XCITY COMPUTE · SERVED FROM XCITY POD
Reserve GPU capacity where
energy is the advantage.
Dedicated Blackwell and Hopper clusters, served bare-metal from XCITY POD modular data centers at energy-advantaged sites. One per-GPU-hour rate covers compute, storage, networking and 24/7 operations — quoted within one business day.
Blackwell · Hopper - NVL72 rack domains
- HGX 8-GPU nodes
- Weeks to first job
- 72GPUs per NVL72 rack, one NVLink domain
- 56 kWrack density inside every XPOD
- 24/7remote operations and on-call engineers
- 1 daybusiness-day reply with a sized quote
The fleet
The machines that matter,
on the grid that matters.
Rack-scale NVLink domains for frontier training, node-scale HGX systems for everything else — every configuration dedicated, never shared.
Rack-scale · NVL72
Full NVLink domains for frontier training — 72 GPUs acting as one accelerator, delivered per rack.
-
GB300 NVL72
Reservations openNVIDIA Blackwell Ultra
- GPUs
- 72× Blackwell Ultra per rack
- Memory
- 288 GB HBM3e per GPU
- Fabric
- NVLink 5 · Quantum-X800 InfiniBand
-
GB200 NVL72
Available nowNVIDIA Blackwell
- GPUs
- 72× Blackwell per rack
- Memory
- 192 GB HBM3e per GPU
- Fabric
- NVLink 5 · Quantum-2 InfiniBand
Node-scale · HGX
Eight-GPU nodes for training, fine-tuning and high-throughput inference — from a single node to hundreds.
-
HGX B200
Available nowNVIDIA Blackwell
- GPUs
- 8× B200 per node
- Memory
- 180 GB HBM3e per GPU
- Fabric
- NVLink · 400 Gb/s InfiniBand
-
HGX H200
Available nowNVIDIA Hopper
- GPUs
- 8× H200 per node
- Memory
- 141 GB HBM3e per GPU
- Fabric
- NVLink 4 · 400 Gb/s InfiniBand
-
HGX H100
Available nowNVIDIA Hopper
- GPUs
- 8× H100 per node
- Memory
- 80 GB HBM3 per GPU
- Fabric
- NVLink 4 · 400 Gb/s InfiniBand
Every configuration is quoted per GPU-hour after scoping — no public price list, no hidden line items.
Why XCITY compute
The rate follows
the power source.
-
Energy is the cost advantage
PODs deploy where power is cheapest — stranded gas, renewables, industrial microgrids. Lower energy cost flows straight into your per-GPU-hour rate.
-
Capacity in weeks, not years
Factory-integrated XPOD units stand up new capacity in weeks. Your reservation maps to real, scheduled hardware — not a waitlist.
-
Placed where you need it
Relocatable modules put capacity near your data, your users or your jurisdiction — and move when your requirements do.
Compute on this page is served from the XCITY POD platform. Explore the POD hardware →
The platform
Built for training.
Run like infrastructure.
-
Bare-metal clusters
Dedicated GPUs with no hypervisor and no noisy neighbors. SSH in and run.
-
Non-blocking fabric
NVLink domains inside the rack, full-bisection InfiniBand across nodes.
-
Co-located storage
Parallel storage sized so checkpoints and dataloaders never gate training.
-
Your scheduler
Slurm or Kubernetes, deployed and burned in before handover.
-
Remote 3D operations
Every POD is monitored on the XCITY remote operations platform — telemetry, alerts, diagnostics.
-
Flexible terms
From one month to multi-year commitments, sized from one node to full campuses.
How it works
Reservation to first job,
in four steps.
- 01
Scope
Tell us the workload — model scale, framework, storage and networking needs.
- 02
Reserve
Lock capacity with a go-live date and a firm per-GPU-hour rate. Terms from one month up.
- 03
Provision
We burn in the cluster, benchmark it and deploy your scheduler before handover.
- 04
Operate
Run your jobs. XCITY engineers watch the fleet 24/7 and report utilization.
FAQ
Straight answers.
How is pricing structured?
Everything is quoted per GPU-hour — compute, storage, networking and operations in one rate. There is no public price list because rates depend on SKU, scale, term and site energy cost; you get a firm number within one business day of scoping.
What commitment terms are available?
From one month to multi-year reservations. Longer terms and larger clusters get better per-GPU-hour rates.
When does GB300 capacity come online?
GB300 NVL72 racks are open for reservation now, with first deployments scheduled as Blackwell Ultra systems ship. Reserving early secures your place in the allocation queue.
Where does the compute physically run?
In XCITY POD modular data centers deployed at energy-advantaged sites. You access clusters over secure networking — nothing ships to you, and nothing needs your own facility.
Is this bare metal or virtual machines?
Bare metal. You get dedicated nodes with direct hardware access, your choice of scheduler, and no virtualization overhead.
What happens when hardware fails?
Hot spares are held at every site. Failed nodes are swapped out of the cluster in minutes and repaired offline; the fleet is monitored 24/7 on the remote operations platform.
How does this relate to XCITY POD?
POD is the modular data-center platform; this page is the compute served from it. If you want to own or host the infrastructure itself, request a POD briefing instead.
Reserve
Five quick questions.
A firm quote follows.
Tell us the machines, the scale and the term. An XCITY engineer replies within one business day with firm availability and a sized per-GPU-hour proposal.
- Firm availability for your target SKU and scale
- Per-GPU-hour rate covering compute, storage and operations
- Go-live schedule from reservation to first job
Prefer email? hello@xcity.ai
XCITY ENERGY DEVELOPMENT INC.
Bare metal. Per GPU-hour. Energy-Aware.
Reservation request received.
An XCITY engineer will reply within one business day with availability and a sized quote. For anything urgent, reach us at hello@xcity.ai.