Deal overview
Dedicated NVIDIA B300 capacity on pure NVIDIA reference architecture in an operational Texas facility. The entry point is a 2,048-GPU dedicated cluster on 5 MW, online about six months after contract; the site then scales in four phases to 18,432 liquid-cooled GPUs on 36 MW of critical IT load (48 MW gross), with later phases landing roughly 8, 10 and 12 months from contract. Pricing is a firm $4.65/GPU-hr on a 3-year minimum term with a deposit of 25% of total contract value — about $6.95M per month for the entry cluster at 730 hours.
Supermicro servers built to the NVIDIA reference design mean no custom cooling, power or interconnect risk, and the buyer chooses the service model: bare metal with full control of firmware, drivers, scheduling and storage, or partially or fully managed. Capacity is single-tenant, or multi-tenant with physical and logical separation, with FedRAMP Moderate designed in for proprietary models and sensitive data. The operating team brings more than 3.5 billion hours of liquid-cooled HPC and secure-deployment experience. The public tile shows the headline specification; the complete data room — site details, full specs, exact location, counterparty and term sheet — is shared under mutual NDA once both sides agree to connect.
Public specification
- GPUs
- NVIDIA B300 · liquid-to-chip cooled
- Servers
- Supermicro · built to NVIDIA reference architecture
- Entry cluster
- 2,048 GPUs · 5 MW · 5 PB NVMe · online ~6 months from contract
- Phase 2
- 4,096 GPUs · 8 MW · 10 PB NVMe · +8 months
- Phase 3
- 4,096 GPUs · 8 MW · 10 PB NVMe · +10 months
- Phase 4
- 8,192 GPUs · 15 MW · 20 PB NVMe · +12 months
- Full site
- 18,432 GPUs · 36 MW critical IT · 45 PB · 48 MW gross, liquid cooled
- Fabric
- 800 Gb/s · topology can be specified on bare metal
- Storage
- DDN AI400X3M with Infinia · NVMe · up to 45 PB
- Service models
- Bare metal, partially managed or fully managed
- Tenancy
- Single-tenant, or multi-tenant with physical and logical separation
- Software
- Your Kubernetes, Slurm and orchestration
- Facility
- Tier III · concurrently maintainable · operational today
- Connectivity
- Multiple on-net national carriers and regional fiber · DWDM live · diverse multi-path routing · four fiber vaults, multiple meet-me rooms
- Compliance
- FedRAMP Moderate designed in · FISMA, NIST, ISO 20000 ready
- Rate
- $4.65/GPU-hr · firm
- Term
- 3-year minimum
- Deposit
- 25% of total contract value
- Branding
- On-site branding available with a take of the full 36 MW
- Location
- Texas, USA
How connecting works
- Submit an inquiry on this page — free, no account needed.
- Ax3 qualifies and routes it within 24 hours.
- Warm introduction with the full spec sheet and exact pricing; identities revealed once both sides agree.
