Home/Colocation
High-density AI colocationHalls engineered for 160 kW racks and the systems you already own
Bring your own NVIDIA HGX, DGX, MGX or OEM GPU platforms into Velyrix data centres built from the ground up for accelerated computing — direct-to-chip liquid cooling, rail-optimised fabric, concurrent maintainability and engineers on site around the clock.
Buy the hardware wherever you like. We take it from the loading dock onwards.
Purchase your NVIDIA GPU servers from any authorized OEM or distributor, ship them directly to our facility, and our engineering team will handle deployment, networking, cooling, cluster configuration, monitoring, and ongoing support. Colocation at Velyrix is not a cage and a power whip — it is an operating team.
We host the servers the AI market actually buys
Velyrix halls are pre-engineered for the mainstream accelerated platforms — from Ampere-era HGX A100 nodes through to rack-scale Blackwell Ultra. Power whips, floor loading, coolant distribution and fabric runs are already sized for each.
| Platform | GPUs per system | GPU memory | Form factor | Typical power | Cooling supported |
|---|---|---|---|---|---|
| NVIDIA GB300 NVL72 | 72 Blackwell Ultra + 36 Grace | 20.7 TB HBM3e per rack | Rack-scale (ORv3 / MGX) | ~120–140 kW per rack | Direct-to-chip liquid (required) |
| NVIDIA GB200 NVL72 | 72 Blackwell + 36 Grace | 13.4 TB HBM3e per rack | Rack-scale (ORv3 / MGX) | ~120 kW per rack | Direct-to-chip liquid (required) |
| HGX B300 — 8 GPU | 8× B300 | 2.3 TB HBM3e | 8–10U air / 4–6U liquid | ~13–15 kW per node | DLC preferred, RDHx assisted air |
| HGX B200 — 8 GPU | 8× B200 | 1.44 TB HBM3e | 8–10U | ~10–12 kW per node | Air with containment, RDHx or DLC |
| HGX H200 — 8 GPU | 8× H200 SXM | 1.1 TB HBM3e | 8U | ~10.2 kW per node | Air, RDHx, DLC retrofit |
| HGX H100 — 8 GPU | 8× H100 SXM | 640 GB HBM3 | 8U | ~10.2 kW per node | Air, RDHx, DLC retrofit |
| HGX A100 — 8 GPU | 8× A100 SXM | 640 GB HBM2e | 4–6U | ~6.5 kW per node | Air with containment |
| NVIDIA DGX H100 / H200 / B200 | 8× SXM | 640 GB – 1.44 TB | 8–10U | ~10.2–14.3 kW | Air, RDHx, DLC per generation |
| PCIe GPU servers | 4–10× PCIe | Varies | 2–5U | ~3–7 kW per node | Standard air |
Power and thermal figures are planning envelopes; the deployed values are taken from your OEM's published specifications during the engineering review. Product names are trademarks of their respective owners and are listed to describe supported hardware compatibility only.
Four thermal regimes under one roof
Most AI estates are mixed: legacy air-cooled inference alongside liquid-cooled training racks. Velyrix halls run both on the same floor, with coolant loops and containment provisioned per pod rather than per building.
- Contained air — up to 20 kW/rack: hot-aisle containment, blanking discipline, in-row CRAH
- Rear-door heat exchanger — up to 60 kW/rack: passive or active RDHx on chilled water
- Direct-to-chip liquid — up to 160 kW/rack: CDU-fed cold plates, N+1 pumps, leak detection and automatic isolation
- Single-phase immersion — on request: for specialised high-density or edge deployments
- Facility water: W3/W4 supply temperatures with heat-reuse offtake where the campus supports it
Representative rack elevation for a liquid-cooled Blackwell Ultra pod. Velyrix engineers produce a stamped elevation and power schedule for your specific hardware before delivery.
From a single cabinet to a dedicated hall
Cabinet
Locking 48U cabinet, 10–40 kW, A+B metered power, shared secure hall. Ideal for a first pod of inference nodes.
- From 1 rack
- Air or RDHx
- Metered or committed power
Private cage
Floor-to-ceiling caged footprint with your own access control, cameras and cable management.
- 4–60 racks
- Up to 160 kW/rack
- Dedicated CDU option
Private suite
A dedicated room with its own entrance, cooling plant and electrical lineup — audited as your space.
- 500 kW – 6 MW
- Dedicated infrastructure
- Customer-badged access
Wholesale / build-to-suit
Powered shell or fully fitted AI hall delivered to your specification on a long-term lease.
