CoreWeave, Nebius show cloud AI infrastructure now hinges on power access

For the past few years, the defining question in cloud AI infrastructure was simple. Who has the GPUs? That question is shifting faster than most people expected, and two earnings reports this week made the transition hard to ignore.

CoreWeave posted first-quarter revenue of 2.08 billion dollars, up 112 percent year over year. Its cloud revenue backlog jumped to 99.4 billion dollars from 25.9 billion a year earlier. Nebius, meanwhile, grew quarterly revenue by 684 percent to 399 million dollars. Both companies also raised capital expenditure guidance sharply, with CoreWeave projecting up to 35 billion dollars in cloud infrastructure spending for 2026 and Nebius lifting its annual capex range to between 20 billion and 25 billion dollars.

The growth figures are striking. However, what sits behind them tells the more important story about where cloud competition is actually heading.

Both companies crossed into gigawatt-scale power territory during the quarter. CoreWeave surpassed 1 GW of active power across its cloud fleet and expanded contracted capacity to over 3.5 GW. Nebius expects to exceed 4 GW of contracted power by year’s end and announced a Pennsylvania cloud campus supporting up to 1.2 GW. Consequently, conversations that once centered on server availability now involve utility planners and national grid capacity.

Analysts describe a clear shift in how cloud AI providers compete. Previously, neocloud operators differentiated on GPU access. Increasingly, they compete on how quickly they can secure power, deploy cooling systems, and build the high-speed networking fabrics that large-scale cloud inference demands. GPU availability alone no longer wins contracts.

This connects directly to how enterprises now run AI inside cloud environments. Training happens periodically. Inference runs continuously. Production-grade cloud AI systems need persistent, low-latency infrastructure operating around the clock, which raises the infrastructure bar considerably compared to early experimental deployments.

CoreWeave’s new 21 billion dollar commitment from Meta further illustrates how hyperscalers now lean on specialized cloud providers to absorb AI capacity they cannot build fast enough internally. In effect, neocloud operators are becoming outsourced cloud infrastructure arms for some of the largest technology companies operating today.

For cloud buyers evaluating providers, the competitive question has fundamentally changed. It is no longer just about which cloud has the chips. It is about which cloud can actually keep them running.