For the third consecutive major evaluation, CoreWeave has earned the top Platinum rating in SemiAnalysis’ ClusterMAX assessment. Only two providers earned Platinum and CoreWeave is the only cloud provider to achieve Platinum in all three major ClusterMAX evaluations, through changing hardware generations, increasingly challenging criteria, and an expanding number of providers in the field. 

That consistency matters because ClusterMAX 3.0 significantly raised the standard for production-ready AI clouds. In high-performance AI computing, achieving a top benchmark score once proves you built a great cluster. Holding that title across three consecutive evaluation cycles and through changing hardware generations, tougher testing criteria, and an expanding competitive field proves you built an unmatched operating system. Against this higher bar, SemiAnalysis found CoreWeave’s clusters “exemplary in all crucial categories” and stated that CoreWeave continues to “set the technical bar” for the industry.

CoreWeave remains the operational benchmark for the industry. Beyond getting the basics right, the team has built its own stack for straggler detection, node validation, and bring-up that goes well past what most providers attempt. CoreWeave's sustained excellence across the stack is unique.

Dylan Patel, SemiAnalysis

What Platinum means for customers

Our engineering efforts to bring the most performant, reliable, and secure AI cloud platform for our customers led to our Platinum rating.  This third consecutive rating reinforces how customers can spend more time advancing their AI and less time qualifying infrastructure, tuning networking, diagnosing faulty nodes, or building their own monitoring and remediation systems.

For customers, CoreWeave’s results translate into:

• Maximized GPU utilization: proactive validation, straggler detection, and automated remediation ensure workloads spend time computing, not waiting on degraded hardware
• Accelerated time to production: integrated and validated clusters that reduce the work required before training and inference can begin
• Predictable performance at scale: compute, networking, storage, and orchestration engineered and tested as one system
• Lower operational burden: hardware qualification, lifecycle automation, observability, failure detection, and repair all managed by CoreWeave
• Future-proof continuity: Trust that CoreWeave, a three-time,  consecutive Platinum rated cloud, will continue to deliver repeatability as hardware generations and workload requirements evolve

A single strong result can capture a moment in time. Three consecutive Platinum ratings demonstrate a consistent ability to bring successive AI architectures into production, operate them reliably, and improve them continuously.

ClusterMAX 3.0 raised the bar

AI infrastructure performance is no longer defined by the GPU alone. Modern training, inference, and agentic AI workloads depend on compute, networking, storage, orchestration, observability, and operational processes working together. A bottleneck or failure at any layer can leave valuable accelerators waiting instead of working.

ClusterMAX 3.0 reflects that reality. SemiAnalysis reviewed 77 providers within a market landscape spanning 323 providers, incorporated feedback from more than 200 end users, and updated its criteria across 10 categories. Only 19 providers  earned a Medallion rating, with many others placed in the Participation and Not Recommended categories.

The evaluation also moved to current-generation infrastructure. Providers were asked to supply NVIDIA Blackwell or AMD Instinct MI355X accelerators, high-bandwidth 800 Gb/s RoCE or NVIDIA Quantum-X800 InfiniBand networking, high-performance file storage, S3-compatible object storage, monitoring, and both Kubernetes and Slurm environments.

Testing covered the complete managed-cluster lifecycle:

• Comprehensive Audit: Deep verification of hardware topology, user experience of managed offerings, firmware consistency, security, networking config, storage capabilities, scheduler support, and infrastructure monitoring
• Stress Performance: An intensive 8-hour burn-in of GPU compute and networking fabric, monitoring clock stability, thermal throttling, collective communication latencies, storage throughput and more
• Failure Injection & Reliability: Deliberate rebooting of active nodes, injection of GPU/fabric faults to evaluate whether the provider could detect, isolate, and recover without breaking customer jobs

This is what makes ClusterMAX 3.0 meaningful: it asks not only how fast a cluster runs when everything is healthy but how effectively the cloud keeps workloads productive when real infrastructure inevitably encounters faults.

CoreWeave remains the technical benchmark

CoreWeave again earned Platinum under these expanded requirements. SemiAnalysis reported excellent reliability, effective health checks, and expected results across nearly every test out of the box. It noted that while other providers must be pushed to handle the basics, CoreWeave has gone further, building a significant independent stack for marginal performance and reliability gains beyond what most providers attempt

SemiAnalysis also reported:  

CoreWeave remains the industry’s sharpest, sturdiest, most proactive neocloud. Our CoreWeave clusters are exemplary in all crucial categories. The health checks work as intended, the reliability is excellent, and nearly all tests reach expected values out of the box.

‍

The rating reflects an operating model built around a simple objective: turn as much infrastructure capacity as possible into productive AI work.

Deep experience with provisioning ~10K GPUs a week  before workloads arrive

CoreWeave provisions roughly 10,000 GPUs per week and has a deep dataset on what makes Blackwell infrastructure production-ready. Its Fleet LifeCycle Controller automates provisioning, testing, and monitoring, combining NVIDIA diagnostics with custom tests, predictive failure data, rigorous burn-ins, and AI-driven analysis. CoreWeave also works upstream with manufacturers to catch defects before systems reach production.

