Three neoclouds, one pattern: revenue up, losses up, capex far ahead of both

CoreWeave, Nebius and Cerebras all reported Q2 2026 results this week, and according to DatacenterDynamics all three accelerated revenue while their losses grew, because all three are still building data center capacity as fast as they can finance it ^[1]. The useful signal is not the growth rate but the ratio of spending to revenue, which tells you that compute is being bought ahead of demand and that sellers, not buyers, still hold the pen on price.

CoreWeave's revenue was $2.575bn against $1.212bn a year earlier, up 112 percent ^[2][1]. Operating income went from $19m to a $49m loss, and net loss widened from $290m to $626m, roughly 2.2 times ^[3][4][2]. Adjusted EBITDA of $1.51bn is 59 percent of revenue ^[5][3]; the roughly $2.14bn gap between that figure and the net loss is what the adjustment leaves out, mostly the cost of owning the machines and the money borrowed to buy them ^[4]. Capex was $9.4bn in the quarter, about 3.6 times revenue, with $35bn to $39bn guided for the year ^[6][5].

The physical numbers say the same thing. CoreWeave added nearly 500MW to reach 1.5GW live across 51 data centers, holds 4.2GW of contracted power and targets 8GW by 2030 ^[7][8]. That leaves about 2.7GW committed but not yet energized ^[6]. Backlog was approximately $104bn at June 30, plus more than $25bn of net new commitments in early Q3 ^[9].

The clearest pricing tell came on the call: CoreWeave said it recently signed an A100 contract extending into 2029 at an "attractive price", for a GPU that launched in 2020 ^[10]. Six-year-old silicon holding price is not a statement about the chip. It is a statement about installed, energised capacity being the scarce good.

Nebius grew revenue 454 percent to $582.3m, with AI cloud up 514 percent to $575m ^[11][12]. Operating costs and expenses rose to $758.2m from $216.3m, which is $176m more than the quarter's revenue ^[13][7]. Adjusted net loss narrowed to $33.2m from $91.5m, while the net income line swung from a $584.4m profit to a $190.4m loss, mostly because last year's figure included gains on revaluing equity investments ^[14][15]. Capex was $5.7bn, above the $4.7bn analysts expected and close to ten times quarterly revenue, and the shares closed 34 percent higher ^[16][17][8]. Nebius says the payback period on that capex is one year and ten months, and that more than half its expenses are financed by customer prepayments ^[18]. Founder Arkady Volozh said Nebius closed four deals worth more than $1bn each at a yield of $20m to $25m per megawatt ^[19][20]; the counterparties named were Cohere, Reflection AI, an unnamed neocloud and a trading firm ^[21]. End-2026 contracted power guidance rose to 5GW ^[22], which at that yield implies $100bn to $125bn of annualised revenue if it is all built and sold ^[9].

Inference, the workload that would eventually justify the buildout, is still small. CoreWeave's managed inference went from $1m to more than $100m ARR and should exit 2026 at $250m or better ^[23]; Nebius said production inference workloads more than tripled ^[24]. Against a $104bn backlog, $100m of inference ARR is a tenth of a percent ^[10].

Watch three things: whether new contracts hold price as CoreWeave's 2.7GW and Nebius's planned 1GW a year from 2027 come online ^[8][25]; whether customers keep prepaying, since prepayment is the buyer conceding pricing power; and where the capacity lands, given Michael Intrator's argument that US data center moratoriums will change where infrastructure gets built rather than the demand for it ^[26]. Cerebras's detail was cut off in the material available here beyond the report that its cloud revenues rose and losses grew ^[1].