RTX PRO 6000 rental in 2026: hourly cloud prices vs a monthly server

In mid-October 2026, the clouds we checked charged $1.13 to $3.03 per GPU-hour for an on-demand RTX PRO 6000, and GetDeploying’s median across 40 providers was $2.33. A monthly server at $899 works out to $1.23 an hour if it runs all month, and costs less than RunPod’s $2.49 rate beyond 361 h of use a month.

The short version

Public prices checked on 15 October 2026, per GPU-hour, on demand, in US dollars and before VAT or sales tax.

Hourly billing wins for short or bursty work and for anything that needs several GPUs at once. A monthly server wins when the GPU stays on for most of the month, as an inference endpoint does, and when bandwidth, storage and the host around the GPU matter as much as the GPU-hour.

RTX PRO 6000 Blackwell: which edition are you renting?

NVIDIA sells the RTX PRO 6000 Blackwell in three editions. All carry 96 GB of GDDR7 memory with ECC on a PCIe Gen 5 card; they differ in power, cooling and memory bandwidth. Cloud listings often say only “RTX PRO 6000”; CoreWeave names the Server Edition, and Hyperstack lists an “RTX Pro 6000 SE”.

The Server Edition and the Max-Q both support Multi-Instance GPU (MIG), which splits the card into up to four isolated instances; NVIDIA’s Max-Q datasheet lists 4 × 24 GB, 2 × 48 GB or 1 × 96 GB. That is how some clouds sell a quarter or half of the card.

Two differences matter when you compare listings. The Max-Q runs at half the board power of the other two editions. The Server Edition has about 11% less memory bandwidth than the workstation cards. If a listing doesn’t say which edition it uses, ask before you compare its price with another.

RTX PRO 6000 price per hour in October 2026

We read each provider’s public price page on 15 October 2026. Prices are per GPU-hour and on demand unless stated, and exclude VAT and sales tax.

Vast.ai

Vast.ai is a marketplace: independent hosts list machines and set their own prices. ComputePrices tracked Vast.ai’s RTX PRO 6000 offers at $1.13 to $1.57 per GPU-hour on 15 October. Compute is billed by the second while an instance runs; storage is billed for as long as the instance exists, even when it is stopped; bandwidth is charged per byte, in both directions, at a price each host sets. Interruptible instances are “often 50%+ cheaper” than on-demand ones, in Vast.ai’s words, but can be paused.

RunPod

RunPod charges $2.49 an hour on Secure Cloud for one RTX PRO 6000 with 16 vCPUs and 188 GB of RAM, and lists $1.69 on Community Cloud; ComputePrices showed the Community Cloud offer as out of stock, at $1.64, the same day. RunPod also sells MIG slices of the card: 24 GB for $0.69 (Secure) or $0.50 (Community), 48 GB for $1.29 or $1.00. Storage is extra: $0.10 per GB-month for container disk, $0.10 for volume disk while the pod runs and $0.20 while it is idle, and $0.07 for network storage under 1 TB.

Hyperstack

Hyperstack lists its RTX Pro 6000 SE at $1.85 on demand, $1.48 on spot and from $1.30 reserved, with up to 31 CPUs and 180 GB of RAM per GPU. It does not charge for bandwidth or data transfer. GetDeploying showed Hyperstack’s on-demand and spot offers as sold out on 15 October.

CoreWeave

CoreWeave lists the RTX PRO 6000 Blackwell Server Edition in 8-GPU instances. The high-memory version, with 128 vCPUs, 1,024 GB of RAM and 7.68 TB of local storage, costs $20.00 an hour on demand, or $2.50 per GPU, and $11.09 an hour on spot in North America, or $1.39 per GPU. Data transfer to and from the internet is free.

Modal

Modal bills serverless GPUs by the second: $0.000842 a second for the RTX PRO 6000, or $3.03 an hour. CPU ($0.0000131 per physical core per second) and memory ($0.00000222 per GiB per second) are billed on top, and choosing a region multiplies prices by 1.15 to 1.75. All Modal functions can be preempted by default, and the non-preemptible option is not available for GPU functions. Network egress costs $0.04 per GiB beyond the monthly allowance: 1 TiB on the Starter plan, 10 TiB on Team.

