NVIDIA B200 (Blackwell) GPU Rental Price in September 2026: Starting from $5/Hour

By Sandeep Narahari, Contributor. Last updated: September 2026. Prices checked September 24, 2026.

On Akash, the B200 GPU rental price starts from $5.00 per GPU-hour as of September 2026, according to Akash GPU pricing. For that you get one NVIDIA B200 with 180GB of HBM3e and 8 TB/s of memory bandwidth, billed on demand with no long-term contract.

TL;DR

• $5.00 per GPU-hour on Akash: At the rate used in this article, one B200 costs $120 a day or $3,600 a month running nonstop. A full 8-GPU configuration costs $40 an hour, or $28,800 a month. Akash’s GPU pricing is dynamic, so verify the current rate before deploying.
• Trillion-parameter models can fit on one 8x B200 node by weight capacity: Kimi K2.6 (1T parameters), GLM-5.2 (753B), and DeepSeek-V4-Pro (1.6T) have published checkpoints that fit within 1,440GB of B200 HBM. Kimi K3 (2.8T) exceeds that capacity and requires a multi-node configuration.
• Lowest bandwidth-adjusted cost of the four GPUs compared: At the rental rates used here, a B200 costs $0.625 per TB/s of peak memory bandwidth per hour, below the H100, H200, and B300 in this comparison. For memory-bandwidth-bound serving, that makes it a candidate for lower cost per token, but actual token economics require benchmarking.
• B300 uses fewer GPUs for memory-heavy workloads; B200 offers more bandwidth per dollar: B300’s 288GB per GPU can reduce the number of GPUs required for some large models. When comparing configurations with more B200 GPUs, the B200 can provide more aggregate memory bandwidth per dollar, but actual throughput depends on the model and serving stack.
• 64.9% below AWS list price: At the rates used here, $5.00 per B200 GPU-hour compares with about $14.24 per GPU-hour for the AWS P6-B200 list rate, while Google Cloud’s A4 8-GPU B200 configuration is $64.44 per hour, or about $8.06 per GPU-hour.

How much does a B200 cost per hour, month, and year on Akash?

A B200 on Akash costs $5.00 an hour, $3,600 a month, and $43,800 a year if it runs nonstop. An 8-GPU node costs $40 an hour, $28,800 a month, and $350,400 a year.

Rate: $5.00 per GPU-hour on demand, per Akash GPU pricing, checked September 24, 2026.

Few workloads run nonstop, though. If you serve a model and scale it down overnight so each GPU runs 16 hours a day, that’s 480 hours a month: $2,400 per GPU instead of $3,600, or $19,200 for a full node. Because billing is on demand, you aren’t charged for hours when the deployment is shut down.

Can one 8x B200 node run a trillion-parameter model?

Yes - an 8x B200 node can hold the published weights of several trillion-parameter-class models in low-precision formats. An 8-GPU B200 system provides 1,440GB of HBM3e memory, with 180GB per GPU. NVIDIA rates the 8-GPU DGX B200 system at up to 144 PFLOPS of FP4 tensor compute.

That is enough HBM capacity for models such as Kimi K2.6 (1T parameters), GLM-5.2 (753B), and DeepSeek-V4-Pro (1.6T) at their published checkpoint precisions, based on their listed checkpoint sizes. Kimi K3, at about 2.8T parameters and roughly 1.56TB for its published MXFP4 checkpoint, exceeds the 1.44TB HBM capacity of a single 8x B200 node and therefore requires a multi-node configuration.

The important distinction is that model weights are only part of the memory requirement. Actual inference also needs memory for the KV cache, activations, framework overhead, and other runtime allocations. Those requirements vary with context length, concurrency, model architecture, quantization, and serving framework. There is no universal 25 percent multiplier that applies to every model.

Weight-only B200 capacity needed is checkpoint size divided by 180GB per B200, rounded up. Example B200 configuration is a practical starting point rather than a universal minimum; the correct GPU count depends on the model, serving framework, context length, concurrency, and parallelism strategy.

Checkpoint and model information comes from the respective model cards: DeepSeek-V4-Flash, Qwen3.5-397B-A17B-FP8, Kimi K2.6, GLM-5.2-FP8, DeepSeek-V4-Pro, and Kimi K3. NVIDIA lists 180GB HBM3e per B200 GPU and 1.44TB across an 8-GPU B200 system.

Monthly cost assumes 720 hours at $5.00 per B200 GPU-hour. Actual Akash pricing can vary with marketplace conditions and provider offers.

