GMO Cloud (GMO Internet Group) review

Egress is free, no cap, which is one of the strongest reasons to pick this service for training. Pulling checkpoints, datasets or finished models to another host adds no transfer charge. That changes the arithmetic for anyone training on a cluster and serving elsewhere.

The lineup covers NVIDIA HGX B300 and NVIDIA HGX H200, so the catalogue is deliberately short. Both are current NVIDIA data-center flagships suited to large training and inference jobs. If a project needs a different architecture, including AMD, this service will not cover it.

Capacity sits in 1 location. That concentration makes sense for workloads that stay in one place, such as sustained training runs. It offers no choice if you need a second region, a second market, or a specific data residency.

Support runs on email only. A written reply suits billing questions and account setup. A cluster down in the middle of a training run is a harder fit, since there is no phone line or live chat to escalate to.

Nodes are connected by NVIDIA Spectrum-X Ethernet. That matters for jobs that cannot scale well over ordinary Ethernet, where inter-card bandwidth becomes the bottleneck. For single-card work, the interconnect is less relevant to your bill or your runtime.

The service belongs to GMO Internet, Inc., the cloud and internet arm of GMO Internet Group. That parent gives the operation a corporate owner rather than a small independent team. It is worth weighing if you sign longer contracts or need formal billing.

GMO Cloud runs NVIDIA HGX B300 and NVIDIA HGX H200, a tight NVIDIA-only range. SberCloud lists B300, H100 and A100, and Salad covers NVIDIA RTX 3060, NVIDIA RTX 3070 and NVIDIA RTX 3070 Ti with more consumer and community cards. Choose GMO Cloud when you want the top end rather than variety.