Key takeaways
- enterprise buyers surveyed across North America and Europe cited hyperscaler capacity bottlenecks and provisioning delays as their initial reason for evaluating CoreWeave.
- Operational performance drives retention: Customers remain on CoreWeave due to bare-metal Kubernetes orchestration, lower inference latency, InfiniBand networking, and dedicated cluster support, rather than simple chip availability.
- The partitioned cloud architecture dominates: Rather than executing full cloud migrations, enterprises keep core transactional databases, ERPs, and governance on incumbent hyperscalers while routing AI workloads to CoreWeave via APIs.
- Inference stacks on training workloads: Workload data shows enterprise inference scaling on top of steady model training and fine-tuning baselines, with enterprise mix shifting toward a 50/50 split over 12 months.
- Governance and compliance tooling remain the expansion bottleneck: While CoreWeave maintains baseline SOC 2 Type 2 and BAA certifications, enterprise buyers seek granular audit logging and real-time dashboards comparable to AWS CloudTrail before migrating regulated data.
- Capital intensity creates balance sheet scrutiny: CoreWeave’s rapid revenue growth ($2.6 billion in Q2, up 112% year-over-year) and $104 billion backlog are counterbalanced by a $10.5 billion half-year CapEx cash gap and high customer concentration.
Enterprise technology leaders evaluating specialized artificial intelligence infrastructure face a critical question: Can CoreWeave convert temporary graphics processing unit scarcity into a permanent enterprise cloud presence? Ground truth evidence from proprietary customer research conducted by Qualitate and SiliconANGLE across 13 enterprise buyers reveals that while GPU access friction initially brings technical teams through the door, superior bare-metal performance, lower inference latency, and responsive Kubernetes support keep them there. CoreWeave is not replacing traditional hyperscalers across general business application estates. Instead, IT organizations are actively adopting a partitioned cloud architecture, running transactional databases and business applications on incumbent clouds while routing latency-critical training and inference workloads to CoreWeave via application programming interfaces.
Why enterprises adopt and retain CoreWeave beyond GPU scarcity
Ahead of CoreWeave’s Fully Connected conference in San Francisco, an examination of 13 in-depth enterprise buyer interviews spanning more than seven hours of structured evaluation sheds light on actual procurement behavior. Conducted with research firm Qualitate, the study analyzed decision-makers across large organizations in healthcare, financial services, retail, technology, and professional services operating across the United States, Canada, the United Kingdom, Germany, and the Netherlands. The cohort comprised senior leaders with technical evaluation and budget authority, including field chief officers, vice presidents, IT directors, and systems engineering heads.
Ten out of the 13 respondents cited hyperscaler GPU capacity constraints or access friction—spanning lead times, regional availability, and rigid cluster configurations—as their primary initial trigger for evaluating CoreWeave. In one representative example, a healthcare IT director reported that CoreWeave provisioned required GPU clusters in weeks, compared to multi-month lead times quoted by Amazon Web Services. However, initial access is merely the catalyst. Customer retention depends on what occurs after deployment. Among the seven respondents actively running production workloads on CoreWeave, all seven reported an expansion-oriented outlook. These organizations plan to either increase spending within their current tier or scale into higher spending bands ranging from under $100,000 to over $5 million annually.
The operational factors cited by current users focus heavily on engineering execution: bare-metal container stability, responsive enterprise technical support, direct Kubernetes integration, and the flexibility to match specific GPU architectures to individual workloads. Crucially, the healthcare IT director noted that CoreWeave’s lower inference latency and cluster reliability would preserve the relationship even if public cloud pricing dropped. CoreWeave is selling operational stability and workload efficiency rather than fungible chip-hours.
