CoreWeave Grew Revenue 112% and Nebius Grew Revenue 514% in One Quarter. Neither Number Means Your GPU Rental Gets Cheaper.
On August 12, 2026, CoreWeave stock jumped 19% and Nebius jumped 34% in a single session, both on the back of Q2 earnings that beat expectations by a wide margin. CoreWeave posted $2.6 billion in quarterly revenue and a $104 billion sales backlog. Nebius posted $575 million in AI cloud revenue, up 514% year over year. Read as a headline, that's "AI demand is unstoppable." Read as someone who pays a monthly GPU bill, it's a warning that the discount you've been waiting for isn't showing up this year.
The numbers, and where they actually came from
I went to the primary sources instead of trusting the wire summaries, because "up 500%" headlines get rounded in weird directions.
CoreWeave's own Q2 2026 press release puts revenue at $2,575 million, up from $1,212 million in Q2 2025. That's 112% year-over-year growth, not a marketing rounding job, it's the real number. Revenue backlog hit approximately $104 billion as of June 30, and that figure doesn't even include the more than $25 billion in net new customer commitments the company added in early Q3. Back in November 2025, CoreWeave reported a Q3 backlog of $55.6 billion. So in about eight months, the backlog nearly doubled. CoreWeave is guiding Q3 2026 revenue to $3.45 billion to $3.6 billion, which would be another jump from Q2.
Nebius told a similar story from a smaller base. Per the company's own shareholder letter, total Q2 group revenue was $582.3 million, up 454% year over year. The AI cloud business specifically, which is about 98% of the company's revenue, brought in $574.9 million, up 514% year over year. Annualized run-rate revenue hit $3.0 billion, up 598% year over year and up 56% from the $1.9 billion reported just one quarter earlier in March. Nebius also disclosed something more interesting than the growth rate: the payback period on new capacity deals shortened to 1 year and 10 months, down from a historical two-to-three-year range, because roughly 70% of deals now include customer prepayments covering half to 60% of the associated buildout cost. On the stock side, NBIS closed around $259 after the 34% single-day pop, well above its 52-week low near $62.
Neither company is profitable on a GAAP basis, for what it's worth. CoreWeave posted a $626 million net loss in the quarter even with 59% adjusted EBITDA margins, because the depreciation and interest expense on all that GPU hardware is enormous. That's a separate conversation from whether the business is growing, but it matters if you're trying to guess how these companies price.
Why this is a supply signal, not a "AI demand is infinite" signal
Here's the part that headline readers skip. A $104 billion backlog doesn't mean CoreWeave has $104 billion sitting around waiting to be spent on discounts. It means $104 billion of future revenue is already contracted, and the company is racing to build the power and floor space to deliver on those contracts. Nebius signed four AI cloud deals in Q2 that averaged over $1 billion in total contract value each, with yields of $20 million to $25 million per megawatt, up more than 30% from Q1 pricing on older-generation GPUs. Prices per unit of capacity went up during the same quarter that revenue went up 514%.
That combination, rising volume and rising unit price, only happens when demand is outrunning what a company can physically build. If neoclouds had spare capacity sitting idle, they'd be discounting it to fill it. Instead Nebius ran its first capacity auction and is now testing $40 million to $50 million per megawatt pricing on short-term deals, a straight-up premium tier for customers who need compute now and can't wait for the standard buildout timeline. That's not what a market with excess supply looks like.
What it means if you're renting from a smaller provider
If you're a solo operator or small team buying GPU time from a neocloud (CoreWeave, Nebius, Lambda, Crusoe, or any of the smaller players), the leverage in that relationship currently sits with the provider, and there's no evidence that's about to flip. These companies are pre-selling 2027 capacity to customers with billion-dollar contracts and investment-grade credit. If you're negotiating a monthly on-demand rate against that backdrop, you are not the priority customer, and you're not going to out-negotiate a hyperscaler-backed contract with $40 billion in commitments behind it (that's Nebius's own disclosed total customer commitment figure, not a made-up number).
The practical version of this: budget for GPU rental costs to stay flat or rise over the next two to three quarters, not fall. If your current provider quotes you a renewal at the same or a slightly higher rate than last year, that's not price gouging, it's consistent with what both companies are telling their own shareholders about deal economics. Where you do have leverage is in commitment length and workload flexibility. Shorter, spot-style workloads are getting priced at a premium right now because that's exactly the capacity neoclouds are keeping scarce on purpose. If you can commit to a longer window or run less time-sensitive batch jobs, you're closer to the pricing tier hyperscalers get, not the one desperate short-term buyers get.
The honest take
Revenue growth curves this steep have a history of not lasting. Triple-digit and quadruple-digit year-over-year growth is exactly the shape of number that shows up right before an infrastructure cycle cools, because it's easy to grow 500% off a small base and much harder to keep growing 500% off a large one. Nebius's own quarter-over-quarter growth (58% for ARR, 46% for group revenue) is still fast but noticeably less explosive than the year-over-year number, which is the first sign the base is getting big enough that the comps get harder. CoreWeave's GAAP losses are also widening in dollar terms even as margins improve, which means the business is not yet self-funding, it's still dependent on debt and equity markets staying open to fund the buildout. If capital markets get skittish about AI infrastructure spend, or if a couple of the anchor AI-lab customers behind these backlogs slow their own spending, both of these growth curves bend fast.
So the honest version of "plan for GPU prices to stay high" comes with a caveat: don't build a multi-year cost model that assumes today's scarcity pricing is permanent either. A solo operator locking into a three-year contract at peak-scarcity pricing because "the trend will obviously continue" is making the same mistake in the other direction. What I'd actually do is keep contracts short enough to reprice within 12 months, keep at least one alternate provider evaluated so switching isn't a research project when you need it, and treat this quarter's numbers as evidence that near-term relief isn't coming, not as a forecast for 2028.
Author
Lukas
@lukcombinator