The race to become the leading infrastructure for training and deploying large-scale AI models is no longer solely measured by TFLOPs: today, the key is—quite significantly—the effectiveness with which a cloud delivers cutting-edge GPUs, orchestrates thousands of nodes, protects sensitive data, and maintains stable service during demand spikes. On that scoreboard, CoreWeave has once again claimed the top honor from SemiAnalysis: the Platinum ClusterMAX™, a distinction that, according to the analytics firm, no other AI cloud provider achieved in its latest ClusterMAX 2.0 evaluation.
Beyond the medal, this news provides a pulse check on a market where generalist hyperscalers compete with specialized clouds optimized from hardware to scheduler for AI workloads. SemiAnalysis claims to have combined independent testing and customer feedback from dozens of providers before concluding that CoreWeave is the only one to meet the “Platinum” standard in its 2025 cutoff.
What is ClusterMAX™ and why does it matter
ClusterMAX™ is a rating system that evaluates critical dimensions for large-scale training and deployment: security, storage, orchestration, reliability, and availability. It’s not just about measuring GPU count or data center bandwidth; the key is verifying whether the platform sustains the complex operation of multi-node clusters with high utilization, managed failures, and best practices for isolation and compliance.
According to SemiAnalysis 2.0, the Platinum level is reserved for providers that “consistently excel across all criteria: from security posture to operational robustness and the quality of their managed Slurm and Kubernetes offerings.” In other words: having cutting-edge GPUs is not enough; they must be integrated into a system that truly allows you to capitalize on them.
The five areas where CoreWeave stands out
The ClusterMAX 2.0 evaluation attributes leadership in:
- Security: enhanced compliance and specific controls for AI/GPU/InfiniBand environments, with penetration testing focused on these layers, granular VPC isolation, and real-time threat detection.
- Storage: systems CAIOS and LOTA are noted for performance and scalability. In AI clusters, the storage subsystem marks the line between smooth training and a bottleneck that reduces Model FLOP Utilization (MFU).
- Orchestration: explicit recognition of Slurm on Kubernetes (SUNK) and the CoreWeave Kubernetes Service (CKS). The coexistence of Slurm (the de facto standard in HPC/AI) on K8s offers flexibility (native cloud services) without sacrificing fine control of distributed jobs.
- Reliability: active and passive health checks with advanced automation for node replacement and failure recovery. In clusters moving hundreds of GPUs, auto-repair is as vital as uptime.
- Availability: large-scale deployments of clusters GB200 and GB300 (the latest NVIDIA accelerators with CPU Grace + Blackwell GPUs), indicating that cutting-edge compute capacity is truly provisioned and ready for clients.
Table — Summary of Evaluation (ClusterMAX™ 2.0)
| Evaluated Dimension | What SemiAnalysis Looks For | CoreWeave’s Verdict* |
|---|---|---|
| Security | Specific pentesting for AI/GPU/IB, isolation, detection | Leadership in controls and isolation (VPC, enclaves) |
| Storage | Performance, scalability, consistency under load | CAIOS/LOTA praised for throughput/latency |
| Orchestration | Job management (Slurm), K8s, elasticity, flexibility | SUNK + CKS rated best-in-class |
| Reliability | Health checks, self-healing, MTTR, resilience | Advanced automation for replacement and recovery |
| Availability | Access to latest GPUs, capacity | Large-scale deployments of GB200/GB300 |
*Based on the report and CoreWeave’s notes.
What do “MFU” and “goodput” mean (and why are they cited)?
In their communication, CoreWeave states that their infrastructure enables clients to achieve up to 20% higher MFU and 96% goodput. In model training jargon:
- MFU (Model FLOP Utilization) measures the percentage of theoretical FLOPs on the GPU that are actually useful for the model (discounting I/O waits, synchronization, and pipeline bubbles).
- Goodput reflects the useful work relative to the total resources consumed (a proxy for end-to-end efficiency).
In large clusters, the difference between a MFU of 45% and one of 55% can translate into weeks of training saved or, put another way, millions of dollars less in computational bills. That said, it’s important to remember these percentages depend on the model, size, topology, framework, and pipeline hygiene; they are approximate figures.
Slurm on Kubernetes: why does this combination matter
The HPC standard for queues and resource allocation—Slurm—has historically coexisted with Kubernetes, which dominates in the cloud-native space. CoreWeave’s proposal with SUNK (Slurm on Kubernetes) and their CKS aims to offer the best of both worlds:
- Slurm for scheduling distributed jobs, GPU/IB affinity, gang scheduling, and HPC-style queue policies.
