The expansion of AI computing power demand shows no signs of peaking and is spreading to broader infrastructure layers.
In a recent interview, CoreWeave, Inc. co-founder and Chief Development Officer Brannin McBee and Vice President of Corporate Development and Investor Relations Nick Robbins stated that AI demand is "finding new ways to intensify every day" with no indication of slowing down. The executives noted that as the wave of agentic AI accelerates, market demand for supporting resources like CPUs and storage is rising significantly relative to GPUs, leading to a fundamental redesign of overall data center architecture.
Regarding the current core bottleneck for AI infrastructure expansion, McBee explicitly identified it as the "powered shell"—the data center building itself with completed power infrastructure—rather than GPU chips or HBM memory. This assessment holds direct implications for evaluating the pace of AI infrastructure investment.
Demand Continues to Accelerate, Agentic AI Emerges as New Driver
The CoreWeave, Inc. executives pinpointed the start of this acceleration in AI demand to the fourth quarter of last year. McBee stated that through deep engineering discussions with clients, the company had already anticipated the trend of agentic AI products being launched en masse in the first quarter of this year. "From a product perspective in the AI market, Q1 was a massive inflection point for inference and AI consumption, and it continues to accelerate."
Robbins offered a more direct description of the current demand state: "It's finding new ways to intensify every day."
This perspective stems from CoreWeave, Inc.'s unique position within the AI ecosystem. According to Robbins, CoreWeave, Inc. is the only independent cloud provider serving top AI organizations like Anthropic, OpenAI, Meta, Google, Microsoft, and NVIDIA simultaneously. This provides forward-looking insights into technological evolution, enabling proactive infrastructure planning.
Architectural Overhaul: CPU and Storage Demand Rises Significantly Relative to GPUs
The rise of agentic AI and inference models is altering the hardware allocation logic within data centers.
McBee noted that CoreWeave, Inc. has been operating CPU resources since 2023, but the current trend is a clear increase in the proportion of CPU and storage demand relative to GPUs. "As agents and inference truly take off in models, storage demand is also increasing significantly relative to previous generations. I believe this trend will continue."
Robbins revealed that CoreWeave, Inc. fundamentally redesigned its standard data center blueprint last year to allocate more space for storage and CPUs.
He confirmed that future deployments will see many NVIDIA Vera CPU racks placed side-by-side with Vera Rubin GPU servers.
Regarding CPU supplier choices, Robbins stated the current fleet is predominantly AMD-based, but he expects NVIDIA Vera CPUs to become a key early adoption direction as client needs evolve, noting "significant interest" in Vera CPUs. McBee added that over 98% of CoreWeave, Inc.'s revenue is contract-driven, with clients explicitly specifying their required infrastructure configurations. "They are defining what we build."
Powered Shells: The Most Pressing Expansion Bottleneck
When asked about the biggest current constraint, McBee directly named "powered shells," highlighting the shortage of electricians as a complicating factor. He stated that CoreWeave, Inc. currently has 49 operational sites and has accumulated extensive practical experience dealing with supply chain issues, "knowing which suppliers to work with and which not to."
Robbins addressed concerns over HBM memory costs and shortages, explaining that the company's business model is designed to isolate price volatility risk. By locking in customer pricing when signing GPU purchase orders, it protects profit margins, allowing cost increases to be passed through to clients. He also noted, "Component acquisition is not the biggest bottleneck right now; powered shells are. But at some point in the future, that answer might flip."
Vera Rubin Production: 2027 Set to Be a Major Year
Regarding the production timeline for NVIDIA's next-generation Vera Rubin (VR) platform, Robbins provided a relatively clear expectation.
He stated that CoreWeave, Inc. is already the first company globally to have a VR rack online and fully validated. VR servers are expected to begin delivery in the second half of this year, but the large-scale production ramp will extend throughout 2027.
Robbins compared this pace to the rollout path of the previous GB200/GB300 platform, which began appearing in 2025 but saw its true mass production in 2026. "I expect VR to follow a fairly similar pattern over the next 12 to 18 months."
Building a Moat Through Execution and Ecosystem Depth
Facing competition from hyperscale cloud providers (Azure, AWS, Google) and other emerging cloud service providers (SpaceX, Nebius, Oracle), the CoreWeave, Inc. executives attributed their competitive edge to three dimensions: execution speed, performance, and ecosystem depth.
McBee cited third-party validation data indicating that nine of the world's top ten AI labs (excluding China) use the CoreWeave, Inc. platform, with research firm SemiAnalysis awarding it a unique platinum rating. He believes NVIDIA's priority allocation of GPU resources to CoreWeave, Inc. stems from high trust in its engineering execution capabilities, describing NVIDIA as "a supplier that has deep confidence in our execution track record and engineering capabilities."
Robbins explained the competitive logic from a client segmentation perspective: against hyperscalers, CoreWeave, Inc. wins with extremely fast deployment and stable operation; against research institutions, it wins with the highest performance and per-token efficiency; for enterprise clients, it leverages the best inference and development tool orchestration layer to help them transform data into models and agents, thereby enabling cross-selling of CoreWeave, Inc. cloud services.
Comments