ClusterMAX 3.0 ranks 77 GPU clouds on reliability, security and performance

ClusterMAX is SemiAnalysis's rating of companies that rent out managed GPU clusters: the chips plus the job scheduler, storage, monitoring and failure handling, run by the provider for a customer that has no infrastructure team of its own. ClusterMAX 3.0 rates 77 providers and raises the bar, so only 19 earn a medal tier from Platinum down to Bronze. Nebius joins CoreWeave in Platinum, Google Cloud joins Oracle in Gold, Microsoft's Azure moves to Silver and Crusoe drops to Bronze. A new Participation Ribbon tier below Bronze holds 15 providers that, in SemiAnalysis's view, do the bare minimum.

The ratings lean hardest on what happens when hardware breaks. SemiAnalysis asked each provider for 32 current-generation GPUs with fast networking, storage and a monitoring dashboard, then tested in three phases: a roughly 15-minute configuration audit, released free on GitHub; performance tests from raw chip math up to real training and inference jobs; and reliability. The reliability phase runs an eight-hour burn-in that stresses GPUs and network together, then deliberately injects GPU faults and times how long the provider's own health checks take to notice, pull the node out of service and return it. SemiAnalysis calls this its most critical test, because the providers that draw reliability complaints from customers usually lack health checks, monitoring and automatic repair, while basic chip benchmarks rarely separate anyone. Hands-on testing is only part of the rating; support, pricing, availability and performance at scale come from other research, including interviews with well over 200 end users.

The tiers answer a specific buyer's question. SemiAnalysis says frontier labs such as OpenAI and Anthropic run their own stacks on bare metal, and it leaves inference endpoints and post-training services out of scope. It also gave extra leeway to providers that delivered Grace Blackwell rack systems. The Not Recommended, Unavailable group is not a test result: it includes providers that declined to test or had no spare capacity, such as Fluidstack, which shifted to bare-metal TPU deployments. Readers should also know that SemiAnalysis sells technical assessments built on the same tooling and promotes its own contract templates in the report. Security got extra attention this round, following SemiAnalysis's earlier preview, Most Neoclouds Suck At Security, and Ilya Sutskever's warning that rogue agents could target poorly secured neoclouds.