| Evaluation Outcome | Respondent Count | Primary Decision Drivers | Operational and Architectural Rationale |
|---|---|---|---|
| Active Production Deployment | 7 organizations | Hyperscaler capacity friction, bare-metal latency, cluster uptime | Workloads span training and production inference; multi-node InfiniBand performance and dedicated Kubernetes tooling outweigh single-cloud consolidation. |
| Active Technical Evaluation | 3 organizations | Workload-specific pricing, cluster availability, API integration | Teams are validating throughput and benchmarking latency against incumbent public cloud instances before committing capital. |
| Completed Pilot Without Rollout | 1 organization | End-customer demand validation and risk management | Technical pilot succeeded, but procurement paused multi-year commitments until internal downstream customer demand proves durable. |
| Retained Incumbent Hyperscaler | 1 organization | Existing enterprise discount agreements, multi-cloud overhead | Evaluator acknowledged CoreWeave technical quality over alternative GPU clouds, but existing Azure commitments and integration overhead prevented migration. |
| Selected On-Premises Buildout | 1 organization | Dedicated facility financing, predictable sustained utilization | Favorable capital expenditure financing for facility renovation made private infrastructure ownership viable for steady-state workloads. |
How bare-metal Kubernetes shifts the performance equation
The technical differentiation underpinning CoreWeave’s retention metrics stems from its foundational infrastructure architecture. In conventional hyperscaler environments, GPU resources are encapsulated within virtual machine instances managed by proprietary hypervisors. While virtualization simplifies multi-tenant provisioning and instance isolation for general-purpose workloads, it introduces latency penalties, virtualization jitter, and resource contention in high-throughput distributed AI operations.
As documented architectural analyses show, bare-metal Kubernetes infrastructure eliminates hypervisor overhead, removing an estimated 65% in computing and latency penalties typical of virtualized cloud stacks. CoreWeave deploys containers directly onto bare-metal hardware, pairing host compute nodes with NVIDIA BlueField Data Processing Units (DPUs) and non-blocking InfiniBand interconnect fabrics. By offloading networking logic, storage encapsulation, and security policies to hardware DPUs, host central processing units and GPUs remain completely unencumbered, dedicated entirely to matrix multiplication and tensor calculations.
This architectural distinction becomes pronounced during distributed training and multi-node inference. Large language models require synchronized parameter exchanges across hundreds or thousands of interconnected chips via Remote Direct Memory Access (RDMA). Within virtualized clouds, minor network jitter on a single virtual machine can stall an entire distributed cluster. On bare-metal infrastructure, deterministic network paths minimize straggler nodes. In head-to-head enterprise evaluations, this operational depth gives CoreWeave an advantage over rival neoclouds. For example, a healthcare director of AI noted that CoreWeave edged out Lambda Labs based on enterprise-grade support, cluster availability at scale, and operational flexibility. Similarly, an enterprise evaluator in Germany that remained with Azure described Lambda’s support and compliance responses as less mature than CoreWeave’s, demonstrating that enterprise procurement demands robust operational frameworks rather than mere hardware access.
The partitioned cloud reality: Hyperscalers keep apps while AI moves
A central finding from the enterprise evidence is that CoreWeave’s success does not require displacing incumbent hyperscalers from the broader enterprise technology estate. All seven current-use respondents retain their primary public cloud relationships with AWS, Microsoft Azure, or Google Cloud. A retail IT director articulated the prevailing enterprise architecture: core enterprise applications run on incumbent hyperscalers, while latency-sensitive AI models hosted on CoreWeave are invoked through high-throughput API connections.
Hyperscalers possess decades of enterprise lock-in across relational databases, analytical data warehouses, enterprise identity frameworks, and negotiated commercial discount structures. As major enterprise software vendors illustrate with their own enterprise integrations with AWS and Google Cloud, organizations maintain deep roots in incumbent hyperscaler ecosystems for data governance, access controls, and core business applications. Rather than undertaking complex migrations, enterprise architects partition workloads: general business logic stays with the hyperscaler, while specialized AI infrastructure is isolated in CoreWeave. GPU spending reflects this division; one enterprise technology buyer allocates 10% to 25% of its total GPU budget to CoreWeave, while a specialized healthcare buyer directs over 50% to the platform.
Furthermore, the data refutes the assumption that the market’s shift toward inference will diminish the need for dedicated training clusters. Chief Executive Michael Intrator has emphasized that CoreWeave treats compute holistically as AI infrastructure rather than maintaining rigid operational silos between training and inference. Customer workload telemetry confirms this continuity. All seven active users run production inference alongside ongoing model training, domain-specific fine-tuning, or architectural evaluation. One healthcare organization reported that its workload mix shifted from 90% training and 10% inference in its first year to 60% training and 40% inference currently, projecting a balanced 50/50 distribution over the next 12 months. Inference is stacking on top of persistent model adaptation cycles, providing a recurring infrastructure pipeline.
