GPU dedicated servers in Dallas, Ashburn and Hillsboro.
GPU Rent Hub rents dedicated GPU servers from three datacenters in the United States: DFW-1 in Dallas, Texas, IAD-1 in Ashburn, Virginia and PDX-1 in Hillsboro, Oregon. The price is identical in all three, terms are weekly or monthly, and payment is in Bitcoin, Ethereum, USDT, USDC, Monero, Solana or Litecoin with no identity verification.
What is a GPU Rent Hub region?
One cage, in one building, with our hardware in it.
We lease space, power and cross-connects from operators who are good at buildings. We do not lease compute from anyone. All three sites are in the United States and we have no capacity anywhere else, which is deliberate: the legal position for a no-KYC GPU host is clearest here, and one jurisdiction means one answer about where your data physically sits. More on that on the about page.
Same in every region
Price, specification, images, API, 20 TB of included outbound, support queue, and the rule that we never inspect your workload.
Different per site
Latency to your users, which cards are free right now, and which 8-way node types exist there. That is the whole decision.
What is in DFW-1, Dallas, Texas?
The first site and about 60% of the fleet.
DFW-1 opened in March 2021 with forty RTX 3090s in one rack. It is now four cages and 11 MW holding roughly 3,800 GPUs on the ERCOT grid, with 2N UPS and N+1 diesel behind it. Cooling is chilled water CRAH with rear-door heat exchangers on the HGX racks, which is why the 1 kW-per-card platforms land here first: every MI300X we own and every whole 8-way B200 node is in Dallas.
Carriers are Cogent, Lumen and Hurricane Electric, with a fourth diverse path from Zayo ordered after a contractor cut the Lumen conduit in July 2026. Best for multi-week training, anything needing a whole NVSwitch node, and users spread across the continent. Hardware bought on the GPU sales page ships from here, and most colocated customer hardware lives here.
- Opened
- March 2021
- Cages · power
- 4 · 11 MW
- GPUs installed
- ~3,800
- Free today
- 534 cards, all 15 SKUs
- Carriers
- Cogent · Lumen · Hurricane Electric (Zayo ordered)
- Power
- 2N UPS · N+1 diesel · ERCOT
- Cooling
- Chilled water CRAH · rear-door HX on HGX racks
- Latency
- Chicago 22 ms · Los Angeles 34 ms · New York 38 ms
- Only site with
- MI300X · 8-way B200 nodes
What is in IAD-1, Ashburn, Virginia?
Seven milliseconds to New York. The serving site.
IAD-1 opened in November 2022, two days later than announced because a cross-connect was late. Two cages, 5 MW, roughly 1,600 GPUs on the PJM grid, 2N UPS and N+1 diesel, chilled water CRAH throughout. There are no rear-door heat exchangers in Ashburn yet, so there is no OAM capacity: MI300X stays a Dallas product until that changes.
Carriers are Lumen, Zayo and GTT; we ran on one upstream here for three months in 2022 and it cost us a degradation, so every cage has had at least two since. At 7 ms to New York, 74 ms to London and 88 ms to Frankfurt, an endpoint here covers the eastern seaboard and most of Western Europe. Best for production inference on H200 SXM, H100 SXM, L40S and RTX 4090.
- Opened
- November 2022
- Cages · power
- 2 · 5 MW
- GPUs installed
- ~1,600
- Free today
- 331 cards, 13 of 15 SKUs
- Carriers
- Lumen · Zayo · GTT
- Power
- 2N UPS · N+1 diesel · PJM
- Cooling
- Chilled water CRAH
- Latency
- New York 7 ms · London 74 ms · Frankfurt 88 ms
- Not here today
- MI300X · B200
What is in PDX-1, Hillsboro, Oregon?
Hydro power, and outside air doing the cooling.
PDX-1 opened in July 2024: one cage, 3 MW, roughly 900 GPUs. Power comes off the Bonneville Power Administration hydro system, and the facility runs a free-air economiser about eight months a year, so mechanical cooling only carries the summer. It is 2N UPS with N+1 diesel like the other two. That economiser is why our largest free L40S pool and a deep L4 fleet sit here rather than in Texas.
Carriers are Lumen, Hurricane Electric and Zayo. Seattle is 5 ms away, San Francisco 14 ms and Tokyo 92 ms, so Hillsboro is the pick for Pacific and Asian users, and for always-on services where power cost matters. Two B200 cards are free here today but no whole 8-way node. H200 and A40 stock is zero at the moment.
- Opened
- July 2024
- Cages · power
- 1 · 3 MW
- GPUs installed
- ~900
- Free today
- 246 cards, 12 of 15 SKUs
- Carriers
- Lumen · Hurricane Electric · Zayo
- Power
- 2N UPS · N+1 diesel · BPA hydro
- Cooling
- Free-air economiser ~8 months/yr
- Latency
- Seattle 5 ms · San Francisco 14 ms · Tokyo 92 ms
- Not here today
- H200 · A40 · MI300X
How do the three datacenters compare?