That operational experience, combined with close collaboration with NVIDIA and Dell, helped CoreWeave become the first provider to announce a Vera Rubin NVL72 system passing L11 diagnostics. CoreWeave designed and built rack-scale infrastructure through Racky, Valvey, and the Rack Lifecycle Controller, treating power, cooling, hardware, and software as one integrated software-defined system. Our approach on limiting failure domains, recovering quickly, and keeping infrastructure operating despite inevitable component failures demonstrates production-scale capabilities.

Find the failures traditional health checks miss

Complete hardware failures are disruptive, but silent degradation can be more difficult to diagnose. In a large distributed workload, a single GPU can remain online while operating slowly enough to delay every other GPU in the job. These “stragglers” may not generate a conventional error, yet they reduce model FLOPs utilization and waste GPU hours.

SemiAnalysis specifically highlighted CoreWeave Mission Control GPU Straggler Detection. The capability applies proprietary algorithms to fine-grained NVIDIA Collective Communications Library (NCCL)  telemetry to identify the rank, GPU, and node falling out of sync.

Instead of forcing customers to search logs, manually compare nodes, or repeatedly restart jobs, CoreWeave turned a vague performance issue into an actionable diagnosis. Customers can identify the affected infrastructure, drain the node, and requeue the workload to healthy capacity.

Another example is how CoreWeave uses LLMs to analyze fleet-wide operational data, uncover patterns that can predict failures earlier, and continuously refine diagnostic procedures. CoreWeave developed a NodeBot that adds a lightweight reasoning layer to automated alerts, summarizing incidents and recommending next steps so engineers can diagnose issues faster. 

In one case, a thermal interface material pump-out alert indicated that a node should be removed from the cluster for inspection, while a simultaneous power management unit halt alert triggered a competing instruction to reboot it. NodeBot reconciled the conflicting signals and identified the correct course of action. 

Turn fleet experience into automation

Operating infrastructure at scale produces something just as valuable as capacity: operational intelligence.

CoreWeave continuously analyzes failure patterns across its fleet and turns what it learns into automated validation, monitoring, and remediation. At the center of this process is the Fleet LifeCycle Controller, which manages the path from initial power-on through provisioning, testing, production handoff, and ongoing operations.

Deploying new architectures at scale also reveals a long tail of failure modes that standard diagnostics may not cover. CoreWeave supplements industry-standard testing with its own validation suite, including additional large scale testing developed from its experience operating advanced infrastructure.

CoreWeave also works upstream with manufacturers to identify problems earlier in the hardware lifecycle. The principle is straightforward: 

1. Find on the factory floor what should never reach cluster bring-up
2. Find during bring-up what should never reach a customer workload

Every issue becomes an opportunity to improve the system. Engineers identify a pattern, turn it into a test or automated response, and apply that learning across future deployments. Customers benefit from the experience of the entire CoreWeave fleet, not only the infrastructure assigned to their workloads.

Engineering for the next generation of AI infrastructure

As AI infrastructure evolves, CoreWeave’s experience compounds with operating GPU infrastructure at massive scale.  There is a long tail of obscure failure modes that are not comprehensively covered anywhere. CoreWeave tracks these failures, identifies patterns that may predict future faults, and codifies the appropriate responses in our  fleet ops that paves the path to operating future generation platforms. 

As liquid cooling, ultra-dense power architectures, and optical fabrics become standard, the boundary between the data center and the compute cluster has dissolved. Engineering for platforms like NVIDIA Vera Rubin NVL72 requires full-stack integration from facility software down to kernel schedulers.

This operating model applies across every way customers consume CoreWeave, including long-term bare metal contracts. As SemiAnalysis noted, CoreWeave manages the scale-out and frontend networks, burn-in, repair and replacement, and monitoring in all of its bare metal deployments, unlike providers who hand over hardware they cannot fully operate. The rating measures the operational system, and that system underpins every customer commitment we make. 

Built to accelerate the AI loop

AI development is no longer a linear process. Models and agents continuously run, reason, generate data, learn from results, and improve. Every interruption whether caused by a straggling GPU, a network bottleneck, or slow infrastructure recovery delays the next cycle of innovation.

CoreWeave’s third consecutive Platinum ClusterMAX rating validates the infrastructure and operational discipline required to keep this AI loop moving. By optimizing the complete system for goodput, proactively identifying degradation, and automating complexity customers should not have to manage, CoreWeave helps teams turn more GPU capacity into productive AI work.

ClusterMAX 3.0 raised the bar for what customers should expect from an AI cloud. CoreWeave rose with it, providing the reliable, high-performance foundation customers need to run the AI loop faster and build more capable models and agents with every iteration.

Learn more

Watch now: Many Workloads, One Cluster: Training and Inference Without Idle GPUs 

Register now: Engineered for Agentic AI: Exploring NVIDIA Vera Rubin on CoreWeave

Read the blog: What It Takes to Bring Up a Multi-Rack NVIDIA Vera Rubin NVL72 Cluster