TensorBay builds dedicated bare-metal AI clusters — no hypervisor, no noisy neighbors, no oversubscribed fabric. Every cluster ships with a published acceptance test: NCCL all-reduce numbers you verify before you pay.
On the factory floor right now
Zero preemptions in six months. The acceptance report matched what we measured on day one — that's never happened with any other provider.
01Hardware
Every SKU below is a whole machine — full HBM, full NVLink, full NIC bandwidth. Reserved rates published for 12- and 36-month terms; capacity is allocated in whole nodes. If a price says "on request", it's supply, not secrecy.
Reserved · yearly · from
$39/node-hr
Reserved · yearly · from
$24/node-hr
Reserved · yearly · from
$15/node-hr
Reserved · yearly · from
$45.00/node-hr
Reserved · yearly
On request
Reserved · yearly
On request
Reserved · yearly
On request
02Why bare metal
Virtualized GPU clouds skim 3–8% of your FLOPs before your job starts and add jitter your all-reduce can feel. TensorBay hands you the machine — firmware to fabric.
F-01
Jobs run on the metal. Measured 5–7% higher sustained MFU versus virtualized instances on identical HGX H100 hardware — and zero p99 latency spikes from neighbor VMs.
F-02
Your cluster is your hardware: dedicated nodes, dedicated leaf switches, dedicated storage lanes. Isolation you can trace on a wiring diagram, not a compliance PDF.
F-03
Every cluster passes a 72-hour burn-in: NCCL all-reduce sweeps, HBM bandwidth checks, thermals under sustained load. You get the report. Sign-off is yours, not ours.
F-04
Custom kernels, your own drivers, NIC firmware tuning, RDMA settings. Bring Slurm, Kubernetes, or bare SSH — we pre-stage images but never lock the BIOS.
F-05
Every reserved cluster includes ≥2% hot-spare nodes on the same fabric. Node failure means a reschedule, not a support ticket: 15-minute replacement SLA, automated.
F-06
Per-minute billing, published rates, zero ingress, 20 TB of egress included per node-month, no "IB surcharge" surprise. The number on this page is the number on the invoice.
03The factory
We don't broker other people's GPUs. TensorBay designs, builds, and operates its own halls — liquid-cooled, high-density, and instrumented down to per-GPU power draw. When output is measured in tokens, the plant matters.
1 GW
Contracted across 5 sites — US, Philippines, Brazil, Norway and Switzerland — anchored on hydro-rich grids.
PUE 1.12
Direct-to-chip liquid cooling on 130 kW racks. Every watt saved is a watt on silicon.
24,000+ GPUs
New capacity commissioned in 6-week blocks — 72-hour burn-in included, report published.
97.4%
Cluster goodput, trailing 90 days. 0.4% monthly node failure rate, hot-swapped from in-rack spares.
Factories publish yield. So do we — live, at status.tensorbay.com
04Fabric & storage
Training throughput dies in the network and the filesystem long before it dies in the GPU. Ours are specified like the rest of the machine — exactly, in writing.
| Layer | Specification |
|---|---|
| Training fabric — Hopper | 8×400 Gb/s InfiniBand NDR per node (3.2 Tbps), rail-optimized non-blocking fat-tree, SHARP in-network reduction |
| Training fabric — Blackwell | Quantum-X800 InfiniBand XDR, 800G per link; GB200 racks add a 72-GPU coherent NVLink 5 domain at 1.8 TB/s per GPU |
| Oversubscription | 1:1, all tiers. No blended east-west fabric, no shared spine with other tenants |
| Local scratch | 30 TB NVMe Gen5 per node, ~55 GB/s read — checkpoint staging without touching the network |
| Parallel filesystem | Managed WEKA, dedicated per cluster: up to 720 GB/s aggregate read per pod, POSIX + GPUDirect Storage |
| Object storage | S3-compatible, NVMe-cached, co-located with compute — $0.055/GB-month hot tier |
| Data transfer | Ingress $0 · 20 TB/node-month egress included, then $1.25/TB · inter-node $0. Free 100G Direct Connect on yearly reservations |
| Front-end network | Dual 100 GbE per node, DDoS-protected, BYO-IP supported |
05Pricing
Reserved capacity, priced in the open: monthly and yearly terms with node-replacement and goodput SLAs written into the contract, not the FAQ. On-demand and spot are on the roadmap — reserve is how you buy today.
