Nebius's Q2 2026 shareholder letter contained something more useful than another AI revenue-growth percentage: a glimpse into the economics of individual capacity deals.
The company said:
Those are company estimates rather than guaranteed realized returns.
But they provide a much better framework for analyzing Nebius than simply saying “GPU demand is strong.”
At the physical level, Nebius has to acquire:
power
GPU systems
networking
storage
buildings
cooling
It then has to turn those assets into sellable compute.
An idle GPU is not just unused technology.
It is capital earning little or no return.
Power determines how much compute can operate at a site.
That makes megawatts a rough bridge between physical infrastructure and commercial capacity.
A simplified model is:
Deployed MW
×
Revenue per MW
=
Potential annual revenue capacity
The real calculation is far more complex, because different GPU generations and workloads produce different economics.
Still, MW provides a common language for comparing infrastructure scale.
Nebius says the new Q2 deals generated annual contract value above $20 million per MW, with four large deals around $20 million–$25 million.
This should not be read as:
Every Nebius MW will permanently generate $25 million every year.
Different contracts can involve:
It is a deal-level metric, not a fixed utility tariff.
A megawatt filled with older accelerators is not economically equivalent to a megawatt filled with the latest systems.
New architectures can produce more useful compute per MW.
Supply scarcity can also affect pricing.
Nebius said Q2 pricing strengthened not only for new-generation GPUs but also for older-generation capacity.
Imagine two identical data centers.
One has 95% of its sellable GPU capacity contracted and operating.
The other has 50%.
They own similar infrastructure.
They do not have similar economics.
High utilization spreads fixed costs across more revenue.
Low utilization leaves expensive hardware depreciating without enough customer income.
Nebius estimated an approximately 1 year and 10 month payback for the capex and related operating costs associated with Q2 new deals.
That was shorter than its previous two-to-three-year range.
A shorter payback can be powerful because AI hardware becomes obsolete quickly.
The faster a GPU cluster earns back its investment, the less exposed the owner is to the next architecture making the existing equipment less competitive.
This distinction is important.
A deal can have an attractive estimated payback while Nebius as a whole still consumes large amounts of cash.
Corporate spending also includes:
Deal economics and consolidated free cash flow should not be confused.
If a customer funds 50% of the required capex in advance, Nebius has less corporate capital at risk.
That can materially improve the return on Nebius's own invested capital.
This is why prepayment structure deserves almost as much attention as contract price.
Sarah Chen, MEXC senior crypto industry analyst, says the current Nebius numbers show why AI cloud economics cannot be judged from gross revenue growth alone. A 454% increase in group revenue is eye-catching, but the more durable question is how much capital was required to create the additional revenue and how quickly that investment can be recycled into the next deployment. Sarah's work is available through her MEXC author profile.
Chen finds the shorter estimated Q2 deal payback and higher customer prepayments encouraging, but she would not extrapolate them automatically across the entire future 5 GW pipeline. Scarcity pricing can be unusually attractive when customers are desperate for compute. As new capacity enters the market, prices could normalize. The real test is whether Nebius can maintain strong utilization and differentiated software value after GPU supply becomes less constrained.
Nebius's July secured debt facility provides another clue to the economics.
The loan is backed by deployed GPU infrastructure and contracted customer cash flows, with pricing at SOFR + 2.50%.
Nebius says that the customer cash flow plus the financing covers more than 100% of the relevant deployment capex.
If this model is repeatable, capital can be raised at the asset level rather than entirely at the corporate-equity level.
The same factors creating exceptional current returns can reverse.
More GPU supply can lower rental prices.
New generations can depreciate older hardware.
Large customers can negotiate better terms.
Financing can become more expensive.
Utilization can fall.
A business that looks exceptional at $25 million ACV per MW can look very different if future capacity clears at much lower prices.
The most useful operating dashboard includes:
| Metric | Why It Matters |
|---|---|
| ACV per MW | Pricing power |
| Utilization | Asset productivity |
| Prepayment % | External funding of capex |
| Payback period | Speed of capital recovery |
| AI Cloud EBITDA margin | Operating economics |
| Capex | Cost of growth |
| Contracted power | Future scale |
| Debt/equity issuance | Funding cost |
MEXC's broader provides additional context for how Nebius compares with the wider AI capex cycle.
Above $20 million on average, with the four landmark deals around $20 million–$25 million per MW.
Approximately 1 year and 10 months for the Q2 new deals under the company's assumptions.
Nebius said prepayments on relevant deals covered roughly 50%–60% of associated capex.
No.
Because idle GPU infrastructure still incurs depreciation and other costs.
ACV per MW, payback periods and other deal economics are company estimates and may not match future realized returns. Pricing, utilization, hardware costs, depreciation, financing and competitive conditions can change materially.

Summary NBISON and NVDAON sit in the same AI investment story, but at different places in the value chain. NVDAON is linked to NVIDIA. NVIDIA designs and sells accelerated computing platforms,

Summary The easiest way to understand Nebius Token Factory is to ask what Nebius does after it has built the GPU cluster. If the answer is only: “rent the GPU” then the business can eventually become

Summary For an AI cloud, GPUs are only useful after electricity arrives. That sounds obvious. In 2026, it has become one of the most important constraints in the entire AI infrastructure market.

Overview As the computational requirements for frontier artificial intelligence foundation model training and high-throughput inference expand along an exponential trajectory, a multi-tier

Overview Global artificial intelligence compute leader Nvidia has executed one of the most consequential corporate capital allocations in semiconductor history, committing $3.5 billion into overseas

Overview Following more than a year of semiconductor dominance where graphics processing units and physical data center hardware captured the vast majority of artificial intelligence capital flows,

Robinhood Chain's ecosystem surged as daily DEX volume exceeded $560M, while institutional adoption of DeFi and cross-chain infrastructure accelerated. Meanwhile, regulators advanced crypto legislatio

BlackRock Global Head of Digital Assets Robert Mitchnick said Bitcoin’s market sentiment had improved in a “clear but subtle” way as the asset began showing signs of separating from equities. His obse

Summary NBISON and NVDAON sit in the same AI investment story, but at different places in the value chain. NVDAON is linked to NVIDIA. NVIDIA designs and sells accelerated computing platforms, network

Summary “Asset-light” can be a dangerous phrase in an industry filled with billion-dollar data centers. Nebius is not suddenly becoming a software company with no physical infrastructure. What changed

Summary The easiest way to understand Nebius Token Factory is to ask what Nebius does after it has built the GPU cluster. If the answer is only: “rent the GPU” then the business can eventually become

Summary For an AI cloud, GPUs are only useful after electricity arrives. That sounds obvious. In 2026, it has become one of the most important constraints in the entire AI infrastructure market. Nebiu