Where to rent an RTX PRO 6000 at the cheapest price

Three cards share the RTX PRO 6000 name, the same chip and the same 96 GB, but not the same speed. The Workstation Edition runs at up to 600 W with 1,792 GB/s, the Max-Q keeps that bandwidth under a 300 W cap with a blower cooler made for four cards in one tower, and the Server Edition is passive, configurable from 400 to 600 W, and rated at 1,597 GB/s^9111213. Two limits follow. Bandwidth sets the ceiling for single-stream token generation, so the Server Edition is up to about 11 percent slower than the other two on a one-user chat with a large model. Power sets the speed of compute-bound work such as diffusion, rendering and batched prefill, so the Max-Q gives up the most there: Puget Systems measured it 5 to 14 percent slower than the 600 W card in most content-creation tests, and 14 percent behind on Unreal Engine and DaVinci Resolve GPU effects⁴⁷.

The table of fits shows why so many renters pick this card despite the edition question. Weights alone take about 2 bytes per parameter in 16-bit, 1.06 in 8-bit and 0.58 in 4-bit, and a model fits comfortably when they leave room for activations and cache. A 70B model fits in 8-bit with a modest cache and in 4-bit with room to spare, gpt-oss-120b fits in its native 4-bit form, Qwen3 32B, FLUX.1 dev and Wan 2.2's two 14B experts fit at any precision, and Mixtral 8x22B fits only in 4-bit, with about 14 GB left for context³¹³².

NVIDIA's RTX PRO 6000 Blackwell carries 96 GB of GDDR7 with ECC, fifth-generation Tensor Cores that compute in FP4, and a PCIe Gen 5 host interface, and it has no NVLink bridge, so work spread over several cards talks over PCIe⁹. It is the GB202 die of the GeForce RTX 5090 with more of it switched on: 24,064 CUDA cores, 752 Tensor Cores, 188 RT cores and a 512-bit memory bus, with three times the memory¹⁰. Multi-Instance GPU can split one card into four 24 GB or two 48 GB slices, which is why Runpod also sells 48 GB and 24 GB slices at $1.29 and $0.69 an hour, and why a cheap listing is worth checking for being a whole card⁵. It launched in 2025 in three editions built around the same chip and the same memory. The Workstation Edition has a double flow-through cooler rated at 600 W and 1,792 GB/s of memory bandwidth, the Max-Q Workstation Edition keeps the bandwidth but is capped at 300 W, and the Server Edition, passively cooled for data-centre chassis, is configurable from 400 to 600 W and rated by NVIDIA at 1,597 GB/s and 4 PFLOPS of FP4¹¹¹²¹³. That one-ninth difference in bandwidth and the factor of two in power are why the edition matters, a point the section at the end of this page works through.

Being a professional card is what puts it in clouds at all. NVIDIA's GeForce licence rules out data-centre deployment of consumer cards such as the RTX 5090, while the RTX PRO line is built and licensed for it¹⁴, which is why AWS sells it in its G7e instances, Google in G4, Oracle in an 8-GPU bare-metal shape and CoreWeave in 8-GPU nodes, alongside specialised clouds and marketplaces^15161718. The people renting it are mostly teams that want a 70B-class model or gpt-oss-120b on one card instead of two or four smaller ones, creators running Wan 2.2 video and large image models without offloading weights to system RAM, and fine-tuners who want 96 GB for LoRA or QLoRA on large models. The price range follows from that mix: the same card is sold as a shared slice, a marketplace listing, a bid, a reservation, a serverless worker and one-eighth of a hyperscaler server, and each of those carries a different price for reasons that have little to do with the silicon.

A managed on-demand cloud sells a fixed machine, the GPU with its vCPUs, memory and local disk, at one hourly price, and the differences between clouds lie in what else that price covers. GPUBank includes the IPv4 address and all traffic and bills only a persistent disk on top³, Beam includes 16 vCPU, 128 GB of RAM and a 1 TB NVMe drive², Hyperstack bills only the GPU line with CPU and RAM included up to its stated maxima¹⁹, and Runpod adds container and volume disk charges⁵. A cheaper managed hour can still produce a larger bill, so the cost section below prices whole months rather than hours. One offer does not fit any of the usual boxes: packet.ai's $0.66 Dynamic tier is a full card's memory on a card shared between tenants under scheduler isolation, a different product from a dedicated hour at the same price; its page includes up to 2 TB of local NVMe per node and charges no egress, but publishes no price for storage that outlives an instance²⁰.

On a peer-to-peer marketplace independent hosts post their own GPU rate and add disk and bandwidth as separate lines, and the machine changes from one listing to the next. On Vast.ai on 16 October the cheapest listing was a 300 W Max-Q at $0.95 with 8 vCPU on a PCIe Gen 4 host¹. Spot or interruptible capacity is cheaper because it can be taken back: Vast.ai bids start at $0.33, GPUBank's interruptible tier is $0.69 with five minutes of notice, and Nebius floats from $0.79¹³⁷. Reserved capacity runs from week-long terms to three-year commitments and is often sold by the 2-GPU or 8-GPU server, serverless workers bill active seconds at two to three times a dedicated hour, and the hyperscalers price the card at $3.36 to $5.50 per GPU, usually as part of a larger machine.

