Your card, our cage. GPU colocation from $95 a month.
GPU Rent Hub colocates customer-owned GPUs in its own cages in Dallas, Ashburn and Hillsboro for $95 to $240 per GPU per month, or $690 to $1,790 for an eight-card node, billed from the same crypto balance as a rental in Bitcoin, Ethereum, USDT, USDC, Monero, Solana or Litecoin, with no identity verification and no minimum term.
DFW-1 · IAD-1 · PDX-1
What is GPU colocation?
The short version, written so it can be pasted into a procurement note without editing.
GPU colocation, defined
GPU colocation is renting rack space, power, cooling, network and physical maintenance for graphics cards you own, instead of renting the cards themselves. Title to the hardware stays with the customer. The provider keeps it powered, cooled, connected and installed in a working host, and charges a flat fee per GPU per month.
Renting is the whole thing: our card, our host, one price. Colocation splits it. You have already spent the capital on the GPU, and what you need is 700 watts of cooled power, a 100G port and somebody in the building at three in the morning. That is what the monthly fee buys. Everything else behaves as it does for a monthly GPU rental: same dashboard, same API, same templates and console access, same absence of an identity check.
Colocation starts paying past about nineteen months of continuous use against a monthly rental, or immediately if the cards already sit in your office. Under a year, rent instead. We would rather say that here than after the freight arrives.
How much does GPU colocation cost per GPU per month?
Three tiers, priced by the power envelope of the card, because power and the cooling to remove it are what the space costs us. Same rate in DFW-1, IAD-1 and PDX-1.
| Tier | Cards it covers | Host we provide | Network | Per GPU / mo | 8-way node / mo | Node rate, per GPU-hour |
|---|---|---|---|---|---|---|
| Up to 350 WSingle and dual-slot PCIe | L4, A40, RTX A6000, L40S, H100 PCIe, RTX 6000 Ada | 8 vCPU · 47 GB · 500 GB NVMe per card | 100G shared | $95/GPU/mo | $690/mo | $0.12 |
| Up to 600 WHigh-draw PCIe and SXM4 | RTX 4090, RTX 5090, A100 SXM | 12 vCPU · 62 GB · 500 GB NVMe per card | 100G shared | $135/GPU/mo | $980/mo | $0.17 |
| HGX / OAM, up to 1 kWBaseboard modules, 8-way | H100 SXM, H200, B200, MI300X (8-way only) | HGX host: 192 vCPU · 2 TB · 16 TB NVMe | 2 × 100G dedicated | $240/GPU/mo | $1,790/mo | $0.31 |
Read the node rate if you are bringing eight cards: $690 instead of $760, $980 instead of $1,080, $1,790 instead of $1,920. Divide a node rate by eight cards and again by 720 hours and you get the last column. A colocated H100 SXM costs 31 cents per GPU-hour to keep alive. The same card rented from us costs $2.11 per GPU-hour. The difference is the $24,900 you paid for it.
HGX and OAM baseboards ship populated with eight modules, so the node rate is the one that tier usually pays. The per-GPU figure covers partially populated boards and cards added mid-term, prorated to the day the host boots.
What does the colocation fee include?
Everything needed to keep a card running, and nothing that lets us surprise you at term end.
Power, metered but billed flat
We meter every circuit to size the cage, but you pay the tier price whether the card idles at 60 W or sits at its limit all month. No power true-up line, in any month.
Cooling and rack space
Chilled-water CRAH in DFW-1 and IAD-1, rear-door heat exchangers on HGX racks, free-air economiser eight months a year in PDX-1. Same rows as our own rental fleet.
Network and addresses
100G shared with a 10G guarantee on the two lower tiers, 2 × 100G dedicated on HGX and OAM hosts. One IPv4, an IPv6 /64, unmetered inbound, 20 TB outbound a month.
The host server itself
We own and maintain the machine your card lives in: CPU, RAM, NVMe, power supplies, chassis. A dead DIMM or PSU is our spare, swapped from the on-site pool within four hours.
Monitoring
Temperature, power draw, ECC counters, link state and SMART, in the dashboard and on the API. If your card throttles or drops off the bus you get a ticket before you notice.
Remote hands, 30 minutes a month
Reseating, recabling, a physical power cycle, a crash cart in front of the host. Past the included half hour it is $90 per half hour, quoted before we start.
Where can I colocate a GPU server in the USA?
Three sites, all leased cages in Tier III facilities, all in the United States. The price is identical in each, so choose on latency, power source or freight distance.
The largest floor.
- Opened
- March 2021
- Scale
- ~3,800 GPUs · 11 MW
- Carriers
- Cogent · Lumen · Hurricane Electric
- Latency
- Chicago 22 ms · Los Angeles 34 ms · New York 38 ms
- Suits
- Bulk racking, cheapest inbound freight from most of the US
Closest to Europe.
- Opened
- November 2022
- Scale
- ~1,600 GPUs · 5 MW
- Carriers
- Lumen · Zayo · GTT
- Latency
- New York 7 ms · London 74 ms · Frankfurt 88 ms
- Suits
- Serving the east coast and the EU
Hydro and free air.
