ClusterMAX
ClusterMAX is a rating and ranking system for GPU cloud providers published by the research firm SemiAnalysis. It scores providers against ten criteria and sorts them into medallion tiers (Platinum, Gold, Silver and Bronze) plus a "Not Recommended" band split into Underperforming and Unavailable [1]. The first version appeared in March 2025, and by the November 2025 revision it covered 84 rated providers out of 209 tracked, which SemiAnalysis bills as the industry standard and which buyers, vendors and third-party pricing analyses now routinely cite [2][3][18].
Origins
SemiAnalysis had been writing about the economics of GPU rental since a December 2023 report, and its October 2024 "AI Neocloud Playbook and Anatomy" popularized the term neocloud. The rating system followed from a gap the firm identified in that market: before ClusterMAX, it wrote, there was no how-to guide for renting a GPU and no independent evaluation of GPU clouds [2]. The first ClusterMAX report, "The GPU Cloud ClusterMAX Rating System | How to Rent GPUs," was published on 26 March 2025 by Dylan Patel, Kimbo Chen, Daniel Nishball and two other authors, and described twelve months of testing behind it [2][7].
The stated purpose was twofold: to help buyers choose among a fragmented supply base, and to raise the standard of the industry itself. SemiAnalysis wrote that "the bar across the GPU cloud industry is currently very low" and positioned the criteria list as guidelines that providers could build against [2]. The firm claimed the first rating covered roughly 90% of the GPU rental market by GPU volume [2].
| Version | Published | Scope |
|---|---|---|
| 1.0 | 26 March 2025 | 26 providers rated, 169 tracked; five tiers ending at UnderPerform [2][3] |
| 2.0 | 6 November 2025 | 84 providers rated, 209 tracked, 140+ end users interviewed, 46,000+ words; ten itemized criteria published [3] |
| 2.1 | April 2026 | Incremental update: new entrants added, four moved into Bronze, no full re-test [4][8] |
| 3.0 | Testing underway as of mid-2026 | Re-test of all providers with expanded benchmarks, focused on B300 with 800G networking; submission deadline of 1 August [4][17] |
Version 2.0 was written by Jordan Nanos, Daniel Nishball, Michelle Shen and four other authors, and was accompanied by the launch of a dedicated website, clustermax.ai, holding the criteria, expectations documents and per-provider reviews, one page per rated cloud [3][10].
Rating tiers
SemiAnalysis stresses that ClusterMAX is a relative system: each provider is judged against its peers, and "Gold and Platinum providers rise to the top by introducing features and functionality that others do not have" [1]. Checking every box on the criteria list is described as a starting point rather than a route to the top tier.
| Tier | Official definition (abridged) |
|---|---|
| Platinum | "The best GPU cloud providers in the industry. They consistently excel across evaluation criteria, are proactive and innovative, and maintain an active feedback loop with their users. In practice they command a pricing premium because their total cost of ownership is better even when the raw $/GPU-hr is higher." [1] |
| Gold | "Strong performance across all evaluation categories with some opportunities for improvement. Gold-tier providers may have small gaps or inconsistencies but are responsive to feedback." [1] |
| Silver | "An adequate offering with noticeable gaps compared to Gold or Platinum. Some users will not consider a Silver provider despite attractive pricing and availability." [1] |
| Bronze | "Fulfills our minimum criteria, and the last tier we directly recommend. Common issues include inconsistent support, subpar networking, unclear SLAs, limited Kubernetes/Slurm integration, or less competitive pricing." [1] |
| Underperforming | "Based on hands-on testing, these providers can quickly rise to Bronze or Silver by fixing one or more critical issues, such as offering only older GPUs, missing basic security attestation (SOC 2, ISO 27001), misconfiguring key server features (leaving PCIe ACS enabled, or failing to enable GPUDirect RDMA), or charging for GPU hours during cluster creation or hardware downtime." [1] |
| Unavailable | "An interesting service we are excited about but cannot verify yet: not launched publicly, sold out with no plans to add capacity, government-only, or otherwise untestable." [1] |
The tier names changed slightly between versions. Version 1.0 used five tiers, with the bottom one called UnderPerform [2]. Version 2.0 renamed the bottom band "Not Recommended" and split it into Underperforming, meaning tested and found wanting, and Unavailable, meaning not testable at all [3][6].