Enterprise hurdles: Compliance maturity, on-prem economics, and capital exposure
Despite technical momentum, CoreWeave faces clear operational boundaries, financial trade-offs, and compliance constraints that enterprise technology leaders must factor into architectural roadmaps.
The primary operational hurdle centers on enterprise governance and compliance visibility tooling. While CoreWeave holds baseline SOC 2 Type 2 certifications and signs Business Associate Agreements (BAAs) for healthcare workloads, enterprise directors emphasize that meeting an audit standard is distinct from providing real-time compliance management tooling. A healthcare director noted that while CoreWeave satisfies baseline requirements, it lacks native, real-time audit dashboards and granular logging equivalent to AWS CloudTrail. Consequently, regulated organizations segment their deployments: anonymized model training pipelines run on CoreWeave, while sensitive protected health information (PHI) remains quarantined within AWS. However, data locality can occasionally favor CoreWeave; one technology provider selected CoreWeave for regional inference specifically to satisfy strict client data residency requirements when the local hyperscaler lacked GPU capacity.
A second constraint involves the total cost of ownership (TCO) calculus surrounding on-premises infrastructure. Breakeven models provided by surveyed enterprise evaluators indicate that private infrastructure requires sustained GPU utilization rates between 50% and 80% to justify the capital expense over cloud hosting. While on-prem ownership can be appealing for organizations with favorable data center buildout financing, empirical data from enterprise proof-of-concept deployments indicates that early private AI clusters routinely languish at 10% to 15% sustained utilization. When accounting for specialized electrical engineering, liquid cooling infrastructure, long procurement lead times, and rapid silicon obsolescence, on-prem repatriation remains restricted to a narrow subset of organizations with highly predictable, continuous workloads.
Finally, CoreWeave’s commercial contract structure and financial commitments present strategic considerations for procurement officers. Second-quarter financial disclosures revealed revenue of approximately $2.6 billion—up 112% year-over-year—backed by a contracted revenue backlog exceeding $104 billion. However, capital expenditures reached $14.1 billion over the first six months against operating cash flow of $3.7 billion, creating a $10.5 billion cash requirement funded through aggressive debt and equipment financing. Furthermore, 98% of second-quarter revenue derived from fixed take-or-pay contracts, with three anchor clients accounting for 72% of total revenue. As CoreWeave introduces shorter-term contracts, on-demand pricing, and managed inference services (which expanded from $1 million to $100 million), enterprise buyers must navigate the balance between locking in discounted multi-year capacity and retaining architectural flexibility as model efficiency evolves.
Bottom line
The ground truth evidence from enterprise buyers demonstrates that CoreWeave has successfully transitioned from an emergency overflow provider during peak GPU shortages to a credible, high-performance AI cloud. Its bare-metal container architecture, low-latency InfiniBand clustering, and responsive engineering support provide defensible technical differentiation against both traditional hyperscalers and rival neoclouds. For enterprise IT leaders, managing this shift requires a deliberate, phased approach rather than a wholesale cloud migration.
- Audit workload profiles and latency dependencies: Benchmark multi-node training throughput and inference response times against incumbent virtualized instances. Workloads with strict latency budgets or high distributed communication overhead will extract immediate performance gains from bare-metal architecture.
- Implement a partitioned cloud architecture: Maintain core application stacks, transactional databases, and primary enterprise governance frameworks within established hyperscaler environments, invoking CoreWeave clusters via APIs for dedicated model execution.
- Assess data sensitivity against compliance tooling: Review whether existing compliance requirements permit anonymized data processing on specialized clouds. Where real-time audit streaming is mandated, implement proxy logging layers or retain sensitive data on established cloud platforms until native audit dashboards mature.
- Validate realistic utilization before building on-premises: Conduct rigorous utilization modeling before committing to private data center builds. If sustained baseline utilization cannot reliably exceed 70%, managed bare-metal cloud infrastructure remains more cost-effective than private capital ownership.
- Structure contract durations around model lifecycles: Avoid overcommitting to rigid multi-year take-or-pay agreements for volatile inference workloads. Leverage hybrid commitments that combine predictable baseline reservations with flexible burst capacity or managed inference tiers.