Why low-precision models fit on B200

Low-precision weights dramatically reduce the memory required to store large models. Kimi K2.6 is distributed as an INT4 model, while DeepSeek-V4-Pro uses a mixed low-precision format. That allows models with more than one trillion total parameters to fit within a single 8x B200 node by weight capacity.

The B200 is designed for low-precision AI workloads and supports FP4 Tensor Core computation. NVIDIA rates an 8-GPU DGX B200 system at up to 144 PFLOPS of FP4 performance (DGX B200 specs).

However, fitting the weights does not guarantee that the model will run comfortably at every context length or concurrency level. For production inference, you need to benchmark the model with your actual serving configuration and account for KV-cache and runtime memory.

What matters when sizing B200s

For MoE models, total parameters primarily determine model-weight memory, while active parameters influence the amount of computation performed for each token. This is why a model with a very large total parameter count can still be practical to serve when only a smaller subset of parameters is active for each token. See total vs active parameters and GPU sizing for the full breakdown.

For multi-GPU and multi-node deployments, the serving framework also matters. Tensor-parallel size is constrained by the model architecture and implementation rather than by a universal requirement to use powers of two. Common configurations include 1, 2, 4, and 8 GPUs, but the correct value should be verified for the specific model and framework.

The practical takeaway: an 8x B200 node can run many trillion-parameter-class models when they are sufficiently quantized, but the number of GPUs you actually need depends on more than the model’s parameter count.

Is a B200 or B300 cheaper for trillion-parameter models?

It depends on whether you’re paying for hours or for throughput. A B300 has 288GB per GPU versus 180GB on B200, so it can fit some models on fewer GPUs and lower the total hourly bill. A B200 configuration provides more aggregate memory bandwidth when you use more GPUs, which may improve cost per token for bandwidth-bound serving.

Same sizing method as the table above. Akash rates used here: B200 $5.00, B300 $6.00, H200 $4.45 per GPU-hour. Actual marketplace prices can vary.

At these example configurations, the B300 has the lower hourly GPU cost for every model in the table because its larger memory capacity can reduce the number of GPUs required.

Throughput tells a different story. Eight B200s provide 64 TB/s of combined peak memory bandwidth for $40/hour. Four B300s provide 32 TB/s for $24/hour. Decode-heavy LLM inference can be strongly influenced by memory bandwidth (here’s why), so the B200 configuration may offer better throughput per dollar in some workloads. However, actual results depend on the model, batch size, parallelism, and serving stack, while B300 also provides Blackwell Ultra architectural improvements.

As a starting point: for a memory-capacity-constrained job, B300 can lower the hourly bill; for high-throughput serving, benchmark both configurations before committing. For more on the larger-memory GPU, see B300 rental pricing.

How much does a B200 cost per token compared with H100, H200, and B300?

On Akash, the B200 is priced at $5.00 per GPU-hour, equivalent to about $0.625 per TB/s of memory bandwidth per hour. That is the lowest bandwidth-adjusted cost among the four GPUs in this comparison.

For workloads where LLM decode is primarily limited by GPU memory bandwidth, this provides a useful estimate of relative token economics. Using the B200 as the baseline, the H100 costs about 22% more per unit of memory bandwidth, the B300 about 20% more, and the H200 about 48% more.

*Estimated relative cost per token is calculated from hourly rental cost divided by peak memory bandwidth, with the B200 set to 1.00x. It is a bandwidth-based proxy, not a measured tokens-per-second benchmark. The comparison assumes decode is primarily memory-bandwidth-bound. Specs from NVIDIA DGX B200 and NVIDIA H200.

The B200 and B300 both offer up to 8 TB/s of memory bandwidth, but that does not mean they will deliver identical token throughput in every workload. Actual inference performance depends on factors such as model architecture, quantization, batch size, KV-cache usage, software, and GPU utilization. NVIDIA also gives B300 architectural improvements beyond memory bandwidth, so bandwidth alone should not be used to predict token throughput.

The B300’s main advantage is memory capacity: 288 GB versus 180 GB on the B200. That extra capacity can allow larger models, longer context windows, or higher batching to fit on a single GPU. On a simple bandwidth-adjusted basis, however, the B200 has the lower hourly cost per unit of memory bandwidth.

For that reason, the table should be read as a memory-bandwidth cost comparison, not as a claim that the B200 will always produce tokens at the lowest dollar-per-token rate.

How much does it cost to fine-tune a model on B200s?

