GB300 NVL72 · HGX B300 · United States
Alder Compute designs, builds and operates dedicated NVIDIA Blackwell Ultra clusters — liquid-cooled, rail-optimized, handed over burned-in.
01 — Footprint
Sites in operation and in development across the U.S., each chosen for firm power and liquid-cooling headroom. We design, build and operate every cluster ourselves.
02 — Campus
Tier III and Tier III+ facilities, concurrently maintainable from the utility to the rack. Every watt is accounted for before a GPU is racked.
03 — Scalable unit
Eight GB300 NVL72 racks, a rail-optimized InfiniBand fabric and in-row liquid cooling make one unit. We build them two at a time.
04 — GB300 NVL72
Seventy-two Blackwell Ultra GPUs and thirty-six Grace CPUs in a single liquid-cooled rack, wired to behave like one accelerator.
05 — The spine
Thousands of NVLink copper cables run behind every rack. Supply and return manifolds carry roughly 90% of its heat away as liquid.
06 — Compute tray
Two Grace CPUs, four Blackwell Ultra GPUs, four ConnectX-8 SuperNICs at 800 Gb/s and a BlueField-3 DPU — cold plates on everything that matters.
07 — Blackwell Ultra
208 billion transistors across two reticle-size dies, joined at 10 TB/s and wrapped in eight stacks of HBM3e.
One GB300 NVL72 rack
Source: NVIDIA
Blackwell Ultra GPUs
in one NVLink domain
HBM3e
fast memory per rack
NVLink bandwidth
all-to-all, per rack
Dense FP4
per rack
01 — Ground to tokens
We don't resell someone else's cloud. Alder selects the sites, designs the liquid cooling, racks and cables every cluster, and operates it — so there is one accountable owner for every layer.
01 — Firm power first
Every deployment starts with megawatts we can stand behind: Tier III and Tier III+ U.S. facilities with utility power and backup generation.
02 — Two platforms
GB300 NVL72 when your model wants one enormous NVLink domain; HGX B300 when you want eight-GPU building blocks you can carve up. We run both, side by side, on the same fabric and storage.
Platform 01
Hover to open
A rack that behaves like one GPU — for frontier training and long-context reasoning.
| GPUs | 72 Blackwell Ultra |
|---|---|
| CPUs | 36 Grace (2,592 Arm cores) |
| GPU memory | 20 TB HBM3e |
| Scale-up | 130 TB/s NVLink · 72-GPU domain |
| Dense FP4 | ~1.1 exaFLOPS |
| Networking | 72 × 800 Gb/s ConnectX-8 |
| Power | ~132–142 kW per rack |
| Cooling | Direct-to-chip liquid |
Best forFrontier pre-training · reasoning & long-context inference · MoE at scale
Platform 02
Hover to open
Eight Blackwell Ultra GPUs per node — the flexible building block for training and serving.
| GPUs | 8 Blackwell Ultra |
|---|---|
| CPUs | 2 × x86 (Xeon 6 / EPYC 9005) |
| GPU memory | 2.1 TB HBM3e |
| Scale-up | 14.4 TB/s NVLink · 8-GPU domain |
| Dense FP4 | 108 petaFLOPS |
| Networking | 8 × 800 Gb/s ConnectX-8 |
| Power | ~14 kW per node |
| Cooling | Liquid (4U) or air (8–10U) |
Best forFine-tuning & post-training · inference fleets · multimodal & physical-AI pipelines
03 — Fabric
Three networks, each doing one job: NVLink inside the rack, a rail-optimized InfiniBand fat tree between racks, and a separate Ethernet fabric for storage.
Every tray’s NIC n lands on rail n. Traffic between GPUs of the same rank stays one switch hop away; everything else crosses the spine — with no oversubscription anywhere. 16 leaves, 8 spines, 1,152 GPUs.
Isolate a rail
Follow a packet
04 — Liquid cooling
Around 90% of a GB300 rack's heat leaves through cold plates. Drag the supply temperature to see how flow, return temperature and heat rejection trade off.
Warmer supply water needs more flow, but the return stays near 55 °C — hot enough to reject heat with dry coolers instead of chillers. That is where the efficiency comes from.
Flow per GB300 NVL72 rack (OEM guidance). Water, ~90% of heat to liquid. Chiller threshold illustrative.
05 — Workloads
The same infrastructure serves very different jobs. We size each cluster — platform, fabric, storage and software — around the work you actually run.
Pre-training on rack-scale NVLink domains; long-context and reasoning inference that keeps the KV cache in 20 TB of HBM.
Learn moreSimulation, synthetic data and world-model training for robotics and autonomy — thousands of parallel worlds per cluster.
Learn moreVideo, audio and image tokens at once. Fast storage and wide fabrics keep data-hungry pipelines fed.
Learn moreDedicated clusters on dedicated fabric, isolated from other tenants, with the controls your security team asks for.
Learn more06 — Scalable unit
Eight GB300 NVL72 racks, two network racks, in-row cooling and four rails of 800G InfiniBand overhead. Open the explorer to walk every rack, tray and cable.
07 — How we deliver
One team owns the whole path — so the schedule we give you is the schedule we keep.
Tell us the workload, scale and timeline. We propose a platform, fabric and storage design — and a delivery date we can stand behind.
We secure power and space, receive and rack the hardware, run the coolant loops and cable every rail ourselves.
Every node, link and cold plate is stress-tested. NCCL all-reduce and storage throughput are measured against targets before handover.
Bare metal, Kubernetes or Slurm on day one, with telemetry, health checks and engineers watching the fleet.