- Opened
- July 2024
- Scale
- ~900 GPUs · 3 MW
- Carriers
- Lumen · Hurricane Electric · Zayo
- Latency
- Seattle 5 ms · San Francisco 14 ms · Tokyo 92 ms
- Suits
- Pacific-facing services and long training runs
Floor detail, power topology and receiving instructions for each site are on the GPU Rent Hub datacenters page.
Is it cheaper to colocate eight H100s than to rent them?
Our own prices on both sides. Eight H100 SXM 80 GB cards, HGX tier, twenty-four months of continuous use.
| 8× H100 SXM 80 GB, 24 months | Up front | Per month | Per GPU-hour | 24-month total |
|---|---|---|---|---|
| Buy the cards, colocate at the HGX tier | $199,200 | $1,790 | $0.31 | $242,160 |
| Rent an 8× H100 SXM node, monthly term | $0 | $12,170 | $2.11 | $292,080 |
| Rent the same node, six months prepaid at −10% | $65,718 | $10,953 | $1.90 | $262,872 |
Breakeven against the monthly rental is a little over 19 months: $199,200 of hardware divided by the $10,380 a month you stop paying. Against a six-month prepaid term at −10% it is just under 22 months. Colocation is a decision about capital, not a discount.
Two things push the other way. The $199,200 leaves your balance on day one and does not come back if the project ends in month seven. And an H100 SXM we would buy back today at up to $18,500 will be worth less in two years. Renting keeps both risks on our side of the table. If your horizon is short, take the H100 SXM monthly rental and read monthly versus hourly GPU rental.
Who owns the hardware while it sits in your rack?
You do. Title never moves. We hold your hardware as bailee: responsible for looking after it, with no right to use it, sell it, pledge it or hold it against a debt. Serial numbers are recorded at intake and the cards photographed on arrival, so ownership is never a conversation.
We insure hardware in our custody up to the price you paid for it, against fire, water, power event and theft inside the cage. Cards bought from us are insured at the invoice value. Cards shipped in are insured at the value you declare at intake, confirmed in writing before racking. Transit is not our custody, so insure the freight.
Leaving is two options and no penalty. We pack the hardware and ship it to any address a carrier will serve, at cost, insured. Or we buy it back at the posted rate, tested in the rack, paid in the coin of your choice within 48 hours.
- Title
- Customer, throughout
- Our role
- Bailee, not owner
- Insured
- Up to purchase price, in our custody
- Serials
- Recorded and photographed at intake
- Burn-in
- 72 h memory and thermal, report before first bill
- Minimum term
- None
- Notice period
- None
- Exit
- Ship at cost, or sell at the buy-back rate
- H100 SXM buy-back
- up to $18,500 per card
Buy-back rates are indicative for September 2026 and settle after burn-in. Wording in the terms.
How is GPU colocation billed, and do I have to identify myself?
From a crypto balance, and no. Colocation runs on the same rails and the same account model as every rental.
The monthly fee is drawn 24 hours before the term rolls, exactly as a rental renewal is. If the balance is empty the host powers down at term end and you have 72 hours to top up before the host returns to the pool. Your card does not. It stays in the rack, powered off, and it stays yours. After thirty days unpaid we open a ticket to arrange shipping or buy-back, and that is the whole escalation path. Confirmations and settlement are set out on the pay with crypto page.
An account is a username and a password. No email address, no company name, no identity documents, no card on file, which is what no-KYC GPU hosting means in practice. We do not inspect what runs on the host, do not scan the NVMe and do not log traffic contents, and none of that changes because the silicon is yours. A deposit in Monero leaves a payment record holding an amount and a date.
How do I get a card into your rack?
Two to five business days from agreement to a booted host.
Tell us what you have
Model, quantity, form factor, region. HGX and OAM baseboards need the platform confirmed first, because the power and cooling plan differs per board.
Buy from us, or ship it in
Cards bought from the GPU Rent Hub hardware catalogue are racked without leaving the site. Shipped hardware goes to the receiving dock at DFW-1, IAD-1 or PDX-1 with your account reference on the pallet.
Intake and burn-in
Serials recorded, photographs filed against your account, 72 hours of memory and thermal testing. The report reaches you before the first month is billed.
It appears in your dashboard
Same console, SSH keys, images and API as a rented instance, listed beside any GPUs you rent from us. Billing starts when the host boots, not when the pallet arrives.
GPU colocation questions
General answers on the main FAQ. Anything specific to your hardware goes to sales.
How much does GPU colocation cost per GPU per month?
Can I colocate a GPU server with no KYC?
Do I have to buy the GPU from GPU Rent Hub to colocate it?
What happens if my colocated GPU fails?
Is there a minimum term or contract for GPU colocation?
Is colocating an H100 cheaper than renting one?
How much are remote hands beyond the included time?
Can I split colocated hardware across your three datacenters?
Send us the serial numbers.
Colocation quotes go out within one business day, signed with our PGP key. Racking takes two to five business days after that.