Lambda

Lambda’s public on-demand price list did not include the RTX PRO 6000 when we checked on 15 October 2026.

What the aggregators show

GetDeploying covered 50 providers and 364 price points on 15 October 2026. Its medians were $2.33 per GPU-hour on demand, across 40 providers with an on-demand price, $1.29 on spot and $2.00 reserved; it converts monthly-only plans to an hourly rate. ComputePrices listed 76 offers from $0.79 to $8.05 per GPU-hour, a mix of on-demand, spot and committed prices, with a median of $1.99 by our count; it notes that its prices exclude CPU, disk and network factors. GetDeploying warns that a provider’s own page may quote a different figure.

About “$0.50 an hour”. The figure is real, but not for the whole card: on 15 October 2026 it was RunPod’s Community Cloud price for a 24 GB MIG slice, a quarter of the GPU.

Monthly vs hourly: the break-even math

The comparison is one division. The monthly price divided by an hourly rate gives the number of hours per month above which the monthly server costs less. A month has 730 hours on average: 8,760 hours a year divided by 12.

The same comparison by utilization, the share of the month the GPU is actually rented:

The monthly server costs $899 in every row; the last column divides that by the hours you actually use. Three readings:

• Light use favors hourly. At 25% utilization, the four hourly options cost $208 to $553 for the month.
• Always-on means 100%. An inference endpoint that stays up bills every hour it exists, whether it serves one request or a million. For that pattern, the comparison is the bottom row.
• The cheapest marketplace listings stay cheap. At $1.14 an hour, a full 730-hour month costs $832.20. Against those offers, the case for a monthly server rests on what comes with the GPU rather than on the GPU-hour, which the next section covers.

Costs the hourly rate leaves out

The rest of the machine

A GPU-hour price buys a share of a host as well. RunPod’s $2.49 includes 16 vCPUs and 188 GB of RAM, and Hyperstack allows up to 31 CPUs and 180 GB per GPU. On Modal, CPU and memory are separate meters: 8 physical cores and 64 GiB of memory add about $0.89 an hour at list price, taking the total to about $3.92. Our RTX PRO 6000 server includes a 32-core EPYC 9355P, 384 GB of DDR5 ECC and 2 × 3.84 TB of Gen5 NVMe, for one customer.

Storage that keeps billing

On RunPod, 500 GB of volume disk costs $50 a month while the pod runs and $100 while it sits idle. Vast.ai bills storage for as long as an instance exists, stopped or not, and Modal charges $0.09 per GiB-month for volumes beyond 1 TiB. Model weights are part of this: gpt-oss-120b alone is about 65 GB, so you either pay to keep it on a volume or download it again each time. Our gpt-oss-120b hardware guide covers its sizes.

Egress

Policies range from free to metered. Hyperstack and CoreWeave do not charge for data transfer. Modal charges $0.04 per GiB beyond the plan allowance, so sending 20 TiB out in a month on its Team plan costs $409.60 for the 10 TiB above the allowance. On Vast.ai, each host sets a per-byte price for traffic in both directions. A PowerDedi server has its own 10 Gbps unmetered port, with no quota and no overage fee.

Idle hours

Hourly billing runs while an instance is up, busy or not. Stopping it saves compute but, as above, not always storage. Budget on the hours the instance will exist, not on the hours it will be busy: for a service that has to answer at any time, those are all 730.

Availability and interruptions

A low list price helps only if the capacity exists. On 15 October 2026, GetDeploying showed several RTX PRO 6000 offers as out of stock or sold out, including Hyperstack’s on-demand and spot capacity, and ComputePrices showed RunPod’s Community Cloud offer as out of stock. Spot and interruptible capacity can be reclaimed, and on Modal every GPU function is preemptible.