Everything that differs, plus the row that does not.
| DFW-1 Dallas, TX | IAD-1 Ashburn, VA | PDX-1 Hillsboro, OR | |
|---|---|---|---|
| Opened | March 2021 | November 2022 | July 2024 |
| Cages · power | 4 · 11 MW | 2 · 5 MW | 1 · 3 MW |
| GPUs installed | ~3,800 | ~1,600 | ~900 |
| Cards free today | 534 | 331 | 246 |
| Grid | ERCOT | PJM | BPA hydro |
| Power topology | 2N UPS · N+1 diesel | 2N UPS · N+1 diesel | 2N UPS · N+1 diesel |
| Cooling | CRAH · rear-door HX | CRAH | Free-air ~8 months/yr |
| Carriers | Cogent · Lumen · HE | Lumen · Zayo · GTT | Lumen · HE · Zayo |
| 8-way nodesNVSwitch or Infinity Fabric | B200 · H200 · H100 · MI300X · A100 · L40S · 5090 · 4090 | H200 · H100 · A100 · L40S · 5090 · 4090 | H100 · A100 · L40S · 5090 · 4090 |
| Best for | Training and whole nodes | East coast and EU serving | West coast and Asia serving |
| Price | identical | identical | identical |
What does a GPU dedicated server cost in Dallas, Ashburn or Hillsboro?
The same to the dollar in each region. The last three columns are cards free right now, not cards installed.
| GPU | Per month | Per week | Per GPU-hour | DFW-1 | IAD-1 | PDX-1 |
|---|---|---|---|---|---|---|
| B200180 GB HBM3e | $3,890/mo | $1,090 | $5.40 | 6 | 0 | 2 |
| H200 SXM141 GB HBM3e | $2,290/mo | $640 | $3.18 | 11 | 8 | 0 |
| H100 SXM80 GB HBM3 | $1,690/mo | $475 | $2.35 | 32 | 19 | 12 |
| MI300X192 GB HBM3 | $1,590/mo | $450 | $2.21 | 8 | 0 | 0 |
| H100 PCIe80 GB HBM2e | $1,490/mo | $420 | $2.07 | 14 | 9 | 4 |
| A100 SXM80 GB HBM2e | $990/mo | $280 | $1.38 | 40 | 22 | 16 |
| RTX 509032 GB GDDR7 | $540/mo | $155 | $0.75 | 48 | 30 | 22 |
| L40S48 GB GDDR6 ECC | $520/mo | $148 | $0.72 | 26 | 18 | 30 |
| RTX 6000 Ada48 GB GDDR6 ECC | $460/mo | $130 | $0.64 | 9 | 12 | 5 |
| RTX 409024 GB GDDR6X | $390/mo | $110 | $0.54 | 120 | 84 | 61 |
| RTX A600048 GB GDDR6 ECC | $290/mo | $82 | $0.40 | 34 | 20 | 11 |
| L424 GB GDDR6 · 72 W | $250/mo | $70 | $0.35 | 40 | 28 | 36 |
| A4048 GB GDDR6 ECC | $240/mo | $68 | $0.33 | 22 | 15 | 0 |
| RTX 309024 GB GDDR6X | $230/mo | $65 | $0.32 | 96 | 52 | 38 |
| RTX A500024 GB GDDR6 ECC | $190/mo | $55 | $0.26 | 28 | 14 | 9 |
How do I choose a region?
Two questions decide it. If neither points anywhere, take DFW-1: deepest stock.
Latency to whoever calls your endpoint
Eastern seaboard or Europe: Ashburn, 7 ms from New York, 74 ms from London. West coast or Asia: Hillsboro, 5 ms from Seattle, 92 ms from Tokyo. Everywhere at once: Dallas, 22 ms to Chicago and 38 ms to New York.
Whether the card you want is free there
Stock differs by site and moves hourly. MI300X and whole 8-way B200 nodes are Dallas only, H200 is Dallas and Ashburn, everything else exists in all three in different quantities.
Nothing else, because nothing else differs
Same price, specification, images, API, included outbound and support queue, and the same policy of not looking at your workload. Region is an inventory and network decision, not a product tier.
Instances do not migrate. A rental is a physical card in a physical cage, so changing region means releasing one instance and deploying another; snapshots and network volumes stay inside their region. Every site takes the same seven coins: see paying for GPUs with crypto and Monero GPU hosting.
What happened at DFW-1 in August 2024?
On 14 August 2024 a storm dropped the utility feed into DFW-1 at 01:00 CDT. The generators started as designed. The B-side UPS string failed to carry cage 3 through the transfer, and 412 instances lost power for up to four hours and forty-five minutes. No customer reported data loss. The facility replaced the string; we credited seven days to every affected account without being asked.
The costly part was the decision afterwards: 2N UPS on every cage in all three sites by October 2024, rather than live with a single-string failure mode inherited from a building. It is why the power topology row above reads the same for a 2021 cage in Texas and a 2024 cage in Oregon. Every incident since April 2021 is in the incident history. Trailing twelve-month uptime is 99.97%.
Questions about the three sites.
Where are GPU Rent Hub's GPU datacenters?
Is a GPU dedicated server in Dallas cheaper than Ashburn or Hillsboro?
Does GPU Rent Hub own its datacenters or resell someone else's capacity?
Which region has the most GPUs in stock right now?
What backup power do the datacenters have?
Do you have GPU servers outside the United States?
Can I colocate my own GPU server in Dallas?
Pick a city, deploy in about ninety seconds.
A username and a password, a deposit in any of seven coins, then choose DFW-1, IAD-1 or PDX-1 at deploy time. No email, no documents, no sales call.