Where to rent a B200 SXM at the cheapest price

The chart below uses one MLPerf round, Inference v5.0, so that the B200, H200 and H100 rows were measured on the same benchmark rules and the same TensorRT release⁴¹. The language-model rows compare each card as you would deploy it, FP4 on the B200 against FP8 on Hopper, so they show the platform's gap rather than the silicon's alone, while the Stable Diffusion XL row, at FP8 on all three, is the like-for-like test and its 1.6 times the realistic floor of what a B200 hour buys over an H100 hour.

The B200 is NVIDIA's first Blackwell datacenter GPU, and in its SXM form it is not a card you can slot into a server: it ships on the 8-GPU HGX B200 baseboard that goes into NVIDIA's DGX B200 and into OEM servers from Supermicro, Dell, HPE, Lenovo and others. NVIDIA's DGX B200 page gives 1,440 GB of HBM3e and 64 TB/s of memory bandwidth for the eight GPUs, which is 180 GB and 8 TB/s per GPU, 14.4 TB/s of aggregate NVLink bandwidth (1.8 TB/s per GPU over fifth-generation NVLink), two Xeon Platinum 8570 processors, 2 TB of system memory and a maximum draw of about 14.3 kW⁹. Its FP4 figure of 144 petaFLOPS and FP8 figure of 72 petaFLOPS for the whole system are sparse numbers, and dense throughput is half of each. NVIDIA's MLPerf write-up describes each GPU running at a 1,000 W thermal design power¹⁰.

You will see both 180 GB and 192 GB quoted. The 192 GB figure comes from early HGX B200 material that some providers still reproduce, including packet.ai, Hyperstack and Spheron on their pricing pages, while shipping DGX systems and every cloud that states the usable memory, from Lambda and Verda to Deep Infra and Oracle, expose 180 GB to software^9117612. Size models, KV caches and batches on 180 GB per GPU. The renters fall into three groups: training teams that want whole nodes or multi-node clusters with InfiniBand for weeks or months, inference operators who serve large models in FP8 or FP4 on one to eight GPUs, and engineers who want one card for a few hours to test or fine-tune a model that does not fit on an H100. For years the market served only the first group, which is why so many B200 offers still start at eight GPUs, and the single-GPU on-demand offers that appeared in 2026 are where the cheapest entry prices now sit.

Every price on this page is in US dollars per GPU-hour excluding VAT, and where a provider sells only an 8-GPU node we divide the node price by eight. That gives a fair per-GPU number, but you still pay for eight, so the minimum unit matters as much as the rate. A managed cloud sells a fixed machine at one hourly price, and the question is what that price leaves out: Nebius bills disks separately, Northflank bills CPU and memory as their own lines, and AWS bills EBS volumes and egress on top, while GPUBank, Beam and Deep Infra put the CPU, memory, local disk and traffic inside the hour^1135146. A marketplace lets independent hosts set their own price on machines that vary from listing to listing, and on the B200 the marketplaces were not cheaper than the cheapest managed offers in mid-October, which reverses the pattern seen on consumer cards.

The other three kinds of offer trade something for the price. Spot or preemptible capacity is the cheapest hour on the market, from Nebius's dynamic floor of $0.99 to published fixed tiers between $2.49 and the $5.25 Spheron showed on GetDeploying, and the terms that matter are the notice period and whether the disk survives¹⁵¹⁵. Reserved capacity brings the B200 to between $2.99 and $4.49 when the contract runs one to four years, and to $6.79 to $9.86 on the short cluster reservations of Together AI and Lambda¹⁶¹⁷¹¹. Serverless platforms bill by the second of work from $5.50 to about $10 and suit bursty inference rather than long jobs¹⁸¹⁹. Stock is the last filter: on 15 October GetDeploying marked Lambda, Massed Compute, Hyperstack's on-demand flavour, Sesterce and Runpod Community out of stock, and Nebius's default B200 quota for VMs without a reservation is zero¹⁵³.

