Earnings Roundup: Neoclouds Shift From GPU Race to Power Wars

extract

Earnings Roundup: Neoclouds Shift From GPU Race to Power Wars CoreWeave and Nebius earnings show the AI infrastructure race increasingly centers on power, networking, cooling, and deployment speed – not GPU supply alone. The AI infrastructure race is starting to look less like a scramble for Nvidia GPUs and more like a fight over electricity. Earnings reports this week from CoreWeave and Nebius offered a snapshot of how rapidly the neocloud market is evolving from a niche GPU cloud segment into a capital-intensive infrastructure layer built around power, networking, cooling, and deployment scale. Both companies posted explosive revenue growth alongside surging infrastructure spending as they expanded AI cloud platforms for hyperscalers, model developers, and enterprise inference workloads. Originally viewed largely as specialized GPU cloud providers built to capitalize on early generative AI demand, neocloud companies are increasingly positioning themselves as long-term AI infrastructure operators alongside hyperscalers. “The AI infrastructure market is bifurcating,” said Dave McCarthy, research vice president of cloud and edge infrastructure services at IDC. “Hyperscalers still dominate general-purpose cloud services, but specialized providers like CoreWeave and Nebius are emerging as major suppliers of high-performance AI compute.” The evolution reflects how AI economics are changing as enterprises move from experimental training workloads toward persistent inference deployments that require sustained utilization, dense GPU clusters, and industrial-scale infrastructure coordination. “The narrative is shifting from model training to large-scale inference,” McCarthy said. “Enterprises are moving from pilots to production-grade agentic workflows, driving demand for low-latency, cost-effective inference infrastructure.” That transition is forcing operators to compete less on raw GPU availability and more on execution – securing power capacity, building high-speed networking fabrics, deploying advanced cooling systems, and bringing infrastructure online fast enough to meet hyperscaler and enterprise demand. CoreWeave Expands GPU-as-a-Service (GPUaaS) provider CoreWeave reported first-quarter revenue of $2.08 billion, up 112% year over year, while revenue backlog jumped to $99.4 billion from $25.9 billion a year earlier. Adjusted EBITDA reached $1.16 billion with a 56% margin, though net loss widened to $740 million from $315 million a year ago. The company crossed 1 GW of active power during the quarter and expanded contracted power capacity by more than 400 MW to over 3.5 GW. Capital expenditures hit $6.8 billion as the company continued building AI infrastructure at industrial scale. CoreWeave raised the low end of its projected 2026 capex range to as much as $35 billion, citing persistent infrastructure component inflation tied to AI demand. CEO Michael Intrator framed the spending surge as part of a long-term infrastructure buildout rather than a short-term market cycle, telling Reuters: “What I’m doing is I’m building a company.” The company also disclosed a new $21 billion Meta commitment tied to long-term AI deployments. Nebius Scales Up Nebius posted even steeper growth, though from a smaller base. First-quarter revenue climbed 684% year over year to $399 million, while Nebius AI revenue surged 841% to $390 million. Adjusted EBITDA swung to a $129.5 million profit from a $53.7 million loss a year earlier. The company raised annual capex guidance to between $20 billion and $25 billion, up from a prior range of $16 billion to $20 billion. First-quarter capex reached $2.5 billion, compared with $544 million a year earlier. Nebius also said it expects to exceed 4 GW of contracted power capacity by year’s end and announced plans for a Pennsylvania AI campus designed to support up to 1.2 GW of power. CEO Arkady Volozh told Reuters that demand continues to exceed available GPU supply as enterprises shift from AI experimentation into production deployment. Analysts’ Take: Power Becomes the Challenge The numbers point to a broader shift underway across the AI infrastructure market. During the first wave of the generative AI boom, cloud providers competed primarily for access to Nvidia GPUs. Increasingly, operators are competing on how quickly they can industrialize AI capacity itself. “The strongest neoclouds are starting to look like a durable infrastructure tier, not just a temporary GPU arbitrage layer,” Sameh Boujelbene, vice president at Dell’Oro Group, told Data Center Knowledge. “The key shift is that competition is moving from ‘who has GPUs available’ to who can industrialize AI capacity fastest – power, land, cooling, networking, financing, and customer commitments.” At gigawatt scale, operators run into a different class of bottlenecks: utility interconnection queues, transformer shortages, transmission access, substation construction, and backup power systems. “What we are hearing in our conversations with vendors and their downstream users is that inference is transforming AI infrastructure from a periodic training exercise into continuous production,” said Don Gentile, analyst, data platforms and resiliency at HyperFrame Research. “Utilization and coordination efficiency now directly drive the economics of these deployments.” Gentile said GPU supply alone no longer determines how quickly operators can scale. “Our HyperFrame Lens research confirms that GPU availability alone is no longer the defining constraint,” he said. “Power procurement, cooling readiness, and deployment velocity now determine how quickly AI infrastructure becomes productive capacity.” Costs Spread Beyond GPUs The spending surge also suggests infrastructure inflation is spreading deeper into the physical AI stack. CoreWeave and Nebius are pouring capital into cooling systems, networking gear, electrical distribution equipment, and power infrastructure as they build denser AI environments optimized for sustained inference workloads rather than intermittent training jobs. “Hyperscalers are evolving quickly, but neocloud providers fill an immediate and durable role that general-purpose cloud environments were not originally designed around,” Gentile said. Utilities and grid planners are already feeling the pressure. Several operators across the US have warned in recent months that large AI data center proposals are straining load forecasts, transmission planning, and long-term generation models. At the same time, hyperscalers continue hunting for AI capacity beyond what they can build internally. CoreWeave’s expanding relationship with Meta reflects how neocloud providers increasingly function as outsourced AI infrastructure operators for giant technology firms chasing faster deployment timelines. That model may also shift bargaining power over time. Long-term hyperscale agreements provide financing leverage and stable demand, but they may also leave neocloud operators dependent on a relatively small number of giant tenants. “Specialized cloud providers are capturing demand by building AI-native architectures without the legacy overhead of traditional clouds,” McCarthy said. “That creates a price-performance advantage for token-heavy inference workloads.”However, McCarthy said the model depends heavily on maintaining high utilization rates as operators continue pouring billions into debt-financed infrastructure expansion. Execution Takes Over Boujelbene said networking architecture is emerging as a central competitive differentiator as AI clusters scale into tens of thousands of GPUs. “At AI scale, the buyer is not just purchasing GPUs; they are purchasing cluster-level performance – how efficiently tens of thousands of GPUs can act like one system,” he said. That dynamic favors operators that design around high-speed interconnects, low-latency fabrics, storage orchestration, and liquid-cooled rack density from the start, Boujelbene said. Gentile said the next competitive divide may depend less on hardware acquisition and more on operational coordination. “AI infrastructure competition is shifting from hardware acquisition toward execution quality,” he said. “The differentiator is becoming the ability to coordinate power, networking, storage, and inference workloads at sustained scale.” The broader message from both earnings reports signals that the next phase of the AI buildout may hinge less on who can secure GPUs and more on who can secure electricity fast enough to keep giant AI clusters running.