Sustainable Metal Cloud

What makes Sustainable Metal Cloud different

Sustainable Metal Cloud (SMC) is built on a sustainability-first infrastructure model centered around liquid cooling technology at massive scale. Unlike traditional GPU clouds that rely on air cooling, SMC’s data centers achieve up to 48% lower CO₂ emissions when training AI models compared to legacy air-cooled GPU instances in equivalent PUE 1.30 facilities. This is not merely a marketing claim—the efficiency gains translate directly to lower power costs and improved reliability without sacrificing performance.

The provider targets enterprises and AI platforms seeking dedicated bare-metal GPU clusters rather than shared virtualized instances. SMC’s offerings range from startup-friendly self-service metal instances to wholesale hyperscale deployments of 32,000+ GPUs in a single cluster, customized for private use or new public availability zones. All clusters are co-designed with NVIDIA and leverage full-rail RDMA-enabled InfiniBand with BlueField-3 DPUs for high-speed GPU-to-GPU communication. This architecture is especially valuable for distributed AI training workloads that demand sub-millisecond latency and high throughput.

SMC is an NVIDIA Preferred CSP Partner and has been ranked alongside AWS and above Google Cloud by SemiAnalysis for GPU cloud performance across real-world AI workloads.

Pricing model

SMC publishes transparent, itemized pricing:

• Fast File Storage (Weka): $0.065/GB per month—optimized for high IOPS and low-latency training I/O
• Object Storage (S3-compatible): $0.025/GB per month—suitable for model checkpoints and archival
• GPU instances: Hourly usage-based pricing (exact per-GPU rates not listed on public pages but available via the self-service console)

The cost advantage stems from SMC’s energy efficiency. By reducing data center cooling overhead by up to 48%, the provider can undercut typical cloud GPU pricing without compromising margin, making it particularly competitive for long-running training jobs and large-scale inference clusters.

When it fits

• AI model training at scale: Teams training large language models or vision models on H100/H200 clusters with distributed training frameworks (PyTorch, Megatron-LM, etc.)
• High-performance inference: Applications requiring dedicated GPU clusters with guaranteed SLA (99.50% uptime) and 24/7 critical support
• Colocation expansion: Enterprises adding GPU capacity to existing on-premise infrastructure without the CapEx and lead time of private procurement
• Hyperscale deployments: CSPs and large platforms needing 10,000–32,000 GPU clusters with custom networking and rack-level control
• Carbon-conscious organizations: Teams with sustainability mandates seeking measurably lower-emission cloud infrastructure

When it doesn’t

• Multi-cloud or region-agnostic workloads: Currently limited to Singapore and India; not suitable for teams requiring global redundancy across multiple cloud providers.
• Small-scale or bursty GPU usage: SMC focuses on dedicated metal; teams needing fractional GPUs or elastic on-demand capacity may find shared-instance offerings from larger hyperscalers more cost-effective.

Inclusion criteria

✅ Transparent pricing: Published per-GB storage costs and hourly GPU pricing via public console
✅ Self-service signup: Self-service console login available; GPU instances available for immediate provisioning
✅ Public SLA and status page: 99.50% uptime SLA published; critical support available 24/7, 365 days per year