The world's largest high-end distributed GPU network: 5,000+ B200s, H200s, and H100s across 200+ locations in 90+ countries. Bare metal. B300 available. Per-second billing. Authorized partner.
NVIDIA H100, H200, B200 — bare metal, on-demand, available now
| Specification | H100 | H200 | B200 | B300 |
|---|---|---|---|---|
| Architecture | Hopper | Hopper | Blackwell | Blackwell |
| Memory | 80 GB HBM2e | 141 GB HBM3e | 192 GB HBM3e | 3TB RAM / 30TB NVMe |
| Bandwidth | 3.35 TB/s | 4.8 TB/s | 8 TB/s | 6.4 Tbps Infiniband |
| Power (TDP) | 700W | 50% less for LLM inference | 1,000W | Next-gen efficiency |
| Inference Speed | 22,290 tok/s | 31,712 tok/s (+42%) | ~45,000 tok/s | ~60,000 tok/s est. |
| Precision | FP8 / FP16 / INT8 | FP8 / FP16 / INT8 | FP4 / FP6 / FP8 | FP4 / FP6 / FP8 |
| Interconnect | NVLink 4.0 | NVLink 4.0 | NVLink 5.0 | NVLink 5.0 |
| Best For | Training + inference | LLMs, large datasets | Gen AI at scale | Next-gen AI training |
| Availability | ✅ On-demand | ✅ On-demand | ✅ On-demand | 🆕 Oct 2026 |
Decentralized GPU cloud that outperforms traditional providers on cost, availability, and flexibility
GPU containers in 93 countries across 200+ locations. Deploy workloads at the edge or in major data center hubs — wherever your latency requirements demand.
Up to 80% less than AWS, Azure, and GCP for equivalent GPU compute. No hidden fees, no egress charges, no markup on data transfer.
Per-second billing. Scale up during training runs, scale down during idle periods. No minimum commitment. No lock-in. Pay only for what you use.
99.31% uptime across the network. 150+ enterprise clients. 1.5 billion compute hours delivered in 2025 alone. Battle-tested infrastructure.
Direct hardware access — no hypervisor, no virtualization overhead. H100, H200, B200, B300 available. Every FLOP counts.
Spec Trading runs managed AI agent infrastructure in Japan & USA. Fully configured, monitored, and optimized for your workloads.
From startups training foundation models to enterprises running inference at scale
TensorOpera trained a 750-million parameter model in 30 days on Aethir infrastructure — at a fraction of traditional cloud costs.
750M params · 30 daysLlama 2 70B runs nearly 2× faster on H200 vs H100. Serve millions of tokens with sub-100ms latency on globally distributed nodes.
2× faster · 31K tok/sMeta acquired 350,000 H100 GPUs. Digi Tech secured 5,120 H200s. The demand is real — Aethir makes it accessible without the CapEx.
350K+ GPUs deployedSpec Trading runs an 84-agent multi-agent system on Aethir compute — 24/7 trading, market scanning, and strategy execution.
84 agents · 24/7Actual Aethir enterprise rates vs traditional cloud. H100 from $1.55/hr. B200 from $2.91/hr.
| GPU Model | Per GPU/Hour | Server Spec | Best Term | Location | Status |
|---|---|---|---|---|---|
| B300 (Blackwell) | $3.25/hr | 8× B300 SXM, 3TB RAM, 30TB NVMe, 6.4 Tbps Infiniband, 100Gbps | 3yr: $3.25/hr | Washington, USA | 🆕 Oct 2026 |
| B200 (Blackwell) | $2.91/hr | 8× B200 SXM, 2TB RAM, 8× 3.84TB NVMe, 3.2Tbps Infiniband | 12mo: $2.86/hr | Texas, USA | ✅ Available |
| H200 (Hopper) | $1.70/hr | 8× H200 SXM 141GB, 3TB DDR5, 15TB NVMe, 3.2Tbps RoCE v2 | 12mo: $1.70/hr | Iceland · Atlanta, USA | ✅ Available |
| H100 (Hopper) | $1.55/hr | 8× H100 SXM, 1TB RAM, 19.2TB NVMe, 3.2Tbps Infiniband | 12mo: $1.55/hr | Seattle · SLC · Virginia | ✅ Available |
| A100 80GB | $1.15/hr | DGX A100, 2TB RAM, 8TB NVMe, Infiniband/Ethernet | Month-to-month | Midwest, USA | ✅ Available |
| L40S | $1.00/hr | 8× L40S PCIe, 96-core AMD, 1.5TB RAM, 96TB NVMe | 12mo: $1.00/hr | Atlanta, USA | ✅ Available |
| RTX 5090 | $0.48/hr | 8× RTX 5090, 512GB RAM, 30TB NVMe, 10Gbps | 12mo: $0.48/hr | Nebraska, USA | ✅ Available |
| RTX 4090 | $0.34/hr | 8× RTX 4090, 512GB RAM, 30TB NVMe, 10Gbps | 12mo: $0.34/hr | Nebraska, USA | ✅ Available |
Hourly H100 Cost Comparison (USD)
* H100 on-demand pricing. Prices shown are for 12-month terms where applicable. M2M available at slightly higher rates.
"If you're using AWS capacity blocks, you could be paying 3X more than necessary. Aethir gives us stable, high-performance infrastructure at a fraction of the cost, plus the fastest NFS we've used."
"Scary thing about working with GCP or AWS — we never know if our bill is going to randomly be two times higher than last month."
Enterprise-grade Service Level Agreements — not best-effort support
| Priority | Service Hours | Description | Response | Resolution | Updates |
|---|---|---|---|---|---|
| P0 | 24/7 | Service Outage | 15 min | Up to 4 hours | Every 15 min |
| P1 | 12/5* | Significant service degradation | 15 min | Up to 8 hours | Every 30 min |
| P2 | 12/5* | Service reconfiguration / requirement change | 15 min | Up to 24 hours | Every 2 hours |
| P3 | 8/5* | Information or feature request | 15 min | Up to 72 hours | Every 12 hours |
* Extended hours coverage available for enterprise contracts
Spec Trading is an authorized Aethir compute partner. We don't just provision GPUs — we deploy, manage, and optimize your entire AI infrastructure.
From a single GPU container to a 100-node cluster running autonomous agents, we handle the infrastructure so you can focus on building.
Authorized Partner
Tell us about your compute needs. We'll design a solution within 24 hours — no obligation, no sales pitch.
London-based team · Response within 4 hours during UK business hours · Serious inquiries only