The captures below were taken on 16 Oct 2026 from each provider's own RTX PRO 6000 or pricing page, so every figure can be read next to the conditions printed beside it.

GPUBank rents the RTX PRO 6000 at $1.15 per GPU-hour on demand, $0.69 interruptible and $0.92 reserved for a week or a month, each with 16 vCPU, 188 GB of RAM, a 2 TB NVMe disk, one IPv4 address and all inbound and outbound traffic, in machines of 1, 2, 4 or 8 GPUs³. Billing is per second with no minimum, the meter starts after the health check, a failed launch is not billed, and the only extra line is a persistent disk at $0.03 per GB-month. Payment is a prepaid balance topped up in crypto from $25, and the top-up bonus brings the on-demand hour to about $0.96 from $25, $0.85 from $100 and $0.72 from $1,000³. It is listed in Paris, Ashburn and Singapore, and the median launch to a machine that has passed its health check takes 48 seconds²⁹. Where it wins is the all-in month: no egress, no IP line and a cheap stopped disk made it the cheapest dedicated option in all three scenarios below, beaten on paper only by a shared card or a floating spot floor before their storage is counted, or by a marketplace bid whose disk is deleted between runs.

Where it loses is the headline hour. Vast.ai bids at $0.33 and Max-Q listings at $0.95, packet.ai's shared $0.66, HyperAI's $0.80 and Beam's $1.09 dedicated machine are all below its $1.15 list price, and its interruptible tier at $0.69 is twice the cheapest Vast.ai bid¹²⁰²¹². Its reservation at $0.92 needs only a week but is prepaid and cannot be refunded once it starts. It takes no cards for now, which rules it out for a buyer who can only pay by card. And its pricing page lists 96 GB GDDR7 ECC at 1.8 TB/s without naming the edition or power limit behind the rental, so a buyer whose job depends on sustained power should ask before reserving³.

Several offers below GPUBank's $1.15 list price could only be partly checked. packet.ai's $0.66 is on its own page but sits on a shared card, with the dedicated tier listed as launching soon²⁰. HyperAI's pricing page lists the card at $0.80 an hour without stating the vCPU, RAM or disk that come with it, and GetDeploying shows a one-month reservation at $0.48 that we could not find on HyperAI's site²¹⁴. Hostinger at $0.65 and GPUhub at $1.10 appear only on GetDeploying, which saw GPUhub in stock with 22 vCPU and 110 GB of RAM on 15 October⁴. Clore.ai's own documentation carries a dated snapshot of 24 August 2026 that put the card at $0.92 to $1.38 with a median of $1.21, from only four free servers, and no live price could be read since²². Lium's Server Edition pods were asking $1.60 to $1.99 per GPU on 8-GPU machines, billed per second²³.

Above GPUBank's list price a group of European and US clouds sells dedicated single cards between $1.74 and $2.41: Nova Cloud at $1.74, UpCloud at $1.88, EcoHash at $1.89, Cloudzy at $1.93, Lyceum at $1.99, JarvisLabs at $2.19, Koyeb at $2.20 and Sesterce at $2.41, all from GetDeploying's crawl of 15 October, and Verda at $2.30 on its own page with 30 CPU threads and 90 GB of RAM per GPU⁴⁶²⁴. Massed Compute's own page lists the card at $2.19 with 16 vCPU, 144 GB of RAM and 725 GB of storage, and Exoscale sells it at $2.00 per GPU only as an eight-card node²⁵⁴. Their case rests on something other than price: a specific European jurisdiction, card payment, or an existing account. On the spot side, Verda's interruptible instances cost $1.15 and its serverless containers $2.53, or $1.27 on spot, while CoreWeave sells 8-GPU Server Edition spot nodes at $9.56 an hour, $1.195 per GPU⁶¹⁸. CloudRift's own page quotes the card only reserved, $1.16 for a month and $1.10 for three, with on-demand capacity on request²⁶. The long-term reservations seen on aggregators, Leaseweb at $0.88 per GPU on a 2-GPU server for a month, Database Mart at $0.90 on a 24-month term and Utho at $1.18 on 36 months, need a quote before anyone plans around them⁴.

AWS and CoreWeave name the Server Edition in their offers, and the large clouds cost two to five times a specialised cloud at list price. AWS sells it in G7e instances from the g7e.2xlarge, one GPU with 8 vCPU and 64 GiB, at $3.363 an hour on demand and about $2.56 on spot, up to the 8-GPU g7e.48xlarge with 192 vCPU and 2 TiB at $33.14, or $4.14 per GPU, with EBS storage and egress billed on top¹⁵²⁷²⁸. Google's g4-standard-48 gives one GPU with 48 vCPU and 180 GB for $4.50 an hour in us-central1, $3.105 on a one-year commitment, $1.98 on three years and about $1.71 on spot¹⁶. GetDeploying lists Azure's full-GPU size at $5.50 per GPU and Oracle's BM.GPU.RTXPRO.8 at $4.50 per GPU, the Oracle shape being a bare-metal server with 8 cards, 144 OCPU and 3 TB of RAM⁴¹⁷.