- Kubernetes for auxiliary services, networking, and the cloud lifecycle (observability, security, autoscaling outside training, CI/CD integration).
For research teams and MLOps groups already familiar with Slurm but wanting to operate in the cloud without rewriting their tooling, this layer is a practical shortcut.
Security and compliance: from checklist to practice
The fact that the report emphasizes specific pentesting for GPU/InfiniBand is not a cosmetic detail. The shift from monolithic training to multi-tenant clusters connected by low-latency networks opens a new attack surface unfamiliar to teams from a web background. Isolation controls, real-time telemetry, and segmentation policies at the VPC/tenant level are today as critical as encryption at rest or SSO in the panel.
What does the map look like against hyperscalers?
Recognition for CoreWeave does not mean a “game over” for AWS, Azure, or Google Cloud; rather, it suggests that a specialized cloud can optimize the entire chain—from GPU selection to scheduler and storage—for massive AI, and thereby outperform in effective efficiency—MFU, goodput, wait times—in certain training and fine-tuning profiles.
On the other side, hyperscalers offer global scale, a broader catalog (data services, analytics, security, DevOps), mature ecosystems, and frame agreements that often weigh as heavily as the MFU metric. The actual client choice is not binary: many organizations combine layers (data in hyperscale + training in specialized clouds) or employ multicloud strategies by region and GPU availability.
What should AI/MLOps teams consider?
- Queue time vs. SLA delivery: the metric that impacts business results is not just the GPU/hour cost but when the training begins and how many restarts or “failures” need to be absorbed.
- Topology and network: what IB/NVLink/Ethernet options are available? What _bandwidth and latency_ can realistically be sustained with the target model and size?
- High-performance storage: verify how well CAIOS/LOTA (or their equivalents) match the pipeline’s I/O pattern (distributed reads, checkpoints, shuffles).
- Orchestration: if your tooling depends on Slurm, assess the maturity of SUNK (plugins, queue contention, preemptions, isolation).
- Security: request pentest details, tenant isolation policies, support for dedicated VPCs, KMS, and audits.
- Future generations: roadmaps for GB200/GB300 and beyond—the availability agreement is as valuable as any theoretical Linerate.
Beyond hardware: capital, ecosystem, and go-to-market
CoreWeave emphasizes that it does not just sell infrastructure: it invests in startups (CoreWeave Ventures) and sequences services and tools—Weights & Biases (experiment tracking), OpenPipe (RL), Marimo (Python model development), and its recent acquisition of Monolith AI (AI for physics and engineering)—which establish a client ecosystem. For labs or scale-ups, having the compute + tools + support accelerates value. For large enterprises, the key remains SLA, security, and total cost of ownership (TCO).
Reasonable cautions
As with any assessment, ClusterMAX™ is a snapshot at a specific moment using a proprietary methodology. The MFU/goodput metrics come from provider communications and vary significantly based on model, framework, and pipeline hygiene. The advice for CTOs and MLOps is to replicate tests with their own workloads, sign Proof of Concept (POC) with KPIs, and agree on availability clauses/penalties that reflect the business’s true risk.
Frequently Asked Questions
What is the practical difference between “having more GPUs” and achieving a Platinum ClusterMAX™?
The rating focuses on the ability to use them effectively: specific security, storage that doesn’t strangle, orchestration that scales, auto-healing, and credible SLA. It’s not just hardware quantity; it’s a system.
How do GB200/GB300 affect this assessment?
Having clusters of GB200/GB300 suggests early access to NVIDIA’s latest accelerators. For the client, the key questions are real availability (delivery timelines, queues) and stack maturity that makes them usable.
What is “SUNK” and why should I care?
The Slurm on Kubernetes (SUNK) approach enables deploying HPC/AI workloads with Slurm—widely used for research and distributed training—on a Kubernetes substrate. This provides flexibility for auxiliary services without losing Slurm’s fine control over queues and allocations.
Is CoreWeave an alternative to hyperscalers for all scenarios?
Not necessarily. For large-scale training or high-performance inference, a specialized cloud can gain in efficiency and boot time. For analytics, cold storage, generalist DevOps, or complex global deployments, hyperscalers might still be more convenient. Multicloud strategies are becoming increasingly common.
via: coreweave