Nebius's $40 Billion Number Is Backlog, Not Spending, and the Gap Is the Story

Nebius Group says it will spend $20 billion to $25 billion on capital expenditure in 2026, up from a March guidance range of $16 billion to $20 billion. The $40 billion figure attached to recent headlines is a different number: it describes contracted capacity backlog, the total value of customer commitments stacked up over multiple years, not cash going out the door this year. Some tallies of that backlog run closer to $49 billion depending on which options get counted. The distinction matters because Nebius plans to fund its actual capex with customer prepayments and debt, and the size of the backlog is the collateral story investors are being asked to trust.

The company expects more than $9 billion in customer prepayments in 2026. Nearly 70% of its Q2 deals included money paid upfront, before any GPU was delivered. That is the core of what SemiAnalysis has called the neocloud playbook: sign a hyperscaler offtake contract, borrow against it at loan-to-value ratios of 60% to 70%, then use the proceeds to build the next data center. CoreWeave built this model first. Nebius is now running the same play at larger scale, with Microsoft and Meta as the anchor tenants and Nvidia as both supplier and investor.

The deal chronology shows how fast the backlog grew. In September 2025, Nebius signed a five-year GPU supply agreement with Microsoft worth $17.4 billion, with options that could push it to $19.4 billion, tied to a site in Vineland, New Jersey, according to TechRepublic. In March 2026, Nvidia invested $2 billion directly in Nebius as part of a plan to deploy 5 gigawatts of Nvidia systems by 2030, per Nebius's own announcement. Five days later, Meta signed a deal worth up to $27 billion, with $12 billion in confirmed dedicated capacity and $15 billion in conditional purchase options, delivery starting in 2027 on Nvidia's Vera Rubin platform, according to the-decoder. Meta reportedly gets whatever capacity is left after Nebius serves other customers first, a detail that puts a hyperscaler at the back of its own supplier's queue.

That queue is the part of the story the TechCrunch headline points to: delays at the Microsoft-linked data center site raise the question of whether Nebius can deliver capacity on the schedule its contracts assume. Prepayments and loans fund construction today on the promise of delivery tomorrow. If delivery slips, the gap between cash collected and infrastructure built widens, and the company has to cover it with more borrowing.

Nebius has already gone back to debt markets once this year to cover that gap. In August 2026, it upsized a convertible notes offering to $5.75 billion from an initial target of $4.5 billion: $3 billion at 0.50% due 2030 and $2 billion at 4.50% due 2034, according to Bisnow. The company also swapped $800 million of existing 2029 and 2031 notes for equity, which diluted shareholders. That pushed total convertible debt to around $12 billion. The stock fell about 6% on the news, driven by concern over the size of the raise and the dilution, not over demand for compute.

The revenue base behind that debt is still thin relative to the capex it is meant to support. Nebius posted $582.3 million in revenue for the second quarter of 2026, up 454% year over year, against full-year guidance of $3.0 billion to $3.4 billion. That guidance is six to eight times smaller than the $20 billion to $25 billion capex plan for the same year. A year earlier, in the third quarter of 2025, Nebius reported revenue of $146.1 million, up 355% year over year, but missed analyst estimates. Its GAAP net loss was $119.6 million and its adjusted net loss was $100.4 million, 153% wider than the year before, according to Zacks. The stock fell more than 7% that day, the same day the Meta deal was announced. Nebius reports third-quarter 2026 results this week.

Nebius also started selling spot GPU capacity at public, listed prices in October 2026, announced at Nvidia's GTC event, and began offering contracts as short as three to six months alongside its long-term offtake deals, according to Nebius's pricing announcement. A company locking in $27 billion, multi-year commitments from Meta does not need to compete on short-term spot pricing unless it expects gaps in utilization between those large contracts, or unless rivals are already discounting idle capacity. Either reading points to the same underlying fact: long-term offtake contracts don't guarantee that every gigawatt built gets used the day it comes online.

Critics of the neocloud model point to a closed loop: Nvidia invests in companies like Nebius and CoreWeave, those companies use the cash to buy Nvidia GPUs, and Nvidia records the purchase as revenue. Comparisons to vendor financing in the dot-com era, and to mortgage securitization before 2008, have circulated in coverage such as 247wallst. The counter-argument is that Nvidia's balance sheet is large enough to absorb a default from any single customer, unlike the thinly capitalized vendor financiers of past bubbles. Shares of both CoreWeave and Nebius have fallen 41% and 30% respectively from their highs even as the contract backlog has grown, a split that marks skepticism in the stock price even as the deal headlines accelerate.

One detail rarely mentioned alongside the capex figures is where Nebius came from. It is the international remainder of Yandex, the Russian technology company. After Russia's 2022 invasion of Ukraine and the sanctions that followed, Yandex's Russian assets were sold in July 2024 to a Kremlin-aligned buyer group for a reported $2.4 billion to $5.4 billion. The remaining international business was renamed Nebius Group and based in Amsterdam, controlled by founder Arkady Volozh through the LASTAR Trust, according to Yahoo Finance. That origin doesn't change the financing math investors are weighing now, but it means the company running this $40 billion bet started as a Russian search and ad business reconstituted abroad after sanctions, not a startup built from zero to chase the GPU boom.