The captures below were taken on 16 Oct 2026 from each provider's own B200 or pricing page, so every figure can be read next to the conditions printed beside it.

The provider card above gives GPUBank's rates; what matters beside them is how the hour behaves. The meter starts only once a machine has passed its health check, which GPUBank gives as a median of 48 seconds from launch, a failed launch is never billed, and a stopped machine costs only its persistent disk at $0.03 per GB-month⁵. Machines come with 1, 2, 4 or 8 GPUs and the vCPU, RAM and NVMe scale by the same multiple, any Docker image can be launched, and ready-made templates cover vLLM, SGLang, Ollama, Open WebUI, llama.cpp, Unsloth, Axolotl, JupyterLab and PyTorch 2.8 with CUDA 12.8, the stacks most B200 renters reach for first. Every account and every team member can be given a monthly spend cap, which is the simplest guard against an 8-GPU machine left running over a weekend at $33.20 an hour.

Where GPUBank wins: its $2.49 is the lowest fixed spot price with published terms we found, its $3.32 week or month is lower than every short reservation we could read, and on demand only Deep Infra's $3.69 container and packet.ai's shared $3.75 tier undercut it at list, with Beam close at $4.11⁶²⁰¹⁴. Payment is a prepaid balance topped up in crypto from $25 to $5,000, and the top-up bonus of 20 to 60 percent brings the effective on-demand hour to about $3.46 with $25, $3.07 with $100 and $2.59 with $1,000⁵. Where it loses: Nebius's dynamic spot can be far cheaper when capacity is idle, Vultr's and Civo's multi-year contracts are lower per hour, Deep Infra is cheaper at list for one GPU, there are only two regions, machines stop at eight GPUs so multi-node training needs a cluster provider, and card payments are not accepted for now, which rules GPUBank out for some companies whatever the price.

Below GPUBank's list price, packet.ai sells a B200 Dynamic tier at $3.75 in US East, which its own pricing page describes as shared GPU pods on scheduler-isolated infrastructure where the physical GPU may be shared with other tenants, and a Dedicated pod at $6.99²⁰. Sesterce appeared at $4.11 for an 8-GPU node and Hostinger at $4.50 for one GPU on GetDeploying, the first out of stock and the second with stock not reported, and neither price could be read on the provider's own page¹⁵. The same aggregator listed Lyceum's single B200 at $6.49 and its spot tier at $3.29, both out of stock, Daytona at $3.59 for the GPU alone on a waitlist, and Spheron and EmpirioLabs in stock but at $11.06 and $11.90 per GPU, close to the hyperscaler rate¹⁵. GMI Cloud now lists the B200 from $5.00 with limited availability through sales, up from $4.00 in September²¹. Massed Compute sells an 8-GPU node at $43.46 an hour, or $5.43 per GPU, with 280 vCPU and 2,800 GB of RAM, Northflank gives $5.87 for the GPU line alone, and Hyperstack's 8-GPU flavour is $6.00, marked out of stock^22132315.

The better-known neoclouds sit between $6.69 and $8.60. Lambda's own page lists the B200 at $6.69 per GPU on an 8-GPU instance and $6.99 on one GPU, both out of stock on GetDeploying, Runpod's Secure Cloud is $7.99 with 28 vCPU and 283 GB of RAM, although several of Runpod's own articles still quote $6.79, E2E Networks in India is $7.99 billed per minute, Together AI's clusters are $8.19, and Crusoe and CoreWeave sell 8-GPU nodes at $8.50 and $8.60, CoreWeave's with 128 vCPU, 2,048 GB of RAM and 61.44 TB of local storage^112425172627. Shadeform's directory quoted the B200 from $6.52 with no live instance listed, and the $3.74 Boost Run node that our September survey read there is no longer visible²⁸. UpCloud's page announces new B200 servers from 20 October at €4.50 an hour on demand and €3.38 on spot for one GPU, billed by the started hour²⁹.

