NVIDIA B300 price per GPU-hour in November 2026: OVHcloud vs AWS vs Scaleway, with commitments and egress

Key points:

• Scaleway is the cheapest published B300 today. Its B300-SXM-8-288G node is €60 an hour (€7.50 per GPU-hour), Paris fr-par-2 only, read from Scaleway's public product catalog on November 15, 2026. Its availability API listed it as "shortage" the same day.
• AWS is the most expensive on demand and the most flexible with commitment. p6-b300.48xlarge is $142.416 an hour ($17.80 per GPU-hour) in us-east-1 and $217.54 in London. A 3-year all-upfront EC2 Instance Savings Plan brings us-east-1 down to $53.55 an hour, 62% off.
• OVHcloud has not published a B300 price on its price page. OVHcloud says B300 "arrive[s] on Public Cloud in November" (OVHcloud). Its public order catalog API already contains a b300-1990 flavor (8x B300, 1,990 GB RAM) at €85.60 an hour, but it is not tagged active and not on the price page. Treat it as provisional.
• Egress can flip the ranking for inference that serves the internet. OVHcloud includes outbound traffic outside Asia-Pacific. AWS charges $0.09/GB for the first 10 TB a month in us-east-1 after 100 GB free.
• GPU commitment discounts are uneven. AWS has Savings Plans and Capacity Blocks for B300. Scaleway's and OVHcloud's self-service Savings Plans both exclude GPU instances.

OVHcloud's B300 announcement brought a very practical question to my inbox. "If we wait for OVHcloud's B300, how much do we save compared with what we can rent today on AWS or Scaleway?" Nobody had put the three price lists side by side, so I did, on November 15, 2026, straight from each provider's own price files.

If you want the broader picture of where OVHcloud's B300 fits among European GPU clouds, Romaric wrote a ranked shortlist of seven GPU clouds. If you are sizing one specific model, Guillaume's Mistral Large 4 cost breakdown goes instance by instance. This article compares B300 prices, and only that.

How did I compare B300 prices across OVHcloud, AWS and Scaleway?

I compared the same hardware shape on each cloud, an 8x NVIDIA B300 node, using only the provider's own public price source, read on November 15, 2026. Every provider sells B300 as an 8-GPU node, so per-GPU-hour is the node price divided by eight.

• AWS: on-demand prices from the EC2 pricing data behind aws.amazon.com/ec2/pricing/on-demand, Savings Plans from AWS's Savings Plans price list (published November 13, 2026), and the Capacity Blocks price page. Linux, USD.
• Scaleway: the public product catalog API, which lists retail prices in EUR excluding VAT.
• OVHcloud: the Public Cloud price page for what is listed, and the public order catalog API for what is not yet on the page. EUR, excluding VAT.

Where I show dollars next to euros, I use the European Central Bank reference rate of 1.1052 USD per EUR from November 13, 2026. Your invoice will use a different rate, and OVHcloud's own USD price list does not follow the ECB rate either: its h200-1920 is €42.00 an hour in EUR and $49.56 an hour in USD.

A month is 730 hours throughout.

What does an NVIDIA B300 cost per GPU-hour on OVHcloud, AWS and Scaleway?

On published on-demand prices, a B300 GPU-hour costs €7.50 on Scaleway (about $8.29), $17.80 on AWS in us-east-1 and $27.19 on AWS in London. OVHcloud has no published B300 price yet. Its order catalog shows €10.70 per GPU-hour, which I flag as provisional.

Sources: Scaleway product catalog and availability API, OVHcloud order catalog, AWS on-demand pricing. AWS also lists the node in Jakarta ($230.93), São Paulo ($281.13) and GovCloud ($170.90).

On list prices, one AWS B300 node in us-east-1 costs as much as 2.1 Scaleway B300 nodes, and one AWS node in London as much as 3.3. Even OVHcloud's provisional catalog price sits about a third below AWS us-east-1.

What exactly does OVHcloud's catalog say about B300?

OVHcloud's public order catalog contains two plan codes, b300-1990.consumption and b300-1990_reserved.consumption, both at €85.60 an hour excluding VAT in the French, German and Irish catalogs, and £74.50 in the UK catalog. The technical description is 8x "B300 HGX" with 288 GB of HBM3e each, 240 vCores, 1,990 GB of RAM, a 400 GB NVMe system disk plus 8 x 3 TB of NVMe passthrough, and 25 Gbit/s of public and private bandwidth.

Three things make me treat it as provisional rather than a price:

• The price page does not list it. The H200 flavor next to it does.
• The H200 entries in the same catalog carry an active tag and a commercial block. The B300 entries carry neither.
• No region is attached on the price page, and OVHcloud's newsroom has not announced a price.