The ten criteria
Version 1.0 grouped its assessment into nine loosely defined areas, including Slurm and Kubernetes, NCCL/RCCL networking performance, and consumption models [2]. Version 2.0 replaced these with ten named categories and, for the first time, published an itemized checklist of what each one tests [3][5].
| Criterion | What it tests |
|---|---|
| Security | Third-party attestation (SOC 1 or SOC 2 Type II, ISO 27001), compliance regimes such as GDPR, PCI, HIPAA and FedRAMP, multi-tenant isolation on the backend fabric via InfiniBand PKeys or VLANs, driver and firmware currency, and protection against container-escape vulnerabilities in the NVIDIA Container Toolkit [5] |
| Lifecycle | Onboarding and offboarding cost transparency, egress fees, delivery-date accuracy, ease of creating and expanding a cluster, out-of-the-box GPUDirect RDMA configuration, support quality, and audit-log completeness and retention [5] |
| Orchestration | User, group and RBAC management, SSO, and the maturity of managed Slurm and Kubernetes offerings, including topology configuration, the Pyxis container plugin, kubeconfig access and ReadWriteMany storage classes [5] |
| Storage | A high-performance parallel filesystem out of the box (Weka, DDN or VAST), managed S3-compatible object storage, mounts that stay mounted, measured read and write performance, snapshots, backups and published durability SLAs [5] |
| Networking | InfiniBand or RoCEv2 availability, HPC-X MPI, sane default NCCL configuration, SHARP support, and four-node NCCL tests plus PyTorch training runs hitting expected bandwidth and model FLOPs utilization [5] |
| Reliability | Hardware uptime SLAs, 24x7 support response commitments, link-flap and filesystem stability, and detection of GPU fall-off-the-bus events, PCIe errors, ECC errors and NVIDIA XID/SXID codes [5] |
| Monitoring | A pre-installed Grafana or equivalent dashboard, DCGM metrics including SM activity and TFLOPs estimation, job accounting, custom alerting, automated active and passive health checks, and automatic node draining [5] |
| Pricing | Price per GPU-hour at a given quality level, the range of commitment terms offered, whether storage and networking are bundled or charged separately, and contract expansion terms [5] |
| Partnerships | NVIDIA Cloud Partner certification, AMD Cloud Alliance status, chip-vendor investment, a SchedMD relationship, participation in industry events, and ecosystem support and integration [5] |
| Availability | Total GPU quantity and cluster-scale experience, utilization rates and capacity planning, which GPU models are live, and the roadmap for future silicon [5] |
Alongside the criteria, SemiAnalysis published five "expectations" documents describing what a well-configured Slurm cluster, Kubernetes cluster, standalone machine, monitoring stack and health-check system should look like, and encouraged providers to build against them [3].
Methodology
The evaluation combines three inputs: hands-on workload testing, documentation and compliance review, and interviews with customers [1]. For version 2.0 the firm interviewed more than 140 end users of neoclouds, and testing ran from August to September 2025 [3].
The hands-on portion is a roughly five-day engagement per provider, which SemiAnalysis acknowledges cannot measure reliability at scale over time [3]. Within that window it runs a battery of deliberately simple tests treated as proxies for deeper configuration problems: time to install PyTorch, time to pull a 22.1 GB NGC container, time to run "import torch", time to download and load a 15-billion-parameter model, a whois lookup on the machine's public IP to expose aggregators reselling someone else's capacity, and nccl-tests or rccl-tests against expected algorithm and bus bandwidth [3]. SemiAnalysis reported that top-tier providers maintaining local NGC mirrors completed the container pull in under ten seconds while many Bronze-tier providers exceeded four minutes [3].
The process is not blind. SemiAnalysis sent its itemized criteria list to every provider it had working relations with on 6 August 2025, and reported that many then delayed cluster handover while they scrambled to meet the requirements, in some cases installing Slurm for the first time a week before handing over a GPU cluster. The firm said it penalized providers that showed significant handover delays and used customer interviews to separate what was generally available from what was merely planned [3]. Its testing window closed on 15 September 2025 [3].
