CoreWeave co-founder and chief development officer Brannin McBee, together with Nick Robbins, vice president of corporate development and investor relations, discussed the state of AI demand and the neocloud market in a recent interview with Key Context. The original article was written by Tae Kim and compiled in Chinese by Peggy for BlockBeats. Their central message was direct: AI demand appears to be intensifying in new ways every day, while the real bottlenecks in the infrastructure cycle are moving beyond the simple question of whether enough GPUs are available.
The interview presents CoreWeave as a company positioned in the middle of the AI infrastructure chain. Founded in 2017, CoreWeave provides large-scale GPU compute for startups and major enterprises. It has been described as an early and innovative leader in the neocloud market and is the only cloud provider to receive the highest “platinum” rating from AI research firm SemiAnalysis. The company serves major customers including OpenAI, Anthropic, Meta, Google, Microsoft and Nvidia, while also seeing demand from research labs, enterprise customers and hyperscale cloud providers.
Agentic AI and reasoning pushed demand into a new phase
When asked when the wave of demand for agentic AI began to break out, Brannin McBee said CoreWeave saw the real beginning in the fourth quarter of last year. At that time, the company was already holding engineering-level discussions with customers about products they expected to bring to market in the first quarter of this year. McBee said this kind of deeply connected engineering relationship with customers is an important lens for understanding demand, because it allows CoreWeave to see trends before they become visible in the broader market.
From the product side of the AI market, McBee described the first quarter as a major inflection point for inference and AI consumption, adding that the acceleration has continued since then. Nick Robbins gave an even shorter answer when asked whether recent weeks had shown any sign of a slowdown. “It seems to intensify in new ways every day,” he said. That comment framed the broader discussion: AI workloads are changing structurally, and compute demand is no longer only about GPUs.
The rise of agentic AI and reasoning models is increasing the importance of CPUs and storage alongside GPUs. McBee said CoreWeave has been running CPUs since 2023 and has always had a full cloud product. In his view, the question is not whether CoreWeave has only just started adding CPUs, but what customers actually need and whether that need is rising on a relative basis. His answer was clear: it is rising. He also said storage demand is increasing compared with prior generations as agentic and reasoning capabilities become more central to models.
Vera CPUs, Vera Rubin servers and redesigned data centers
Nick Robbins said customers should expect to see a large number of Vera CPUs deployed alongside large numbers of Vera Rubin servers. He said CoreWeave fundamentally redesigned its base data center plan last year to leave room for more storage and more CPUs next to GPUs. The reason, according to Robbins, is CoreWeave’s position across the AI ecosystem. He said no other independent AI cloud provider can say that Anthropic, OpenAI, Meta, Google, Microsoft and Nvidia are all its customers.
Robbins described that customer base as creating a useful flywheel, or positive feedback loop. Because CoreWeave works with many of the most advanced technology users, it can understand where customers are taking the technology and then plan infrastructure around that direction. When asked whether the company will mainly use Nvidia Vera CPUs in the future, Robbins said the answer depends on the workload. CoreWeave is customer-driven, and it has already disclosed that it expects to be an early and significant adopter of the Vera CPU.
Robbins also noted that CoreWeave’s fleet is currently still mainly AMD, but that this can change over time according to customer needs. He said customer interest in Vera CPU is very strong. McBee added that this is a useful reminder of how CoreWeave’s contracts work: more than 98% of revenue is contract-driven. The company is not guessing what infrastructure customers want. Customers specify the configurations they need, and those requirements define what CoreWeave builds.
Competition with neoclouds and hyperscalers centers on execution
The interview also covered CoreWeave’s competitive position against neocloud names such as SpaceX, Nebius and Oracle, as well as hyperscale cloud providers including Azure, AWS and Google. McBee said he prefers to look at differentiation through third-party validation. According to him, nine of the top ten AI labs outside China use CoreWeave’s platform, and SemiAnalysis has consistently placed CoreWeave alone at the highest performance level.
McBee said he does not believe CoreWeave receives GPU allocation because of a personal relationship with Jensen. Instead, he said it reflects suppliers’ confidence in CoreWeave’s execution record and engineering capabilities, and their belief that the company can best represent their products globally. Robbins added that CoreWeave wins hyperscale cloud customers because it is very good at execution: it can build these systems very quickly and keep them running well.
Robbins separated CoreWeave’s value proposition by customer type. For research labs, he said the company wins because it provides the strongest version of the technology and the best efficiency per token. For enterprise customers, it wins because the infrastructure works well and because CoreWeave has built what he described as a best-in-class orchestration layer, one of the sources of recognition such as the platinum rating. He also said CoreWeave has built the most mature layer among AI cloud providers across inference and developer tools, helping enterprises put AI into production.
That enterprise layer, according to Robbins, is intended to help companies with lower levels of technical maturity turn data into models and then turn those models into agents that can run internally. CoreWeave can also cross-sell its cloud services during that process. This shifts the competitive discussion away from chip procurement alone and toward the ability to deliver full engineering systems, operate them reliably and optimize the economics of AI workloads.
Powered shells are the current bottleneck
When asked whether the current bottleneck is powered data center shells, GPUs or electricians, McBee answered that it is powered shells, and more precisely the components inside those shells. He agreed that electricians are part of a complex area. But he emphasized that CoreWeave already has 49 such sites online and running. The company is not relying on one or two sites, he said; it has already done this 49 times.
McBee said that execution record gives CoreWeave deep knowledge of how to manage supply chain issues and which suppliers in the chain are suitable partners. Robbins was also asked about HBM memory cost and shortages, and how those costs are handled with customers. He said CoreWeave’s business model is designed so that when it signs a GPU purchase order and determines what it will pay, it also locks in the GPU price it charges the customer. More broadly, this means the server price, which includes HBM costs.
Robbins said component availability is not currently the biggest bottleneck; powered shells are. He added that this answer can change back and forth at some point in the future. On Vera Rubin deployment, Robbins said CoreWeave was the first company in the world to start up and fully validate VR, meaning Vera Rubin racks. He said the same was true last year with GB200 and GB300.
Robbins expects VR to begin appearing later this year. He expects a truly large-scale and very strong deployment ramp to run throughout 2027. He compared that cadence with GB: GB began to appear in 2025, but the real large-scale ramp runs through 2026. In his view, a very similar rhythm should play out for VR over the next 12 to 18 months.