So the plain answer to "what does OVHcloud charge for B300?" is that OVHcloud has not published it yet. The €85.60 figure tells you where OVHcloud's catalog is pointing, and I would not sign a budget on it.

Is B300 the same machine on every cloud?

Nearly. All three are 8-GPU Blackwell Ultra nodes, and NVIDIA specifies 288 GB of HBM3e per B300 (NVIDIA). The host around the GPUs differs a lot.

Sources: AWS EC2 P6, Scaleway product catalog, OVHcloud order catalog.

Networking explains much of the price gap. AWS charges more partly because the P6 node comes with 6.4 Tbps of EFA networking for training and multi-node inference across many nodes. Scaleway and OVHcloud sell a strong single node with 20-25 Gbit/s links. If your model fits on one node, you are paying AWS for a fabric you will not use. If you train across 16 nodes, the European nodes are the wrong product.

Per gigabyte of GPU memory, using each provider's listed memory and the ECB rate, the hourly cost is about $0.029 on Scaleway, $0.041 on OVHcloud's catalog price and $0.066 on AWS us-east-1.

How much do commitment discounts change the B300 comparison?

On AWS, a lot. On Scaleway and OVHcloud, self-service commitments do not apply to GPUs today, so the list price is what you pay unless you negotiate a contract with sales.

AWS's Savings Plans price list for us-east-1 gives these hourly rates for p6-b300.48xlarge:

Sources: AWS Savings Plans price list and Capacity Blocks pricing.

A few details I found while reading the files:

• No 1-year EC2 Instance Savings Plan rate appears for p6-b300.48xlarge in the us-east-1 file. Only 3-year rates do.
• No Savings Plan rate at all appears for p6-b300.48xlarge in the London (eu-west-2) file. If you want B300 on AWS in Europe with a discount, Capacity Blocks are the published route.
• Capacity Blocks cost $129.168 per instance-hour ($16.146 per GPU) in Oregon, N. Virginia, the Atlanta Local Zone, Seoul, Hyderabad, Jakarta and London. AWS says prices are "updated regularly based on trends in supply and demand" and next in January 2027. The London Capacity Block is 41% cheaper than London on demand.

The 3-year EC2 Instance Savings Plan is the only AWS option that gets under Scaleway's list price per GPU-hour ($6.69 against about $8.29). It locks you to one instance family in one region for three years, for a GPU generation NVIDIA will replace well before then.

On Scaleway, the Compute Savings Plan explicitly excludes "B300, H100, L4, L40S, RENDER" GPU instances, and commitments range from €50 to €9,999 a month over 12 or 36 months (Scaleway docs). One B300 node at €43,800 a month would exceed that cap anyway.

On OVHcloud, Savings Plans cover "B3, C3 and R3 instances", with 15% off for 12 months and 30% for 36 months (OVHcloud Savings Plan). GPU instances are not on the list. The _reserved B300 plan code in the catalog hints at a reserved offer, but it shows the same €85.60 hourly price and no published terms.

Where can you get B300 capacity today?

AWS has the widest regional footprint, Scaleway has one Paris zone with stock marked as short, and OVHcloud has no published region yet. A low list price only helps if the node is available when you click create.

A few practical notes from deploying GPU workloads for customers:

• AWS quotas come first. New accounts rarely have the vCPU quota for a 192-vCPU P instance on demand. Request it before you need it, or buy a Capacity Block.
• "Shortage" on Scaleway is a real signal. Retry across a few hours or talk to Scaleway sales before you promise a launch date on a single Paris zone.
• OVHcloud's existing GPU regions are worth checking first. Its h200-1920 (8x H200) lists EU-WEST-PAR on the price page, and the European GPU cloud shortlist covers its Gravelines and Beauharnois GPU regions.

How much does egress add to the cost of a B300 node?

For inference that serves end users over the internet, egress is a few hundred to a few thousand dollars a month on AWS and zero on OVHcloud in Europe. Next to a $100,000 node, that sounds small, but it scales with your traffic and never goes away.

Sources: AWS Data Transfer price list, OVHcloud price page, Scaleway product catalog.

Two egress costs are easy to miss:

• Moving model weights. A 1.5 TB checkpoint pulled from S3 in another region, or out to another cloud, is billed as data transfer. Keep weights in the same region as the node.
• Hybrid setups. Training on AWS and serving on Scaleway or OVHcloud means paying AWS egress on every checkpoint you ship out.

What does a B300 node cost per month, all in?

For one 8x B300 node running 24/7 and serving 10 TB a month to the internet, the published list prices give roughly €43,800 on Scaleway, about $104,900 on AWS us-east-1 on demand, and about $40,000 on AWS with a 3-year all-upfront EC2 Instance Savings Plan. OVHcloud would be about €62,500 if its catalog price holds, with no egress charge.

