Where can you get NVIDIA B300 GPUs by the hour?
You can get NVIDIA B300 GPUs by the hour, with no contract, from several platforms that publish a self-serve rate. Nodus lists the B300 at $6.943/hr on its pricing page, next to Modal at about $7.10, RunPod at $8.99 and Nebius at $9.50 per GPU-hour. The cheapest rate is not always the cheapest run, because some offers only come as a full 8-GPU machine and some need a specific CUDA version. Below…
You can rent NVIDIA B300 GPUs by the hour without a contract from several platforms, including Nodus, Modal, RunPod, Nebius and Amazon EC2. The hourly rates range from $6.943 to $9.50 depending on the platform and specific offering.
Nodus lists the B300 at $6.943 per hour, Modal at about $7.10, RunPod at $8.99 and Nebius at $9.50 per GPU-hour. The cheapest rate isn't always the best deal though, as some offers require a full 8-GPU machine and others have specific CUDA version requirements.
The B300 is NVIDIA's Blackwell Ultra GPU model, offering up to 288 GB of HBM3e memory and 15 PFLOPS of dense NVFP4 compute per GPU. This makes it a memory-intensive GPU, ideal for training jobs where memory is a bottleneck. On Nodus, the B300 costs 1.26 times the B200's $5.50 per hour, so a BF16 job must be 1.26 times faster to be cost-effective.
When deciding whether to choose a B300 over a B200 or H200, consider if your training state exceeds 180 GB per GPU, if it saves you a second node, or if your training recipe runs in NVFP4. Otherwise, a B200 or H200 is likely cheaper.
Amazon EC2 also offers B300 GPUs, but they are listed as an 8-GPU instance type rather than an hourly rate. For smaller runs, the machine shape matters more than the hourly rate, as you pay for the entire smallest machine you rent. Run a short test on the smallest available B300 configuration to gauge the cost, then scale up for the full run if needed.
To request a B300 from Nodus, specify your required memory rather than the GPU model name. Use the `minMemory` parameter set to at least 200 GiB to ensure a B300 is selected. Checking availability and costs before starting the job is advised, as these platforms charge only for GPU time and additional fees may apply.
Confirming that you've received a B300 rather than a B200 or other GPU can be done by querying your driver for the GPU name and total memory after booting the machine. Remember that NVIDIA's HBM capacity can vary by SKU, so verify the actual memory the driver reports rather than relying on the headline figure. Some platforms like Modal are better suited for Python-based workflows, while others may be more appropriate depending on your team's setup and hardware needs.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
