Private GPU compute,
scaling to 5 megawatts.
Dedicated, secure, compliant GPU infrastructure in US & Europe-based facilities for AI teams whose data can’t sit on shared cloud. We’re growing to 5 MW of capacity — and building it with partners.
B300
First deployment launches on NVIDIA B300
US & Europe
Domestic data residency, no cross-border transfer
1 PB+
Networked storage for long training runs
24/7
Telemetry, on-site staff, remote hands
01 · Why demand is there
Public cloud GPUs are multi-tenant — weights and datasets co-reside with others.
Reserved cloud queues run 3–6 months, past most project windows.
Offshore capacity fails residency and export-control policy.
Consumer cards lack ECC and warranty, and fail on sustained runs.
02 · What we deliver
Training — dedicated multi-node clusters for foundation and domain models.
Fine-tuning — isolated capacity for sensitive, regulated data.
Inference — private endpoints on the same hardware, no model egress.
Hardware — NVIDIA B300 first, plus H200 and H100 matched to each workload.
Phase 1
First deployment: B300 clusters with redundant power, cooling and connectivity. [MW]
1 MW
~550–750 GPUs. B300 anchor clusters alongside H200 and H100.
2.5 MW
~1,400–1,900 GPUs. Mixed B300, H200, H100 across US facilities.
TARGET
5 MW
~2,800–3,900 GPUs. Full training, fine-tuning and inference estate.
Per MW of IT load: B300 8-GPU nodes at ~14 kW, H100/H200 nodes at ~10 kW. Range spans a B300-heavy to an H-series-heavy mix.
04 · Ways to partner