CoreWeave is described as an early and innovative market leader in the neocloud category. Founded in 2017, the company provides large-scale GPU compute to startups and major enterprises, and it is the only cloud provider to receive the top “Platinum” rating from AI research firm SemiAnalysis. In a recent Key Context interview, CoreWeave co-founder and chief development officer Brannin McBee and vice president of corporate development and investor relations Nick Robbins discussed the current state of AI demand, the rise of agentic AI and reasoning workloads, and the competitive dynamics shaping the neocloud market.
Agentic AI and reasoning demand began accelerating in the fourth quarter
Asked when the demand wave around agentic AI began to break out, Brannin said CoreWeave saw the real beginning in the fourth quarter of last year. At that time, the company was having engineering-level conversations with customers about products they expected to bring to market in the first quarter of this year. He said this engineering relationship is central to how CoreWeave understands customer demand: it allows the company to see trends ahead of time rather than respond only after changes have already taken place.
From the product side of the AI market, Brannin described the first quarter as a major inflection point for reasoning and AI consumption, adding that the acceleration has continued. When asked about the current state of AI demand and whether recent weeks had shown any signs of slowing, Nick replied that it seems to intensify in new ways every day. The interview framed this acceleration as being driven by agentic AI, reasoning models and enterprise-level AI applications rather than by a single wave of GPU buying alone.
CoreWeave’s position in the AI infrastructure chain gives it a broad view of this shift. The company serves leading customers including OpenAI, Anthropic, Meta, Google, Microsoft and Nvidia, while also seeing demand from research labs, enterprise customers and hyperscale cloud providers. As a result, its executives argued that the central question is no longer simply whether GPUs are scarce. The workload mix itself is changing, and that change is affecting data center design, server configuration, storage planning and the broader infrastructure delivery process.
CPU and storage requirements are rising alongside GPU demand
Brannin said CoreWeave has been running CPUs since 2023 and has always had a complete cloud product. In his view, the question is not whether the company has only just begun adding CPUs, but what customers actually need and whether the relative demand for CPU capacity is rising. His answer was clear: it is. He also said that as agentic and reasoning capabilities become real components of models, storage requirements are rising compared with previous generations, and he expects that trend to continue.
Nick added that customers should expect to see a large number of Vera CPUs deployed next to a large number of Vera Rubin servers. He said CoreWeave fundamentally redesigned its base data center plan last year to make room for more storage and more CPUs beside GPUs. The reason, he explained, is the company’s unique position in the ecosystem: as an independent AI cloud provider serving Anthropic, OpenAI, Meta, Google, Microsoft, Nvidia and others, CoreWeave can understand where customers are taking the technology and plan infrastructure accordingly.
On whether CoreWeave will mainly use Nvidia Vera CPUs in the future, Nick said the answer depends on the workload. The company operates according to customer demand. It has already disclosed that it expects to be an early and important adopter of Vera CPUs, while its current fleet is still mainly AMD. Over time, that mix can change based on customer needs, and Nick said customer interest in Vera CPU is very strong. Brannin used this point to explain CoreWeave’s contract model: more than 98% of revenue is contract-driven, and the company is not guessing what infrastructure customers want. Customers specify the required configuration, and those customers define what CoreWeave builds.
Execution and engineering capacity define the competitive landscape
When discussing competition with neocloud companies such as SpaceX, Nebius and Oracle, as well as hyperscale cloud providers such as Azure, AWS and Google, Brannin pointed to third-party validation. He said nine of the top ten AI labs outside China use CoreWeave’s platform, and that SemiAnalysis has consistently placed CoreWeave alone at the highest level for performance. He also said he does not believe CoreWeave receives GPU allocations because of any personal relationship with Jensen, but because suppliers have deep confidence in the company’s execution record and engineering capability.
Nick said CoreWeave wins hyperscale cloud customers because it is strong at execution: it can build these systems very quickly and make them run well. It wins research lab customers because it offers the highest-performing version of the technology and delivers the best efficiency per token. It wins enterprise customers because the infrastructure works well and because the company has built what he described as a best-in-class orchestration layer, which is part of the basis for recognition such as the Platinum rating.
He added that an increasingly important part of CoreWeave’s position is the layer it has built around inference and developer tools. Among AI cloud providers, he said, CoreWeave has developed a mature capability set that helps enterprises put AI into production. In practical terms, he described this as building and delivering products that help less technically mature enterprises turn data into models and then into internal agents, while also allowing CoreWeave to cross-sell its cloud services during that process.
Powered data center shells are the main bottleneck
Asked about the current bottleneck in the AI infrastructure buildout, Brannin pointed to powered shells, meaning data center shells with power available. More precisely, he said the constraints involve the components inside those shells. He agreed that electricians are part of the challenge and described the area as complex. However, he emphasized that CoreWeave already has 49 such sites live and running. The company is not relying on one or two sites; it has repeated the process 49 times, which he described as a deep execution record.
That record has also helped CoreWeave accumulate knowledge about supply-chain issues. Brannin said the company has learned how to deal with this supply chain and which vendors in it are suitable partners and which are not. On the question of HBM memory costs and shortages, Nick said CoreWeave’s business model is designed to lock in the GPU price it charges customers at the same time it signs GPU purchase orders and determines what it will pay. More broadly, this means locking in the server price, which includes the cost of HBM. He said component availability is not currently the biggest bottleneck; the biggest bottleneck is powered shell, though that answer can move back and forth at some point in the future.
Nick also discussed the expected deployment ramp for Vera Rubin. He said CoreWeave is clearly the first company in the world to start up and fully validate VR, meaning Vera Rubin cabinets, and that the company did the same last year with GB200 and GB300. He expects VR to begin appearing later this year, while the truly large-scale and very strong deployment ramp will run through 2027. He compared the timing with GB: GB began appearing in 2025, but the real large-scale ramp runs through 2026. A meaningful amount was deployed at the end of last year, but this year is the real year of large-scale GB deployment. Over the next 12 to 18 months, Nick expects VR to follow a very similar cadence.