- 6–40 MW
- 10–15 year terms
- Custom resilience & PUE targets
Priced on power, not on floor space
AI colocation is a power business. Velyrix prices per kW of committed IT load, with cooling method and resilience level as the main variables.
| Product | Density | Cooling | Commitment | Per kW / month | Setup (NRC) |
|---|---|---|---|---|---|
| Retail cabinet | 10–20 kW | Contained air | 12–36 months | $215 | $1,950 |
| High-density cabinet | 20–40 kW | Air + containment | 12–36 months | $190 | $3,250 |
| RDHx cabinet | 40–60 kW | Rear-door heat exchanger | 24–60 months | $170 | $5,800 |
| Liquid-cooled pod | 60–160 kW | Direct-to-chip + CDU | 36–60 months | $155 | Quoted per pod |
| Private cage | 250 kW – 2 MW | Mixed | 36–60 months | $145 | Quoted per pod |
| Private suite | 500 kW – 6 MW | Dedicated plant | 60–120 months | $130 | Quoted per pod |
| Wholesale hall | 2 – 40 MW | Dedicated plant | 120–180 months | $110 | Quoted per pod |
| Service | Unit | Price |
|---|---|---|
| Single-mode fibre cross-connect | per connection / month | $315 |
| Copper cross-connect | per connection / month | $195 |
| Blended IP transit | per Mbps / month (95th) | $0.43 |
| 100G unmetered uplink | per port / month | $1,600 |
| Private cloud interconnect | per 10G / month | $545 |
| Remote hands — scheduled | per hour | $230 |
| Remote hands — emergency P1 | per hour | $375 |
| Smart hands retainer | 10 hours / month | $2,000 |
| Secure receiving, staging & de-trash | per pallet | $235 |
| Certified media destruction | per drive | $43 |
Indicative USD list pricing, exclusive of tax and metered power where applicable. Regional energy tariffs, resilience level, term length and committed load materially change the final rate; all colocation is quoted individually following an engineering review.
Run by engineers, not a ticket queue
Physical security
Fenced perimeter, mantrap entry, biometric plus badge, 90-day CCTV retention, escorted visitor policy and per-cage access logs available to your auditors.
Power resilience
Diverse utility feeds, 2N UPS with N+1 generators, on-site fuel for 48–72 hours under contract, and monthly load-bank tested maintenance windows.
Connectivity
Carrier-neutral meet-me rooms with 20+ networks, diverse conduit entries, dark fibre between campuses and direct cloud on-ramps.
Deployment support
Receiving, unboxing, rack & stack, structured cabling, labelling and asset registration — or hand the whole build to our deployment team.
Customer portal
Real-time per-circuit power draw, environmental telemetry, ticketing, access requests, invoices and capacity planning in one place, with API access.
Sustainability
Renewable power contracting, design PUE of 1.15–1.25, water-conscious cooling design and Scope 2 reporting for your own disclosures.
RMA & spares management
For customer-owned hardware we open the RMA under your support contract, supply logs and failure evidence, track the dispatch and fit the part — with consignment or Velyrix-held spares for a four-hour replacement target.
Asset & serial reporting
Every unit received is photographed, serialised and registered in your portal, with firmware levels, warranty end dates and failure history exportable for your own asset management system.
OEM & reseller friendly
Buying through a partner? We work alongside your OEM account team and reseller, deliver under their brand where required, and register their deal. Partner programme →
Own the GPUs. Let us run them.
Buy your NVIDIA servers from any OEM or distributor you like, ship them to a Velyrix hall, and we handle the rest — deployment, fabric, cooling, monitoring and support. Or rent ours. Either way, you get a plan in one business day.
Frequently asked questions
Can I buy my own GPU servers and have Velyrix run them?
Yes. That is the Bring Your Own GPU model: you buy from any authorized OEM or distributor, ship the hardware to a Velyrix facility, and our engineers handle deployment, networking, cooling, cluster configuration, monitoring and ongoing support, including RMA management with your vendor. You are never required to purchase hardware through Velyrix.
Which GPU servers can Velyrix colocation support?
Velyrix supports the mainstream accelerated platforms: NVIDIA HGX and DGX systems with A100, H100, H200, B200 and B300 GPUs, rack-scale GB200 NVL72 and GB300 NVL72, MGX and OCP ORv3 designs, and PCIe GPU servers. OEM chassis from Dell, Supermicro, HPE, Lenovo, GIGABYTE, QCT, ASUS, Foxconn and Wiwynn are all accommodated.
What rack power density can you deliver?
Contained air supports up to roughly 20 kW per rack, rear-door heat exchangers up to about 60 kW, and direct-to-chip liquid cooling up to 160 kW per rack. GB300 NVL72 racks at approximately 120-140 kW are deployed on direct-to-chip liquid loops with N+1 coolant distribution units.
How is AI colocation priced?
Per kW of committed IT load per month, not per square foot. Indicative rates run from about $215 per kW for retail cabinets down to roughly $110 per kW for wholesale halls, with cooling method, resilience level, term and regional energy cost as the main variables.
Can Velyrix install and cable the hardware for us?
Yes. Our AI Infrastructure Deployment and Commissioning service covers receiving, rack and stack, power and fabric cabling, firmware baselining, burn-in, NCCL validation and a documented acceptance test before handover.
Do you support liquid cooling for Blackwell and Blackwell Ultra?
Yes. Liquid-cooled halls are in production in Ashburn and Dallas with coolant distribution units, W3/W4 facility water, leak detection with automatic isolation, and redundant pumping. GB300 NVL72 and HGX B300 deployments require direct-to-chip cooling and are engineered on that basis.
What connectivity is available in the data centre?
Carrier-neutral meet-me rooms with more than twenty networks, cross-connects to any tenant, blended IP transit, dedicated 100G ports, dark fibre between Velyrix campuses and private on-ramps to AWS, Azure, Google Cloud and Oracle Cloud.