On the marketplaces, Akash quotes a fixed $5.00 for the GPU alone, with CPU, memory, storage and the IP lease priced in each provider's bid, a figure we could read only on ComputePrices because Akash's own page loaded without its table³⁰. Runpod Community lists $5.98, out of stock, and Vast.ai's B200 page said "No current offers" while ComputePrices recorded $6.25 listings the same day; on Vast, storage keeps billing while an instance is stopped and bandwidth is billed per byte^24313032. Other spot tiers include Together AI's preemptible $4.09, CoreWeave's $34.11 node ($4.26 per GPU) and Vultr's $3.20 node, out of stock¹⁷²⁷¹⁵. Among serverless offers, Koyeb is $5.50 with 15 vCPU and 240 GB of RAM, Cerebrium $6.01, Modal $6.25 for the GPU alone, fal $7.99 at list, Runpod $8.64 and Baseten $9.98^18333482419.

The large clouds sell the B200 only as 8-GPU instances, and their per-GPU price is two to four times the neocloud rate. AWS's p6-b200.48xlarge, with 192 vCPU, 2,048 GiB of RAM and eight 3.84 TB NVMe drives, is $113.9328 an hour on demand in US East, or $14.24 per GPU, on AWS's own price list³⁵. Its Capacity Block, which reserves the instance for chosen dates and is paid up front, is $113.666 per instance-hour, or $14.21 per GPU, in Ohio, N. Virginia, Oregon, Mumbai, Hyderabad, Stockholm and Paris, up from $12.36 in our September survey, with AWS's next scheduled price change in January 2027, so a reserved block now costs about the same as on demand and buys only certainty that the capacity exists on your dates³⁶. Oracle's price list gives $14.00 per B200 GPU-hour for the BM.GPU.B200.8 bare-metal shape¹².

Google Cloud does not sell its A4 machine on ordinary on-demand terms: its documentation says you must reserve capacity, use Spot or Flex-start VMs, or create a resize request in a managed instance group³⁷. On Google's own pricing page the a4-highgpu-8g, with 224 vCPU, 3,968 GB of RAM and 12,000 GiB of local SSD, costs $39.6336 an hour on Spot ($4.95 per GPU), $64.44 on Flex-start ($8.06), $90.22 in calendar mode, $88.9272 on a one-year commitment ($11.12) and $56.7072 on three years ($7.09)³⁸. Microsoft's retail price list carries no ND B200 v6 SKU at all; the nearest Azure product is the ND GB200 v6, a four-GPU Grace Blackwell VM at $108.16 an hour in East US, which is $27.04 per GPU³⁹. What the premium buys is the surrounding platform, from IAM and private networking to compliance programmes, and teams that do not need it pay half or less elsewhere.

Next to the B200 sits the GB200 NVL72, a rack that joins 72 Blackwell GPUs and Grace CPUs in a single NVLink domain. CoreWeave lists GB200 NVL72 instances at $42.00 an hour for four GPUs, or $10.50 per GPU, and AWS's u-p6e-gb200x72 UltraServer costs $761.904 an hour on a Capacity Block in its Dallas Local Zone, or $10.58 per GPU²⁷³⁶. That is about two and a half times a single B200 on the cheaper managed clouds, and it buys the one thing an HGX B200 cannot give: NVLink bandwidth across 72 GPUs instead of eight, which large mixture-of-experts training and very large tensor-parallel inference need. For any job that fits in one 8-GPU node, the rack is the expensive way to rent Blackwell.

The B200 earns its price on three kinds of work: serving large models in FP4 or FP8, where 8 TB/s of bandwidth and native FP4 multiply tokens per second; training and full fine-tuning from about 30 billion parameters up, where 1.44 TB of HBM in one node keeps the job on one machine; and long-context inference, where 180 GB per GPU holds a KV cache that would need two H100s. It is the wrong tool for LoRA on small models, image generation at moderate volume, transcription, embeddings and anything that fits comfortably in 80 GB, where an H100 does the same work for less money per output. Our rule: if a job uses less than half of an H100's memory and is not throughput-bound, a B200 costs more and finishes only a little sooner.

