CoreWeave and Nebius Q2: gigawatts, $9.4bn capex, wider losses

What was reported

On 14 August 2026, Datacenter Dynamics (DCD) reporter Dan Swinhoe published second-quarter 2026 results for CoreWeave, Nebius, and Cerebras. This briefing covers the two graphics processing unit (GPU) neoclouds. Cerebras is a wafer-scale vendor and is covered in a separate AGICY piece.

DCD’s summary sentence is simple. Revenues accelerated. Losses widened as the firms built data-center capacity.

CoreWeave

CoreWeave reported revenue of $2.575 billion for the second quarter of 2026, up from $1.212 billion in the second quarter of 2025. Operating loss was $49 million, a swing from $19 million of operating income a year earlier. Net loss was $626 million, up from $290 million. Adjusted earnings before interest, taxes, depreciation, and amortization (EBITDA) was $1.51 billion, up from $752 million.

Capital expenditure (capex) in the quarter was $9.4 billion. Full-year capex is expected at $35 billion–$39 billion. Those figures are DCD’s report of company guidance, not an AGICY model.

Active power rose by nearly 500 megawatts (MW) in the quarter, including 300 MW in June, to 1.5 gigawatts (GW) across 51 live data centers. Eight facilities were added in 2026. The company expects to end 2026 with 1.85 GW live. Contracted power is 4.2 GW. The 2030 target is 8 GW. The first self-build is due later in 2026. Revenue backlog was about $104 billion as of 30 June 2026, not including more than $25 billion of net new commitments in early third quarter 2026.

Chief executive Michael Intrator said the managed inference platform grew from $1 million to more than $100 million of annual recurring revenue (ARR) and that the firm expects to exit 2026 with at least $250 million of managed inference ARR. Executives said demand for older NVIDIA A100 GPUs remains solid. The company said it signed an A100 contract that extends into 2029. The source does not state the contract price.

On US data-center moratoriums, Intrator told the earnings call that the rules will change where infrastructure is built, not whether demand exists. He said operators have to work with host communities. DCD did not list EU moratoriums in that paragraph.

Nebius

Nebius Group reported revenue of $582.3 million, up 454 percent from the second quarter of 2025. Almost all of that is AI cloud: $575 million, up 514 percent. Adjusted EBITDA swung from a $21 million loss to $236.2 million. Adjusted net losses fell from $91.5 million to $33.2 million. GAAP net income swung from $584.4 million to a $190.4 million loss. DCD notes that much of the 2025 profit came from revaluation of equity securities.

Operating costs and expenses rose from $216.3 million to $758.2 million. Founder and chief executive Arkady Volozh said the firm closed four deals valued at more than $1 billion each, at a yield of $20 million–$25 million per megawatt. He said Nebius contracted another 1 GW in the quarter and expects to end 2026 with 5 GW of contracted power. The plan is to deploy more than 1 GW of compute per year from 2027. Production inference workloads more than tripled in the quarter, according to Volozh as reported by DCD.

DCD named three European site moves in the same week as the results: Vantage in Wales, Greenergy in Estonia, and a second self-built facility at the Mäntsälä campus in Finland. The article does not publish megawatt figures for those three sites.

Power is the binding constraint the numbers show

These prints are energy stories as much as they are software stories. CoreWeave is adding hundreds of megawatts per quarter and still guiding toward 1.85 GW live by year end. Nebius is booking gigawatts of contracted power and quoting dollars per megawatt as the commercial unit.

The European Union (EU) already has a separate gigafactory tender. That process is not in this DCD article. What this article does show is private neocloud capex running in the tens of billions of dollars, with live MW as the scarce input.

What it means for EU compute

Wales, Estonia, and Finland appear here as Nebius expansion points. That is speed-to-power in Northern Europe, not a Cyprus live hall. AGICY’s Vasilikos campus is pre-construction. Do not map CoreWeave’s 51 sites or Nebius’s 5 GW year-end contracted target onto AGICY.

For a buyer who needs NVIDIA hours now, GPU Bridge remains partner hourly capacity, not campus inventory. For a buyer who wants a take-or-pay reservation against a planned EU hall, the product is a Sovereign Resource Agreement (SRA). The DCD article does not price either path.

Older A100 demand into 2029 is a useful reminder. Not every workload waits for the newest rack. The source does not say those A100s sit in the EU.

GPU leasing, rent, buy, or colocation

AGICY paths for this silicon: GPU leasing, rent GPU server, buy AI accelerator (client-site finance), or AI colocation / GPU colo (also searched as collocation). Fleet in view: Tenstorrent Galaxy, Cerebras CS-3, AMD Helios, d-Matrix Corsair. Planned Cyprus campus — pre-construction, not a live hall. GPU leasing estimates · hardware roadmap · SRA.