, and $582 million in total revenue, above the $557 million estimate. That growth strengthens the case that its capacity has buyers; the cited results do not establish the margin on its planned expansions. In the coverage record, Nebius appeared in 5 articles in an earlier comparison period and 12 in a later one. Enterprise framing rose from 20% to 50%, while funding framing fell from 20% to zero. That small sample records a shift in attention toward customers, not the disappearance of financing needs.

Nvidia has begun supporting the credit chain that buys its chips

Nvidia is no longer only an equipment supplier to these projects. On February 26, it disclosed $3.5 billion in guarantees for companies leasing land, power and data-center facilities, four times its third-quarter level. A guarantee is contingent support, not $3.5 billion of cash already paid. Its purpose is still consequential: it places Nvidia’s credit behind obligations that a young cloud provider might struggle to support alone.

Nvidia has also promised to rent unused GPUs back from young cloud providers, including Firmus and Sharon AI, for a share of their revenues. That arrangement addresses a different risk from a lease guarantee: GPUs may be installed before paying workloads fill them. In August, Nvidia joined Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR and others on a $500 billion AI-infrastructure funding package. A proposed package is not $500 billion of deployed capital, but the partners identify the institutions now involved in getting capacity built.

Supplier support can make a project more financeable while entangling chip sales with customer borrowing and future cloud demand. The capital cycle sets a limit on that support. AI infrastructure companies borrowed more than $100 billion in 2025, according to Refinitiv figures cited in reporting, while smaller firms faced higher rates. On October 15, CoreWeave’s bond yields neared 15% as its shares fell 15.5%; Nebius shares fell 10%. One trading day does not measure either company’s long-run economics. The borrowing costs show what a leveraged builder may face when capital becomes more expensive before its capacity has matured.

Power and utilization determine what the collateral is worth

A financed GPU produces no cloud revenue while it lacks usable power or sits idle. Google signed demand-response agreements with five US utilities covering up to 1GW, allowing it to reduce data-center consumption during grid peaks. That is a way to manage access to electricity, not 1GW of newly generated supply. Crusoe, meanwhile, said it paused a planned Wyoming data center; sources said it had failed to win customers including Google amid cost and timetable concerns. Some major investors also turned down data-center debt because insurance did not adequately cover risks including natural disasters. Each issue affects what a lender can expect from an apparently valuable site or rack.

Operators are working on the revenue side of the same equation. Nebius agreed to acquire Eigen AI for about $643 million to improve inference-chip performance. An October report said Nebius was acquiring Inferize for $100 million to $150 million to reduce idle GPU capacity and AI-request costs, but the acquisition remained a reported, unconfirmed transaction in the cited record. Nebius also became the first customer for Nvidia’s production Groq 3 LPX inference accelerator. Specialized hardware and better scheduling can improve funded capacity’s economics; none of the cited reports quantifies the resulting margin. Chip choice matters most after the chip can be powered, deployed and kept busy.

Helios offers another rack design, not another source of capital

AMD planned to ship its Helios rack-scale system, estimated to cost more than $5 million, to Microsoft, Meta, OpenAI and others later in 2026. It also secured a deal for more than 500MW of Core Scientific US data-center capacity beginning in 2027, with an option to scale to 2.5GW. The Helios challenge to Nvidia remains commercially meaningful: a second rack-scale supplier gives buyers another hardware path, and AMD’s capacity deal shows that it, too, needs a place to deploy one.

AMD’s contracted megawatts are not yet operating capacity, just as Nebius’s planned Finnish campus is not yet utilization revenue. If customers can sustain their commitments, power remains available and financing terms accommodate the revenue ramp, competing rack designs have a market in which to prove themselves. At Nebius’s $10 billion Finnish campus, a financed rack still needs power and a customer before it can pay its own bills.