MLPerf is the only cross-vendor source of B200 numbers we trust, and it tests whole 8-GPU systems at full load. In MLPerf Inference v5.0, closed division, NVIDIA's DGX B200 (entry 5.0-0056) served Llama 2 70B at 98,443 tokens per second in the server scenario, against 33,072 for NVIDIA's 8-GPU H200 system (5.0-0060) and 31,106 for its DGX H100 (5.0-0057), all on TensorRT 10.8 and CUDA 12.8⁴¹⁴². Under the interactive variant, which caps time to first token at 450 ms and time per output token at 40 ms, the figures were 59,409, 19,424 and 14,009 (HPE, 5.0-0039)⁴¹. The B200 ran FP4 weights and Hopper ran FP8, so part of the threefold gap is the lower precision rather than the chip: on Stable Diffusion XL, where all three ran FP8, the B200 produced 30.4 images per second against 19.0 and 18.4, a gap of about 1.6 times. Rented nodes reach the same figures: in MLPerf Inference v5.1 Nebius's HGX B200 (entry 5.1-0077) served 101,611 tokens per second in the server scenario and Lambda's 99,994, on TensorRT 10.11 and CUDA 12.9, and NVIDIA's data page shows 102,398 in a later round, so software has added only a few percent since v5.0⁴³⁴⁴⁴⁵.

Training results are thinner. NVIDIA's MLPerf Training v4.1 write-up reports 2x per GPU for GPT-3 pre-training and 2.2x for Llama 2 70B LoRA fine-tuning from DGX B200 submissions in the preview category, measured against an HGX H100 submission¹⁰. Those are NVIDIA's own per-GPU ratios from a round in which the B200 was not yet a shipping product, and later rounds have mostly reported GB200 NVL72 systems: NVIDIA's Training v5.0 write-up claims up to 2.6 times Hopper per GPU, but on racks that run at a higher power envelope than an HGX B200⁴⁶. Treat 2 to 2.2 times as the measured range for one HGX node and anything above it as a GB200 result.

Per million output tokens on a 70B model at full batch, the B200 is the cheapest card GPUBank rents. The v5.0 server result is 12,305 tokens per second per GPU, or 44.3 million an hour, which at $4.15 is $0.094 per million and $0.059 at the $2.59 effective rate⁴¹. The H200 at 4,134 tokens per second per GPU makes 14.9 million an hour, $0.167 per million at $2.49, and the H100 at 3,888 makes 14.0 million, $0.132 per million at $1.85. On Llama 3.1 405B offline the gap holds: $6.04 per million tokens on the B200 against $9.87 on the H200 and $9.42 on the H100 (QCT, 5.0-0065). For image generation it reverses: at the SDXL rates above, 1,000 images cost $0.30 on the B200, $0.29 on the H200 and $0.22 on the H100, because FP8 diffusion does not use Blackwell's FP4 advantage.

For training the answer is closer. A fine-tune that takes 80 GPU-hours on H100s costs $148.00 at $1.85; at NVIDIA's measured 2.2x it needs 36.4 B200 GPU-hours, $150.91 at $4.15, so the money is about the same and the job finishes in less than half the time¹⁰. If the 2.5 times NVIDIA reported for eight Blackwell GPUs in a later Training round held on an HGX node, the job would take 32 B200 GPU-hours and cost $132.80, cheaper than the H100 as well⁴⁶. We have no measured H200 ratio from the same round; at H100 speed the same 80 GPU-hours on the H200 would cost $199.20 at $2.49, so the B200 is cheaper per training job than the H200 unless the H200 trains at least 1.32 times faster than the H100. The MI300X, at $1.65 with 192 GB, has no single-node MLPerf v5.0 result on these benchmarks (its only Llama 2 70B entry that round used four nodes), so we do not compute its cost per job; it is the cheaper way to hold a large model in memory, and the B200 is the faster way to push tokens through it.