A B200 fine-tuning run on Akash costs $5.00 per GPU-hour at the rate used here, multiplied by the number of GPUs and the hours the run takes. For example, an 18-hour run on one 8x B200 configuration costs $720.

The hours are examples to show the math, not measured run times. See Akash GPU pricing for current marketplace rates.

To estimate your own run, time a short test, extrapolate to the full wall-clock hours, and multiply by the applicable GPU-hour rate and your GPU count. You pay while the deployment is running; closing the deployment stops the services and returns unused escrow to your available balance.

What should you know before renting a B200 on Akash?

Before you rent, check whether your model fits on one 8x B200 node, how many GPUs you’ll actually deploy, how many hours they’ll run, whether you need GPU interconnect, and whether you need on-demand or reserved capacity.

• Does your model fit on one node? An 8x B200 configuration provides 1,440GB of HBM3e. Compare your model’s weight memory with that capacity and leave additional headroom for KV cache and runtime overhead. KV-cache requirements vary with context length, concurrency, model architecture, and precision. Kimi K3’s roughly 1.56TB checkpoint does not fit on one 8x B200 node.
• How many GPUs will you actually deploy? Tensor parallelism does not universally require a power-of-two GPU count. The valid GPU count depends on the model and serving framework, so verify the supported parallel configuration before budgeting.
• Is the model FP4 or FP8? The B200 supports FP4 Tensor Core operations. Low-precision formats such as FP4 can substantially reduce model-weight memory compared with FP8, potentially reducing the number of GPUs required. Actual savings depend on the model and quantization format.
• Are you paying for hours or for tokens? A B300 can fit some models on fewer GPUs because it has more memory per GPU. A B200 configuration can provide more aggregate memory bandwidth for a larger GPU count, which may benefit bandwidth-bound serving. Benchmark your own model before committing.
• Do you need more than one node? For Kimi K3 or multi-node training, enable GPU interconnect when you deploy in Akash Console. Akash supports RDMA over InfiniBand and RoCE. You can require a specific fabric or let the provider choose.
• How many hours will it actually run? At $5/hour, one B200 running 24 hours a day for 30 days costs $3,600. Running it for 16 hours per day costs about $2,400 over 30 days. You need to actually close the deployment or use a runtime limit to stop billing.
• On-demand or reserved? For bursty or experimental work, use on-demand capacity. For predictable production workloads, Akash Enterprise offers reserved B300, B200, H200, H100, and A100 capacity with dedicated enterprise support and SLA-backed options.

Is a B200 on Akash cheaper than on AWS or Google Cloud?

At the rates used here, yes. A $5.00 B200 on Akash is about 64.9% cheaper than AWS’s $113.93/hour 8x B200 instance and 37.9% cheaper than Google Cloud’s $64.44/hour 8x B200 instance.

For an 8-GPU configuration running 720 hours, that is $28,800 on Akash, compared with $46,396.80 on Google Cloud and $82,031.62 on AWS.

Sources: Akash GPU pricing, Google Cloud accelerator-optimized pricing, and AWS EC2 on-demand pricing. Google lists the A4 a4-highgpu-8g at $64.44/hour, while AWS lists p6-b200.48xlarge at about $113.93/hour in us-east-1.

These are 8-B200 GPU configurations, not identical complete server specifications. Hyperscaler instance prices also include host CPU, RAM, storage, and other infrastructure.

Serving DeepSeek-V4-Pro or Kimi K2.6 on an 8x B200 configuration for 720 hours costs $28,800 at the $5/GPU-hour Akash rate. At the AWS list price above, the difference is $53,231.62.

How can enterprises reserve B200 capacity on Akash?

Enterprise teams can reserve B200s for a fixed term through Akash Enterprise, in addition to renting on demand through the open marketplace. Reserved capacity is available for B300, B200, H200, H100, and A100 GPUs, with bare metal available and an SLA.

On top of GPU-hours, Akash Enterprise provides:

• Dedicated bare-metal servers with no virtualization layer and no shared tenancy.
• High-bandwidth GPU-to-GPU interconnect for multi-node distributed training, including NCCL workloads.
• Hardware-based Trusted Execution Environments (TEEs) that protect data, model inputs and outputs, and model weights.
• Upfront pricing with no egress fees, billed by invoice in USD.
• SOC 2 compliant providers, with HIPAA and other certifications handled case by case.
• Direct access to the technical team for onboarding, architecture guidance, and deployment support, with defined response times.

The usual path is to test a workload at on-demand pricing first, then move steady workloads to reserved capacity. No contract is needed to get started.

FAQ