When hourly billing is the better choice

Hourly billing is the right tool in several cases, including some where we can’t help:

• Short or bursty jobs. A weekend fine-tune, a benchmark or a demo uses a few dozen hours. Below 41% of the month, hourly is cheaper at every on-demand rate in the break-even table.
• Several GPUs at once. Tensor parallelism across 2, 4 or 8 GPUs needs a provider that sells multi-GPU machines: CoreWeave sells the RTX PRO 6000 by the 8-GPU instance, and Vast.ai lists 2-, 4- and 8-GPU offers. We sell one GPU per server, with no NVLink and no multi-GPU clusters.
• Immediate starts. On-demand capacity can be used as soon as it is available. Our RTX PRO 6000 and L40S servers are built to order in 72 hours, the RTX PRO 4000 in 24 hours.
• Interruptible batch work. If your job checkpoints and restarts cleanly, spot pricing helps: GetDeploying’s spot median was $1.29, and Vast.ai’s interruptible instances are often half price or less.
• A fraction of the card. If 24 GB is enough for your model, a MIG slice costs from $0.50 an hour.
• Regions and billing. Our GPU servers are in Amsterdam and Frankfurt, plus New York for the RTX PRO 4000, and we accept payment only in cryptocurrency: Bitcoin, Ethereum, Monero, USDT on TRON and Solana. If you need GPUs elsewhere or card and invoice billing, a cloud is the better fit.

L40S and RTX PRO 4000: the same math

The method is the same for our two other GPU servers, the L40S with 48 GB of GDDR6 and the RTX PRO 4000 Blackwell with 24 GB of GDDR7.

The L40S is widely rented: GetDeploying counted 37 providers, and its cheapest on-demand listing was $0.55 per GPU-hour. The RTX PRO 4000 is rarely rented by the hour; GetDeploying tracked six configurations from two providers.

What the monthly RTX PRO 6000 server includes

Our RTX PRO 6000 Blackwell server is a single bare-metal machine for one customer:

• One RTX PRO 6000 Blackwell with 96 GB of GDDR7, on an AMD EPYC 9355P host (32 cores) with 384 GB of DDR5 ECC and 2 × 3.84 TB of Gen5 NVMe.
• Its own 10 Gbps port, unmetered, with always-on network DDoS mitigation, IPMI/KVM console access, one IPv4 address and a routed /64 of IPv6.
• Amsterdam or Frankfurt, built to order in 72 hours, $899 a month with no setup fee, month to month.
• One GPU per server: no NVLink and no multi-GPU clusters. GPU mining is not permitted.

The three GPU configurations are compared on the GPU servers page, and our AI and GPU compute page covers the workloads they serve. The host around the GPU accounts for part of any GPU server’s price; our article on RAM prices in 2026 covers the memory side.

How we compared

• Date. All third-party prices were read on 15 October 2026 from public pages. They change often, so re-check them before you budget.
• Units. Per GPU-hour, on demand unless stated, in US dollars, excluding VAT, sales tax and setup fees. CoreWeave’s per-GPU figure is its 8-GPU instance price divided by eight.
• Hours. 730 hours per month. Our per-hour figures are the monthly price divided by 730, or by the hours actually used in the utilization table.
• Not included. Negotiated discounts, credits, storage and egress beyond the examples above, and performance differences between editions or hosts.
• Disclosure. PowerDedi sells monthly GPU servers, so we have an interest in this comparison. That is why every third-party price above comes from a public page listed in the sources, with the date we read it.

Sources

• RTX PRO 6000 Blackwell Server Edition — NVIDIA
• RTX PRO 6000 Blackwell Workstation Edition — NVIDIA
• NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition datasheet — NVIDIA
• GPU Cloud Pricing — Runpod
• Pricing — Vast.ai documentation
• Hyperstack AI Cloud Pricing — Hyperstack
• CoreWeave Cloud Pricing — CoreWeave
• Pricing: Pay per second for GPUs and CPUs — Modal
• Preemption — Modal documentation
• GPU cloud pricing — Lambda
• RTX PRO 6000 GPU Price — ComputePrices, checked 15 October 2026
• Nvidia RTX PRO 6000 Rental Prices — GetDeploying, 15 October 2026
• Nvidia L40S Rental Prices — GetDeploying, 15 October 2026
• Nvidia RTX PRO 4000 Rental Prices — GetDeploying, 15 October 2026