| Per node-hour | Monthly term | Yearly term |
|---|---|---|
| HGX B200 | from $56 | from $39 |
| HGX H200 | from $34 | from $24 |
| HGX H100 | from $22 | from $15 |
| HGX B300 | from $64 | from $45.00 |
| GB200 NVL72 | — | On request |
| GB300 NVL72 | — | On request |
| VR200 NVL72 | — | Early buy · delivery Mar 2027 |
| SLA | 99.9% monthly + 15-min node replacement + goodput credits | |
| Scale | 128–10,000+ GPUs · dedicated fabric per tenant | |
| Billing | Per-minute metering inside the reservation · $0 ingress | |
Prices per 8-GPU HGX node-hour; rack-scale systems quoted per rack. Prices exclude applicable taxes. Managed Slurm and Kubernetes control planes: $0. Storage and Direct Connect priced above; nothing else exists to charge for. Long-term and multi-year contracts: on request.
06Use cases
512–10K+ GPU reserved clusters on dedicated XDR fabric with in-rack spares and goodput SLAs. When a 6-week run costs seven figures, 97% goodput versus 91% is the whole margin.
→ Reserved B200 / GB200 / GB300 podsWhole 8-GPU nodes on short reserved blocks — per-minute metering inside the reservation, WEKA-backed datasets that persist between runs. Provisioned in under 4 minutes, returned without a ticket.
→ H100 / H200 nodes from $15/hrMemory-dense H200 and B300 nodes for long-context and 70B+ serving; bare-metal p99s with no virtualization jitter and 20 TB of included egress per node, every month.
→ H200 from $24/hr · B300 from $45/hr07Security
Single-tenant hardware is the security model — everything else is attestation of it. We publish report scope because an acronym without one is marketing.
Scope: bare-metal compute, fabric, and managed storage
Certified — all datacenter sites
BAA available on dedicated clusters
EU region with in-country data residency
08FAQ
Short answers, same numbers as the rest of this page. If something here matters to your contract, it goes in the contract.
Rates are published per 8-GPU node-hour on a yearly term: HGX H100 from $15, HGX H200 from $24, HGX B200 from $39, HGX B300 from $45. Monthly terms are higher. GB200, GB300 and VR200 NVL72 racks are quoted per rack — that is a supply constraint, not a pricing one.
No. Capacity is allocated in whole 8-GPU nodes, or whole NVL72 racks for rack-scale systems. Slicing a node means a hypervisor, and the hypervisor is the thing we removed.
Not today. Reserved monthly and yearly terms are how you buy; on-demand and spot are on the roadmap. Inside a reservation, metering is per minute.
Virtualized GPU clouds skim 3–8% of your FLOPs before the job starts and add jitter your all-reduce can feel. On identical HGX H100 hardware we measure 5–7% higher sustained MFU than virtualized instances, and no p99 latency spikes from neighbor VMs.
Every cluster runs a 72-hour burn-in before handover: NCCL all-reduce sweeps, HBM bandwidth checks and thermals under sustained load. You get the report and you sign off — not us. The quote tells you the exact report format your cluster must pass.
Every reserved cluster carries at least 2% hot-spare nodes on the same fabric. Replacement is automated with a 15-minute SLA, so a failure is a reschedule rather than a support ticket. Node failure runs about 0.4% monthly and cluster goodput has been 97.4% over the trailing 90 days.
No. Nodes, leaf switches and storage lanes are physically dedicated per tenant at 1:1 oversubscription on every tier — there is no blended east-west fabric and no shared spine. Hopper pods run 8×400 Gb/s NDR per node; Blackwell pods run Quantum-X800 XDR at 800G per link.
Ingress is $0, inter-node traffic is $0, and every node includes 20 TB of egress per month, then $1.25/TB. Yearly reservations include a 100G Direct Connect. Managed Slurm and Kubernetes control planes are $0.
1 GW is contracted across five sites — the US, the Philippines, Brazil, Norway and Switzerland — anchored on hydro-rich grids. The EU region supports in-country data residency for GDPR, and a HIPAA BAA is available on dedicated clusters.
Send GPU count, interconnect and timeline and you get a firm quote, a delivery date and the acceptance-test format within 48 hours. Nodes inside an existing reservation provision in under 4 minutes; new capacity is commissioned in 6-week blocks with burn-in included.
09Commission your cluster
Tell us GPU count, interconnect, and timeline — get a firm quote, a delivery date, and the exact burn-in report format your cluster must pass, within 48 hours.
20 TB egress included per nodePer-minute billingLive goodput at status.tensorbay.com
