Nebius’s $10B Test of Useful Inference

In Q2 2026, Nebius, the Yandex spinoff, reported $575M in cloud revenue, up 514% year over year and nearly all of its $582M total. It was also planning a $10B, 310MW Finnish data center—an announced investment more than 17 times that quarter’s total revenue.

Nebius will build the campus in phases, so comparing one quarter with the full investment does not measure affordability. It does show what the company must execute: finance facilities, secure accelerators, fill them with contracted work and operate them cheaply enough to preserve margins.

The disclosed contracts and build plans establish demand and intended scale. They leave three unit-economics measures undisclosed: cost per inference, cluster utilization and throughput per megawatt. Useful inference per financed megawatt is the operating test in this analysis, not a reported Nebius metric.

Outside buyers changed what Nebius is

At Yandex, the infrastructure served products owned by the same company. After the spinoff, Nebius began selling capacity to independent customers that can compare its price, performance and delivery with other clouds. The technical inheritance remained; the buyer changed.

Microsoft supplied the clearest evidence of that transition when it signed an agreement worth up to $19.4B through 2031 for Nebius AI cloud capacity. A contract extending through 2031 can support financing for facilities that take years to deliver. It also makes missed delivery dates and service failures more consequential.

The megawatt now carries a balance sheet

The Lappeenranta campus calls for 310MW of capacity, with phased operations planned for 2027. On March 17, 2026, Bloomberg reported that Nebius intended to raise roughly $3.75B in convertible debt for data-center expansion and customized AI chips.

Before the first phase begins operating, Nebius must commit capital, secure hardware and match customer demand to the delivery schedule. Lenders, suppliers and customers must move together because an unfunded megawatt produces no compute.

Bloomberg reported on March 11, 2026, that Nebius planned to deploy more than 5GW of Nvidia systems by the end of 2030 as Nvidia prepared to invest $2B in the company. That target is more than 16 times Lappeenranta’s stated capacity, making the Finnish campus one part of a much larger financing and delivery plan.

CoreWeave pairs commitments with power on a similar scale. CNBC reported on August 11, 2026, that the company had $104B of revenue backlog and 1.5GW of contracted power. Both companies speak in contracted megawatts because electricity access and customer agreements make each other financeable.

Software changes how much plant Nebius needs

Nebius is buying software expertise and specialized hardware to get more work from each powered cluster. It agreed to acquire Eigen AI for about $643M to improve the performance of chips running inference tasks. Nvidia then put Groq 3 LPX into full production with Nebius as its first customer.

Eigen would bring optimization expertise inside Nebius, while Groq adds a dedicated inference accelerator to its fleet. Together, the moves push Nebius above the basic rental layer, where providers sell access to a familiar GPU.

On October 20, 2025, the South China Morning Post reported Alibaba Cloud’s claim that GPU pooling reduced the number of Nvidia H20s required by 82% while serving dozens of language models with as many as 72B parameters. Alibaba produced that result on its own workload, so it cannot be applied to Nebius. It does show how scheduling can change the hardware required for a serving load.

If Nebius controls the scheduler or compiler, it can initially retain the savings as lower serving costs. Model developers can produce the opposite effect for the cloud provider. OpenAI engineers reportedly found a method that could more than halve inference costs, allowing customers to rent less capacity or demand lower prices.

Long contracts replace utilization risk with delivery risk

Microsoft’s commitment makes future utilization more legible, but Nebius still must finance facilities, procure systems and start service before the contract produces cash. Reuters reported on November 11, 2025, that Nebius had also signed an approximately $3B, five-year agreement to provide Meta with AI infrastructure.

Those buyers cover some future demand while putting more weight on Nebius’s delivery schedule. The announced contract values do not disclose price per unit of compute, utilization commitments or operating margins.

In 2025, AI infrastructure companies borrowed more than $100B, while smaller operators paid higher interest rates as lenders scrutinized unproven businesses. Nebius plans to issue convertibles into that market and must offset financing costs with utilization and operating efficiency.

Crusoe shows how a construction plan can fail before a facility opens. The company paused a proposed Wyoming data center after failing to secure customers including Google amid reported concerns about cost and timetable. Without a buyer willing to accept the price and delivery schedule, Crusoe could not proceed.

Large contracts can reduce utilization uncertainty for a specialist cloud while giving a few buyers leverage over price and delivery. Backlog can finance capacity without establishing the returns that capacity will earn.

The $10B campus still lacks a unit-economics score

Microsoft and Meta have committed demand. Nebius has announced power and financing, while its Eigen and Groq moves target operating efficiency. None of those records supplies a comparable Nebius cost per inference, cluster utilization rate or throughput-per-megawatt figure.

Cheaper inference can lower Nebius’s serving costs. The same advance can reduce rental prices and the amount of capacity customers need. If Nebius raises throughput and utilization faster than prices fall and financing costs rise, 310MW can support margins. If it cannot, the $10B Lappeenranta campus becomes a very large warehouse of increasingly ordinary compute.

The campus already has a price and a power rating. Its unit-economics score remains undisclosed.