Hardware in scope for version 2.0 included H100, H200, B200, GB200 NVL72 and GB300 NVL72 systems on the NVIDIA side, and MI300X, MI325X and MI355X on the AMD side [3]. One of the report's general findings was that among providers deploying both vendors, the AMD offering was consistently worse, typically missing detailed monitoring, health checks with automatic remediation and working Slurm support [3].
Current ratings
The version 2.1 chart, dated April 2026, is the published rating as of mid-2026 [4][9]. Version 2.1 was an incremental update rather than a re-test: SemiAnalysis said it added providers as they became available for testing, moved four newly tested providers into Bronze, and moved several others into Unavailable as they prepared to launch [4]. SemiAnalysis described 2.1 as adding a small set of new providers rather than a full re-test [4][8].
| Tier | Providers listed in ClusterMAX 2.1, April 2026 [9] |
|---|---|
| Platinum | CoreWeave |
| Gold | Oracle, Nebius, Azure, Crusoe, FluidStack |
| Silver | Together AI, Lambda, Google Cloud, AWS, Scaleway, Cirrascale, Vultr, Voltage Park, Gcore, Firmus, GMO GPU Cloud, TensorWave |
| Bronze | CUDO Compute, Hyperstack, Shadeform, Neysa, STN, GMI, RunPod, Prime Intellect, Core42, Bitdeer, FPT AI Factory, Qubrid, Latitude.sh, Lightning AI, Verda (formerly DataCrunch), Denvr Dataworks, IBM Cloud, DigitalOcean, Atlas Cloud, Hot Aisle, Buzz HPC, Vast.ai, Radiant |
| Underperforming | Sharon AI, IREN, Hydra Host, FarmGPU, WhiteFiber, DeepInfra, dstack, PaleBlueDot.AI, Hyperbolic, GPU.NET, Akamai, Hetzner, Clore.AI, Massed Compute, Exabits, Sesterce, E2E Networks, OVHcloud, Aethir, Akash, Salad, Mithril |
| Unavailable | Nscale, HUMAIN, Corvex, Highrise, BluSky AI, ARC Compute, TELUS, Telenor, Mistral AI, Firebird, Tatra SuperCompute, Moonlite, Alibaba Cloud, MegaSpeed, BytePlus, RunSun Cloud, SK Telecom, VESSL AI, Backend, Naver, Indosat, SAKURA, Yotta, NeevCloud, QumulusAI, Boostrun |
The individual review pages on clustermax.ai still carry the version 2.0 write-ups, so Core42, Bitdeer, FPT and Radiant remain labelled Unavailable there even though the 2.1 chart and the April 2026 update place all four in Bronze [4][9].
CoreWeave has been the sole Platinum provider in all three published versions [2][9]. In version 2.0, SemiAnalysis counted 37 clouds holding a medallion rating of Bronze or better [3].
The picture in March 2025 was considerably smaller and differently ordered. Version 1.0 placed Crusoe, Together AI, Nebius, Lepton AI, Oracle and Azure in Gold; AWS, Lambda, Scaleway and Sustainable Metal Cloud in Silver; and Google Cloud, TensorWave, DataCrunch, RunPod and Denvr Dataworks in Bronze [19]. Google Cloud, which SemiAnalysis singled out at the time as being "on a Rocketship path" toward a higher tier, had moved up one tier to Silver by version 2.0 [2][20].
Commercial effect
Ratings have visible commercial weight. SemiAnalysis argued in version 2.0 that its highest-rated providers had collectively booked nearly $400 billion in remaining performance obligations since version 1.0, and that CoreWeave commands roughly 10% to 15% more per GPU-hour for managed Slurm or Kubernetes clusters than direct competitors including Nebius, FluidStack, Crusoe, Lambda and Together AI [3].
Providers use the tiers in their own marketing. CoreWeave issued a press release on 6 November 2025 announcing that it had again earned the Platinum rating and remained the industry's sole Platinum provider. It quoted Dylan Patel, described there as founder, chief executive officer and chief analyst at SemiAnalysis, saying that CoreWeave "continues to set the benchmark for AI cloud performance by demonstrating strong technical execution and operational maturity in managing large-scale AI cloud solutions" [12]. The company also maintains a standing marketing page built around the rating [13]. Third parties have gone further: the networking vendor Hedgehog published a model in 2026 that assigns pricing premiums to each tier and calculates the additional annual revenue a provider could earn by moving up one, though the figures are the vendor's own construction rather than SemiAnalysis data [18].