CoreWeave's 8-GPU Server Edition node with 128 vCPU, 1,024 GB of RAM and 7.68 TB of local storage costs $20.00 an hour on its own page, which is $2.50 per GPU, the cheapest of this group but only by the whole node¹⁸. These clouds make sense when the rest of a stack already runs there, when enterprise discounts bring the effective rate down, or when compliance demands a particular provider. For a team that only needs the card, the hyperscaler premium buys nothing the card itself does faster, and the per-GPU price of an 8-GPU instance is only reachable by renting all eight.

Storage is the line that most often overturns an hourly comparison, because model weights for this card are large and live on disk between sessions. A 500 GB persistent disk costs $15.00 a month on GPUBank and nothing else is billed while the machine is stopped³. The same 500 GB costs about $35.32 a month on Hyperstack at $0.000096774 per GB-hour¹⁹, and on Runpod a volume disk costs $0.10 per GB-month while the pod runs and $0.20 while it is idle, so a mostly idle 500 GB volume approaches $100 a month⁵. Vast.ai hosts set their own rate, from $0.13 to $0.87 per GB-month on the RTX PRO 6000 listings we read, and storage keeps billing while an instance is stopped¹³⁰. Traffic is the second line: AWS bills egress at about $0.09 per GB, so a video service sending 5 TB a month would pay roughly $441 for it after the free first 100 GB, more than half the card's own month on GPUBank²⁰. Google bills it too and Vast.ai hosts charge up to $0.04 per GB, while GPUBank, packet.ai and CoreWeave state none³²⁰¹.

Billing granularity decides the price of short work. A 61-minute job, 3,660 seconds at $0.000319, costs $1.17 on GPUBank, while the same rate billed per started hour costs $2.30 because the sixty-first minute buys a second hour. packet.ai bills hourly²⁰, Runpod and Vast.ai per second⁵¹, and the hyperscalers per second after a one-minute minimum. Deposits and credit matter too: GPUBank's minimum top-up is $25 and its balance never expires, Beam's cheapest plan is $20 a month with $20 of credit², and a top-up bonus is a real discount only on money you will spend. The last cost is lost time. GPUBank restarts a failed host's machine on a healthy one and credits the interrupted session, and a missed spot notice credits 30 minutes³. On a marketplace the hours lost to a host that drops offline, or to a Max-Q sold as a generic card, are usually yours. The scenarios below keep the disk all month, and where a storage price could not be read it is set to zero, which favours that provider.

On a model that fits in 32 GB, the RTX 5090 wins per job. CloudRift's Qwen3-Coder-30B AWQ run gave 4,570 tokens per second on one RTX 5090 against 8,425 on the RTX PRO 6000³⁴; at GPUBank's $0.48 and $1.15 an hour that is $0.029 per million tokens on the 5090 and $0.038 on the RTX PRO 6000³. Trooper.ai's automated runs tell the same story on images, 8.3 seconds per SD3.5 Large image against 12 on the 5090, which is $0.0027 against $0.0016 per image at the same prices³⁹. The larger card earns its price only when the model would otherwise be offloaded, quantised harder or split across cards. Llama 3.3 70B in 4-bit, about 41 GB of weights, does not fit on one 5090, and CloudRift's two-5090 setup at $0.96 an hour on GPUBank produced 1,230 tokens per second against 1,031 for one RTX PRO 6000 at $1.15: close to a tie per token, with the single card keeping 50 GB free for cache and avoiding pipeline parallelism. The L40S at $0.55 has 48 GB and 864 GB/s, enough for a 4-bit 70B model with almost no room for context, and the only shared test we found is Runpod's single-user vLLM run, where the L40S decoded 48 tokens a second against 89 on the RTX PRO 6000⁴⁰.

Against the 80 GB datacenter cards the measurements are few, and they split by workload. CloudRift ran GLM-4.5-Air at 4-bit under heavy batched load and measured 3,140 tokens a second on one RTX PRO 6000 against 2,987 on one H100, so at high concurrency the cheaper card holds its own⁴¹. One user at a time is a different matter: Runpod's vLLM test at one request per second gave 89 tokens a second on the RTX PRO 6000, 96 on the A100 SXM, 109 on the H100 PCIe and 165 on the H100 SXM, which at GPUBank's prices is $3.59 per million tokens against $2.75, $3.44 and $3.11⁴⁰³. Spread over several cards the lack of NVLink shows: in the same CloudRift study four H100s served Qwen3-Coder-480B about 31 percent faster than four RTX PRO 6000s, and eight H100s ran GLM-4.6 in FP8 nearly three times as fast as eight of them⁴¹. For a team serving one large model to many users, or generating video, the RTX PRO 6000's 16 extra gigabytes and lower price usually decide it; for multi-GPU work or one-user latency on a 70B model, the H100 SXM at $1.85 is the better hour.