These are list prices for a single node. Your real number depends on three things a price list does not show: whether you get capacity, how many hours a month the node really works, and what your account team will negotiate. At this price level, negotiate. Every provider in this table has a sales team that discounts GPUs for committed volume.

Is 24/7 the right assumption?

Usually not for development and staging. A B300 node that runs 10 hours a day on weekdays costs about 30% of the 24/7 figure. On AWS on demand in us-east-1 that is roughly $31,000 instead of $104,000 a month. Turning GPU environments off at night is the cheapest commitment discount there is, and it needs no contract.

Should you wait for OVHcloud's B300 or rent today?

Rent today if you need capacity this month, and keep your deployment portable so you can move when OVHcloud publishes its price and region. Waiting only makes sense if you already run on OVHcloud, rely on its included egress, or need its compliance posture.

My take by situation:

• One-node inference in Europe, now. Scaleway B300-SXM-8-288G, if Paris has stock. Lowest published price per GPU-hour, no egress line for instance traffic.
• Multi-node training or wide expert-parallel inference. AWS. The EFA fabric is the product, and Savings Plans or Capacity Blocks take the edge off the price.
• Already on OVHcloud, or egress-heavy public inference. Wait for OVHcloud's B300 to appear on the price page, then compare with Scaleway. Its catalog price sits between Scaleway and AWS, and the included egress helps when you serve a lot of traffic.
• Budget certainty above all. AWS 3-year Instance Savings Plan, if you are sure you will use the node for three years. I would not be, with Rubin already in vLLM's benchmarks.

If you are sizing a specific model, the Mistral Large 4 breakdown shows how the per-node math changes once you fit real weights and KV cache, and the Kimi K3.1 Pro self-hosting guide does the same for a 2.8T-parameter model that needs a full B300 node.

How do you keep the option to switch GPU clouds when prices move?

Run the model server as a Kubernetes workload, keep the weights in the same region as the GPUs, and make the cluster the only thing that changes when you move. B300 prices this month will not be B300 prices in six months, and the provider that is cheapest today may not have stock tomorrow.

Qovery is an Agentic DevOps platform that runs your workloads on Kubernetes in your own cloud account. It does not resell GPUs and does not host your models. The GPU bill, any Savings Plan or negotiated discount, and the data stay in your name.

• Your account on AWS or Scaleway. Qovery creates and operates the Kubernetes cluster, with GPU node pools per service on Qovery-managed clusters. On AWS, Karpenter provisions the GPU nodes.
• Your existing OVHcloud cluster. When OVHcloud's B300 lands in Managed Kubernetes, install Qovery on that cluster (Kubernetes 1.31+) and deploy the same services there.
• Same deployment, different cluster. Moving vLLM from an AWS P6 node to a Scaleway B300 node is a cluster change in Qovery, with the same service configuration and git-push deployments.
• Environment auto-stop and cost visibility. Stop non-production GPU environments on a schedule, and see what each environment costs.
• No dedicated DevOps hire. RBAC, audit logs and cluster upgrades are handled for your team.

How do I deploy this with Qovery from Claude Code or Cursor?

Install the Qovery Agent Skills, then ask your coding agent to create a cluster with a B300 GPU node pool in the cloud you picked and deploy your inference service on it. The skills deploy, and the MCP Server operates what already exists.

curl -fsSL https://skill.qovery.com/install.sh | bash

# Project-scoped instead of global
curl -fsSL https://skill.qovery.com/install.sh | bash -s -- --project

# MCP Server for Claude Code: OAuth, read-only by default
claude mcp add --transport http qovery https://mcp.qovery.com/mcp --callback-port 4242

# MCP Server with write access, only when explicitly wanted (also needs write access enabled in the Qovery Console)
claude mcp add --transport http qovery "https://mcp.qovery.com/mcp?read_write=true" --callback-port 4242

• The first command installs the skills (qovery-onboard, qovery-deploy, qovery-troubleshoot, qovery-optimize, qovery-speedup, qovery-preview, qovery-terraform) for every compatible agent. The agent asks before each action.
• The second limits the install to the current project.
• The third connects Claude Code to the Qovery MCP Server through OAuth in your browser, read-only.
• The fourth adds write access, only if you want it and after enabling it in the Qovery Console.

Then ask, for example: "create a Kubernetes cluster in my Scaleway account in fr-par-2 with Qovery, add a B300-SXM-8-288G node pool, deploy my vLLM service on it, and stop it at night." Swap Scaleway for AWS, or point the agent at your OVHcloud cluster. The AI agent quickstart and the Claude Code integration have the details. For the cost side, qovery-optimize reviews resource usage on what you already run.