SemiAnalysis restricts commercial reuse. The site's legal notice states that the content, methodology and data are SemiAnalysis intellectual property and that using the ratings to create or value financial products, including indices, funds or structured notes, requires prior written consent [17].
Alongside the 2.1 update the firm published a GPU Cluster TCO Calculator and a Goodput Expense Calculator, which break cluster cost into eight components and are intended to be applied to future provider analysis [4][16].
Reception and criticism
ClusterMAX has been endorsed publicly by a long list of buyers and vendors, whose statements SemiAnalysis collects on a dedicated page. They include Peter Hoeschele of OpenAI, Santosh Janardhan of Meta, Michael Dell of Dell Technologies, Charles Liang of Supermicro, and investors including Gavin Baker of Atreides Management and Brad Gerstner of Altimeter Capital [11]. Because SemiAnalysis curates that page itself, it is best read as a promotional artifact rather than as independent evidence of adoption.
The system has also drawn substantive criticism, most of it directed at SemiAnalysis rather than at the criteria.
The sharpest published critique came from Jon Stevens, chief executive of Hot Aisle, a provider that ClusterMAX 1.0 placed in the UnderPerform tier and that later moved to Bronze. In a December 2025 essay he argued that SemiAnalysis's dual role as an independent publisher and a paid consultant to companies it covers creates a structural conflict of interest, and described what he called a "Critical-to-Consultant" pipeline in which negative coverage precedes a consulting engagement and softer follow-up coverage. He also objected specifically to the Not Recommended band added in version 2.0, arguing that a "kinder, alternative approach would have been to end the list at the Bronze category" [14]. Stevens disclosed his own interest as an AMD-only neocloud operator, and his essay ranges well beyond ClusterMAX into unrelated disputes about SemiAnalysis's coverage of AMD, a 2025 compromise of the firm's X account, and attribution practices.
A separate line of criticism concerns whether a pre-announced, provider-cooperative test can be trusted. On the SemiWiki forum in November 2025, one participant argued that because a provider knows which account belongs to the benchmark, it becomes trivial to flag that account for extra performance, elevated priority and availability [15]. SemiAnalysis's own account of the process partly concedes the point: it disclosed that providers rushed to change their offerings after receiving the criteria list, and said it penalized those whose handover slipped as a result [3].
Providers have also disputed the criteria themselves. SemiAnalysis reported that Cirrascale, rated Silver, considers requirements around software orchestration such as Kubernetes irrelevant to its bare-metal customers, a position SemiAnalysis rejected on the basis of what Cirrascale's own customers told it [3]. The firm reported that IBM deactivated its account and blocked further sign-ups during the testing period, and that IBM was rated Bronze pending a proper test [3]. In a post reproduced in full on SemiWiki, SemiAnalysis went further and claimed that IBM's analyst relations team had said it "would only be willing to participate if guaranteed a silver rating or higher"; the forum's founder reported the following day that SemiAnalysis had deleted a critical comment Jon Stevens had left on that post. The same thread carried the counter-argument that benchmarking a service after being denied permission raises problems of its own [15].
On the question of whether ratings can be bought, the version 1.0 report carried an explicit statement that "No part of SemiAnalysis's compensation by our clients was, is, or will be directly or indirectly related to the specific tiering, ratings or comments expressed" [2]. The version 2.0 disclaimer covers trademarks and intellectual property but does not repeat that line, and the clustermax.ai legal notice invites licensing and partnership inquiries to a ClusterMAX address [3][17]. No published evidence has emerged that any provider paid for a specific tier.
Related work
ClusterMAX sits alongside InferenceMAX, a separate SemiAnalysis benchmark launched in October 2025 that continuously measures inference throughput and cost across models, frameworks and accelerators [21]. Where ClusterMAX rates the operator of a cluster, InferenceMAX measures the silicon and software running on it. Other reference points in the same market include Artificial Analysis, which benchmarks model and inference-endpoint performance, and MLPerf, whose training and inference suites measure hardware and software stacks under audited rules rather than rating commercial service quality. Traditional analyst firms cover the same buyers with different instruments: SemiAnalysis has publicly contrasted its own placements with Gartner's, noting that Gartner ranked IBM above Nebius and CoreWeave [15].
