GMO GPU Cloud
GPU cloud for AI/ML in Japan with bare metal servers, InfiniBand, and data sovereignty.
What makes GMO GPU Cloud different
GMO GPU Cloud is purpose-built for high-performance compute workloads that demand low-latency, high-throughput interconnection between compute and storage. Unlike general-purpose cloud providers, it combines bare metal Tesla V100 GPUs with 100 Gb/s InfiniBand EDR networking and Lustre distributed filesystem—infrastructure traditionally found in HPC supercomputing centers, not public clouds. This architectural choice eliminates the noisy-neighbor problem and unpredictable latency inherent to virtualized GPU offerings.
The service reflects GMO Internet’s internal use of GPUs across its ad-tech, FX trading, and e-commerce divisions. Rather than forcing customers to configure their own HPC stack, GMO provides a pre-integrated environment with CUDA, OpenCL, and container engines (Singularity/Docker) ready to use. For organizations in Japan with strict data residency requirements, this is the only option—data never leaves Japan.
Pricing model
GMO GPU Cloud offered three fixed monthly plans at launch (March 2019):
| Plan | GPUs | Initial Cost | Monthly Cost | GPU Memory |
|---|---|---|---|---|
| 1 GPU | 1x Tesla V100 | ¥882,000 | ¥97,000 | 16 GB |
| 2 GPU | 2x Tesla V100 | ¥1,560,000 | ¥142,000 | 32 GB |
| 4 GPU | 4x Tesla V100 | ¥2,880,000 | ¥230,000 | 64 GB |
Initial costs could be paid in installments. The pricing reflects bare metal hardware and managed HPC-grade interconnect; it is significantly higher than hyperscaler on-demand GPU pricing but includes dedicated resources and low-latency networking at no extra charge. The original roadmap mentioned a future hourly billing option, though current status is unclear.
When it fits
- AI/ML model training and inference requiring multi-GPU synchronization with sub-millisecond latency
- Molecular dynamics simulations and drug discovery leveraging Lustre for high-throughput data access
- Scientific computing (genome analysis, physics simulations) where file distribution and HPC toolchains are critical
- Japanese enterprises needing strict data sovereignty and zero overseas data transfer
- Organizations migrating from on-premises HPC who want managed infrastructure without capital expense
When it doesn’t
- Batch inference or inference-only workloads that don’t require tight multi-GPU synchronization (hyperscaler spot GPUs are cheaper)
- Workloads requiring GPUs outside Japan or multi-region failover
- Development and testing where hourly billing would be preferable to fixed monthly commitments
Inclusion criteria
GMO GPU Cloud meets all three inclusion criteria:
- Transparent pricing: Monthly plans with explicit costs published in the March 2019 press release
- Self-service signup: Service available via https://gpu.cloud.gmo.jp with same-day provisioning
- Public SLA/status page: Status monitoring available (specific SLA terms should be verified on the service portal)