The race to build AI infrastructure has hit an unexpected wall: not a shortage of graphics processors or memory chips, but a lack of physical data center buildings with completed power infrastructure and the skilled workers to install it. CoreWeave, the independent cloud provider serving OpenAI, Anthropic, Meta, and other major AI labs, has identified powered data center shells as the single most pressing bottleneck constraining AI expansion today.

What's Slowing Down AI Data Center Construction?

For years, the AI infrastructure conversation has centered on GPU (graphics processing unit) availability and high-bandwidth memory (HBM) shortages. But CoreWeave co-founder Brannin McBee and Vice President Nick Robbins offered a starkly different assessment in recent interviews. The real constraint isn't the chips themselves; it's the buildings that house them and the electricians needed to wire them.

CoreWeave operates 49 data center sites globally and has accumulated hands-on experience navigating supply chain challenges across the infrastructure stack. McBee was direct about the challenge: powered data center shells represent the single biggest expansion obstacle, with electrician shortages compounding the problem. This physical-world constraint is far less visible than GPU allocation battles, but it's proving just as limiting for companies trying to scale AI workloads.

"Component availability is not the biggest bottleneck right now; powered data center shells are. But at some point in the future, that answer could flip," explained Nick Robbins, Vice President of Corporate Development and Investor Relations at CoreWeave.

Nick Robbins, Vice President of Corporate Development and Investor Relations at CoreWeave

The shortage of powered data center shells reveals a fundamental mismatch between semiconductor supply and physical infrastructure capacity. While chip manufacturers race to increase production, the construction and electrical infrastructure needed to deploy those chips at scale cannot keep pace.

How Is AI Architecture Changing to Demand More Power?

The bottleneck is being amplified by a major shift in how AI systems are being built. The rise of agentic AI (AI systems that can plan and execute tasks autonomously) and reasoning models is fundamentally reshaping data center hardware requirements. CoreWeave identified the fourth quarter of 2025 as an inflection point when agentic AI products began moving toward mass market deployment in early 2026.

This architectural shift is creating unexpected demands on data center infrastructure:

• CPU and Storage Growth: As agentic AI and reasoning models take off, CPU and storage requirements are rising significantly relative to GPU capacity, forcing data center operators to redesign their standard blueprints.
• Memory Wall Problem: Traditional GPU architectures struggle with the "memory wall," where data transfer between processors and memory accounts for over 60% of total chip power consumption, leaving GPU utilization below 50% in real-world scenarios.
• Architectural Redesign: CoreWeave fundamentally redesigned its standard data center blueprint to reserve more space for storage and CPU resources, with expectations for large numbers of Nvidia Vera CPU racks deployed alongside GPU servers.

Robbins noted that CoreWeave's business model is structured so that customers explicitly specify the infrastructure configuration they need: "They are defining what we build." This customer-driven approach means data center operators must constantly adapt their physical infrastructure to match evolving AI workload requirements.

Robbins

What Role Are Compute-in-Memory Chips Playing in Solving This?

The semiconductor industry is responding to these power and efficiency challenges with a fundamental architectural shift. Compute-in-memory chips, which perform calculations where data resides rather than shuttling data back and forth between processors and memory, are emerging as a potential breakthrough. These chips achieve significantly better energy efficiency by eliminating the data transfer bottleneck that plagues traditional GPU designs.

Industry experts believe 2026 could become a pivotal year for semiconductor architecture restructuring, with compute-in-memory and application-specific integrated circuit (ASIC) chips offering dramatic efficiency gains. ASIC chips can cut power costs by over 60% compared to GPUs in specific workloads, though they lack the flexibility of general-purpose processors.

However, these architectural innovations don't solve the immediate infrastructure problem. Even if chips become more efficient, the physical buildings and electrical infrastructure needed to deploy them at scale remain the limiting factor. CoreWeave's assessment suggests that the industry will continue to be constrained by powered data center shells and skilled labor availability for at least the next 12 to 18 months.

How Are Global Powers Responding to Infrastructure Constraints?

China's response to these infrastructure challenges reveals how seriously governments are taking the power and space bottleneck. In June 2026, China announced a five-year plan to invest approximately $295 billion to construct a nationwide network of AI data centers, with a mandate that at least 80% of hardware be domestically produced. This massive investment directly contradicts any notion that efficiency gains would reduce the need for new facilities.

China's "Eastern Data, Western Computing" strategy specifically addresses the energy bottleneck by relocating new data center projects from power-constrained eastern regions to western areas with abundant renewable energy sources and cooler climates that reduce cooling costs. The government has reinforced this initiative with strict efficiency mandates, requiring that new data center projects incorporate green electricity and maintain average Power Usage Effectiveness (PUE) below 1.5.

Meanwhile, Amazon is taking a different approach by developing its own custom AI chips. Amazon's internal chip business now generates revenue approaching $50 billion annually, and the company is exploring selling these custom processors to third parties. By controlling both the chips and the cloud infrastructure, Amazon can optimize the entire system for power efficiency and cost, potentially reducing reliance on external power constraints.

The infrastructure bottleneck identified by CoreWeave executives aligns with a broader recognition that AI's future depends less on raw computing power and more on the ability to deliver that power reliably and efficiently. As the industry continues to scale, the companies that can secure powered data center capacity and manage energy consumption most effectively will gain significant competitive advantages.