See also
References
- ^ClusterMAX, "ClusterMAX Overview: GPU Cloud Rating Methodology." clustermax.ai/overview
- ^SemiAnalysis, "The GPU Cloud ClusterMAX Rating System | How to Rent GPUs," 26 March 2025. newsletter.semianalysis.com/...em-how-to-rent-gpus
- ^SemiAnalysis, "ClusterMAX 2.0: The Industry Standard GPU Cloud Rating System," 6 November 2025. newsletter.semianalysis.com/...e-industry-standard
- ^ClusterMAX, "ClusterMAX 2.1: Updated Tier Rankings and New Entrants." clustermax.ai/v2.1
- ^ClusterMAX, "Evaluation Criteria: 10 Dimensions of GPU Cloud Quality." clustermax.ai/criteria
- ^ClusterMAX, "ClusterMAX 2.0: GPU Cloud Industry Standard (2025)." clustermax.ai/v2
- ^ClusterMAX, "ClusterMAX 1.0." clustermax.ai/v1
- ^SemiAnalysis, "How Much Do GPU Clusters Really Cost?", 20 April 2026. newsletter.semianalysis.com/...lusters-really-cost
- ^SemiAnalysis, "SemiAnalysis GPU Cloud ClusterMAX Rating April 2026" (ClusterMAX 2.1 rankings chart). clustermax.ai/...neocloud-ranking-v2.1.jpg
- ^ClusterMAX, "Best GPU Clouds 2026: Top 85 GPU Cloud Providers Ranked, ClusterMAX 2.0 + 2.1." clustermax.ai/cloudreview
- ^ClusterMAX, "Industry Quotes: Researchers, Engineers, Investors on Neoclouds." clustermax.ai/quotes
- ^CoreWeave, "CoreWeave Achieves SemiAnalysis' Platinum ClusterMAX Rating for the Second Consecutive Ranking, Remaining the Industry's Sole Platinum Provider," 6 November 2025 (Business Wire). markets.financialcontent.com/...-platinum-provider
- ^CoreWeave, "CoreWeave ranks as #1 AI Cloud, Backed by SemiAnalysis's Platinum ClusterMAX Rating." coreweave.com/...iss-platinum-clustermax-tm-rating
- ^Jon Stevens, "Influence as a Service: SemiAnalysis Under the Microscope," 2 December 2025. jon4hotaisle.substack.com/...-service-semianalysis
- ^SemiWiki forum thread, "SemiAnalysis Gross Mafia Tactics?", started 21 November 2025. semiwiki.com/...analysis-gross-mafia-tactics.24073
- ^ClusterMAX, "GPU Cloud TCO and Goodput Calculator." clustermax.ai/tco
- ^ClusterMAX, "GPU Cloud ClusterMAX Rating and Ranking System." clustermax.ai
- ^Marc Austin, Hedgehog, "The Sell Math: How ClusterMAX 2.0 Ratings Translate Into Revenue," updated 8 June 2026. hedgehog.cloud/...0-ratings-translate-into-revenue
- ^SemiAnalysis, "SemiAnalysis GPU Cloud ClusterMAX Rating March 2025" (ClusterMAX 1.0 rankings chart). clustermax.ai/...neocloud-ranking-v1.webp
- ^SemiAnalysis, "SemiAnalysis GPU Cloud ClusterMAX Rating November 2025" (ClusterMAX 2.0 rankings chart). clustermax.ai/...neocloud-ranking-v2.jpg
- ^SemiAnalysis, "InferenceMAX: Open Source Inference Benchmarking," October 2025. newsletter.semianalysis.com/...en-source-inference
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.
v1 · 3,111 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Reviewer note: All 89 provider tier placements independently re-derived from the April 2026 rating chart, SemiAnalysis's per-provider review pages and the April 2026 newsletter on 2026-08-01; tier definitions and the ten criteria checked word for word.