One block, built the same way every time.
A 256 GPU AI factory block on NVIDIA B300: dedicated hardware, one fabric, bare metal orchestration and low latency paths to US hubs. It is the standard we replicate at every site.
The block
Every block ships with the same topology, so a design that works on one works on all of them.
Compute
NVIDIA B300 on HGX. 32 nodes per block, 8 GPUs per NVLink domain. Single tenant hardware for every reserved cluster.
Fabric
NVIDIA Spectrum X Ethernet, ~800 Gb/s class inside the POP. Rail aligned scale out sized for the full block.
Storage
Local NVMe on every node, plus parallel file and AI object tiers sized to the cluster you reserve.
Orchestration
Managed Kubernetes on bare metal, GPU ready on day one. Your images, your schedulers, your topology awareness.
Front end and management
A separate front end network for ingest and serving, and out of band management on every node.
Scale
Blocks split as 128 + 128 or 64 × 4, and two blocks make a full NVIDIA reference architecture scalable unit.
Spec sheet
The node and the block, line by line. Pricing is per contract.
NVIDIA HGX B300
| Accelerators | 8 × NVIDIA B300 |
|---|---|
| GPU memory | 288 GB HBM3e per GPU |
| Per node | ~2.3 TB HBM3e |
| NVLink domain | 8 GPUs |
| Scale out | Spectrum X, ~800 Gb/s class |
| Storage | Local NVMe |
| Use | Training and inference |
| Pricing | Per contract |
256 GPU AI factory
| Nodes | 32 × HGX B300 |
|---|---|
| GPUs | 256 |
| GPU memory | ~74 TB HBM3e |
| Fabric | NVIDIA Spectrum X Ethernet, rail aligned |
| Orchestration | Managed Kubernetes, bare metal |
| Terms | 12 to 36 months |
| First block | Live Q2 2027 |
| Roadmap | 1,024 GPUs in 12 months |
Software
A managed platform from bare metal to model, so your team starts on workloads, not on drivers.
Provisioning
Bare metal provisioning and lifecycle are automated, from first boot to firmware and OS updates.
Kubernetes
Conformant Kubernetes clusters per customer on bare metal, with upgrades managed for you.
GPU ready
NVIDIA GPU and network operators, drivers and GPU partitioning in place on day one.
Isolation
Multi tenant isolation and policy enforcement across clusters, so every customer runs on its own hardware.
Self service
A portal and API to create clusters, scale nodes and track usage and performance.
Training and inference
Ready environments for training and serving, with autoscaling and routing to healthy nodes.
How a block is delivered
From term sheet to production, the same five steps at every site.
- 01
Term sheet
Size, term and a delivery window, in writing.
- 02
Topology design
Rack plan, fabric, storage tier and network handoff, designed with your team.
- 03
Burn in
Every node stress tested and the fabric validated before anything is handed over.
- 04
Handover
Credentials, runbooks and named contacts on day one.
- 05
Operate
Monitoring, hardware replacement and maintenance windows agreed in the contract.
Security and compliance
What we commit to, stated plainly.
Commitments
- ✓Single tenant hardware for every reserved cluster
- ✓Facility certifications from the colocation operator
- ✓dense² SOC 2 program in progress
- ✓Every customer screened against US export control and sanctions rules
Disclosures
- →Residency is Mexico, disclosed in every document you sign
- →Not a US data residency region, so not for FedRAMP or US only hosting