4
Engineered systems, repeated at every site
2N
Redundant power and network
90 days
Site control to live workload
SOC 2
Current compliance milestone

The elements of every site

Four systems, engineered together, repeated identically at every location in the network.

Fig. 01 — GB200 inference location / unit mapScale N.T.S.
GB200 inference location unit map Technical diagram of a GB200 inference location with power, cooling, and inference output. GB200 CLOSED COOLANT LOOP 0 GAL CONSUMED GRID + SOLAR · 420 kW DC INFERENCE OUTPUT 5.7× H200 / CAPEX EBITDA MARGIN ~85% MODELED 2N POWER · A/B UPS SIGNED ANCHOR OFFTAKE

Compute

NVIDIA GB300 NVL72

Spacebus sites run on current-generation GB300 NVL72 clusters, purpose-built for large-scale inference and training workloads at rack-scale density.

A defined upgrade path

Next-generation platforms — including NVIDIA Vera Rubin and Groq LPx — are treated as a roadmap upgrade into the same physical footprint, not a separate build. Sites are designed from day one to absorb the next hardware generation rather than be stranded by it.

Cooling

Water-zero, closed-loop

Every site uses closed-loop thermal management with no ongoing water draw. That removes local water availability as a siting constraint and eliminates a growing source of community and regulatory friction around data center growth.

Why it matters

Conventional evaporative cooling ties data center growth to local water systems. A water-zero design means Spacebus can site capacity in more places, including markets where water stress would otherwise rule out new construction.

Power

2N redundant architecture

Every site is built on fully redundant, 2N power infrastructure — no single point of failure between the grid and the GPU.

Owned onsite solar

Each site pairs grid power with owned, onsite solar generation, reducing dependence on grid interconnection timelines and easing the demand each site places on the local grid.

Networking

2N redundant 400GbE fiber

Every site connects out on fully redundant, 2N 400GbE fiber links — no single point of failure between the cluster and the network, matching the same redundancy standard applied to power.

Built for low-latency inference

High-bandwidth, low-latency connectivity keeps inference traffic moving as fast as the GPUs can serve it, so the network is never the bottleneck between a request and a response.

Deployment

The core of the Spacebus model: a standardized build process that takes a site from control to live workloads in 90 days.

Site control & power interconnection

Site secured and power interconnection agreements finalized alongside local utility and permitting partners.

Modular infrastructure build

Prefabricated power, cooling, and enclosure modules installed on a standardized design — minimizing on-site custom construction.

Cluster commissioning

GPU clusters installed, networked, and validated against performance and redundancy benchmarks.

Workload go-live

Contracted customer workloads onboarded and live — typically within 90 days of site control.

Security & compliance roadmap

Built to meet the assurance requirements of regulated and public-sector customers as the network scales.

StageMilestone
CurrentSOC 2 Type I
NextFedRAMP High
TargetDoD IL4 / IL5

See where this design is deployed today.

View the network