Which GPU Clouds Are Actually Good? | ClusterMAX 3.0
ClusterMax 3.0 rates GPU clouds on managed-cluster quality, finds neocloud security and compute availability bad, and shows labs, Nvidia and chip makers reshaping who supplies and buys compute.
Watch on YouTube ↗Charlie's summary
- ▶ 1:34
ClusterMax replaced hands-on benchmarking with an automated, agent-run test suite and now covers 323 providers.
Jordan says ClusterMax 1 was manual benchmark scripts, before coding agents. Over a year and a half the team built a repo of industry-standard and custom benchmarks, plus reliability and burn-in tests, that coding agents launch and fix autonomously. Coverage grew from about 120 providers tracked to 323, and the ranking adds Kubernetes alongside Slurm.
- ▶ 5:53
Most neoclouds are terrible at security, often through basic failures like years-old software and unconfigured network keys.
Dylan says testers could see other tenants' jobs, data and storage, and in one case a country's national intelligence workloads. Jordan says the checks are simple: outdated drivers and software with documented CVEs, and unconfigured InfiniBand keys. “You don't need a Frontier model” to exploit them. The team published a free CLI, CMAX, that checks software versions on your cluster.
- ▶ 9:33
The ClusterMax bar is “latest and greatest”: clouds that skip GB300 NVL72 or fix problems slowly get downgraded.
The bar rises each generation, so downgrades outnumber upgrades. Google rose steadily to gold. Jordan says GB300 NVL72 brings a new network, ARM CPUs, 800 gig networking and mandatory liquid cooling, so skipping it is a risk when Vera Rubin arrives. Providers that deploy it and fix issues in an hour, not two weeks, move up. Reliability is the top criterion for big customers.
- ▶ 12:09
Renting compute is harder than ever because profitable inference providers now compete with labs for GPUs.
Dylan says inference companies like Baseten, Fireworks and Together, and small ones like Morph, run around 60% gross margins on open models, so anyone with a GPU can make money. Labs once wanted 8k-GPU deals and now rent as few as 1,000, while Modal profits on four nodes. Neolabs now compete against buyers who earn money from their GPUs. Even renting single spot nodes is hard.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.