GPU cloud provider CoreWeave transitions from GPU scarcity arbitrage to building a stable, multi-product AI infrastructure platform.

Ahead of CoreWeave Inc.'s Fully Connected conference in San Francisco, we commissioned proprietary customer research with Qualitate based on 13 in-depth interviews spanning more than seven hours of conversation. Rather than rehash earnings calls, we went directly to people evaluating, buying and running the infrastructure, then compared their insights against CoreWeave's financial disclosures and public statements from Chief Executive Michael Intrator and Chief Financial Officer Nitin Agrawal.

Our research shows that GPU scarcity opens the door for CoreWeave, but performance, cost and operating experience give customers reasons to stay. AI inference is growing alongside training—training is not declining at the expense of inference. Inference is growing steeply while training workloads continue to expand. Meanwhile, hyperscalers remain deeply embedded in customer application estates, and not every successful CoreWeave proof of concept converts to a signed customer.

The central question is whether CoreWeave is converting a GPU availability advantage into a durable AI cloud. The customer evidence strengthens that case, though it also reveals exactly where the argument still needs work.

**Survey Methodology**

Our research included nine respondents from large enterprises spanning healthcare, technology, financial services, professional services and retail, plus one other sector. The panel covers the United States, Canada, the Netherlands, the United Kingdom and Germany. Interviewees held senior titles—field chief officer, vice president, directors, functional heads and senior managers—all with influence or decision-making authority over purchasing, technical evaluation and operations.

Critically, this is not a collection of 13 happy customer references supplied by CoreWeave. We worked with Qualitate to reach ground truth. The panel includes users, evaluators and even buyers who evaluated CoreWeave and passed on doing business with them.

The methodology consolidated findings from structured conversations with open-ended responses, not multiple-choice clicks. Qualitate built the response categories after the survey was complete, grouping answers into logical buckets to preserve distinctions between categories.

**Testing Intrator's Demand Expansion Claim**

Of the 13 respondents, seven report significant current CoreWeave use, including one financial-services account still combining pilot evaluation with production workloads. Three are evaluating. One completed a successful pilot but has not rolled out the product. Two decided against CoreWeave—one stayed with Azure, one chose on-premises infrastructure.

Among current users, all seven describe an expansion-oriented outlook for CoreWeave spending or workloads. Some expect to move into higher spending bands; others expect to grow within existing bands or add production usage. Current spending ranges from below $100,000 to above $5 million annually.

This supports management's argument that enterprise demand is becoming real and recurring. But the other six interviews matter equally. Positive AI demand does not automatically become CoreWeave revenue. Buyers can like the technology yet face operating, financing or end-customer-demand constraints.

**GPU Scarcity as Entry Point; Quality as Retention**

Ten of the 13 respondents cited specific hyperscaler GPU-capacity friction or product access issues as part of their buying decision—lead times, regional availability and difficulty obtaining desired configurations.

The more revealing evidence comes after the initial purchase. A healthcare IT director reported that CoreWeave delivered capacity in weeks versus months from Amazon Web Services Inc. That same customer said CoreWeave's superior inference latency and stability would preserve the relationship even if competitors improved prices—evidence of value extending beyond initial availability and supporting Intrator's claims.

Other current users highlighted enterprise support, reliable multi-node performance, Kubernetes fit and the ability to match GPU type to workload as key factors. These are operating benefits. Our research supports the notion that customers are buying more than a chip-hour. However, we would avoid treating this as a universal price advantage, particularly as contract structures and market prices evolve.

With supply shortages and a quality edge, CoreWeave likely has pricing power and will use it. The buying decision will depend on whether customers—like Caterpillar, frequently mentioned in interviews—can monetize AI infrastructure on their end.

**Inference Grows Atop Training, Not Instead of It**

Intrator emphasizes that CoreWeave doesn't separate training and inference infrastructure—it thinks of both as AI Infrastructure. Customer evidence supports this.

All seven current-use respondents run inference on CoreWeave alongside training, fine-tuning or model-development work. One healthcare customer's training-to-inference mix moved from 90-10 in year one to 60-40 today, with a roughly 50-50 mix expected within 12 months. That is a shift in ratio—not evidence that training volumes fell. A separate buyer explicitly stated that training remains steady as new model generations arrive, with inference stacking on top of that baseline.

This matters for the business model. CoreWeave initially engages through training needs and stays involved as models reach production. Other buyers may enter through inference and add model development. This dynamic supports recurring AI workflows.

**Hyperscalers Control Applications; CoreWeave Takes the AI Layer**

Customer data suggests hyperscalers retain the broader application estate while CoreWeave captures the dedicated AI workload. This aligns with Intrator's position that CoreWeave's dominant swim lane is the infrastructure layer required for AI specifically.