GPU Rental 2026: Rent a Cloud GPU by the Hour. Rent NVIDIA H100, H200, B200, B300, A100 & RTX 5090 GPUs from $0.59/hr

Cloud GPU rental runs from under a dollar an hour for a consumer card to roughly ten dollars an hour for a current-generation datacenter GPU. On Spheron, an NVIDIA H100 rents on-demand at $2.65/hr and an A100 80GB at $1.48/hr, billed per minute with no minimum term. The table on this page lists every GPU in the catalog with its live hourly rate, so you can price a workload before you deploy it.

Yes. Renting a GPU by the hour is the standard way to get datacenter GPU access without buying hardware. Spheron goes finer than hourly and bills per minute: billing starts when the instance reports ready and stops the moment you terminate it, so a 20-minute benchmark is billed as 20 minutes rather than rounded up to a full hour. There is no minimum run time and no charge for boot time.

Pick the GPU model, pick a region, click deploy, and SSH in. On Spheron that takes under two minutes with no approval queue and no support ticket. You get a bare-metal VM with root SSH, a dedicated public IP, Ubuntu 22.04 with CUDA preinstalled, and Docker with the NVIDIA Container Toolkit already set up, so a container that runs on your workstation runs on the rented GPU without changes.

A GPU cloud server is a remote machine with one or more datacenter GPUs attached that you rent by the hour instead of buying. You get the same silicon an on-premise cluster would give you, without the capital outlay, the 6 to 12 month lead times on current-generation cards, or the power and cooling bill. On Spheron a GPU cloud server is bare metal rather than a shared virtualised slice: root SSH, a dedicated public IP, Ubuntu 22.04 with CUDA and the NVIDIA Container Toolkit preinstalled, live in under two minutes, from $2.65/hr an hour for an H100. The catalog below lists every model with its live rate.

Spheron rents GPU servers with root SSH and a dedicated IP on every instance, single-GPU through 8x nodes. Access is bare-metal rather than a shared virtualised slice, so you get the full hardware with no hypervisor overhead and no noisy neighbours, which matters for low-latency inference, tight-loop training, and any profiling work where a shared host would corrupt the measurement.

Start from VRAM. For LLM inference you need roughly 2x the model's parameter count in GB for FP16 weights, so a 70B model needs about 140 GB and fits on a single H200 (141 GB HBM3e) or a pair of H100s. Training roughly doubles that pressure from gradients and optimizer state. Then pick on compute: H100 and B200 lead on training throughput, H200 leads on long-context inference, A100 80GB is the cost-effective middle ground for fine-tuning, and L40S, RTX PRO 6000, RTX 5090 and RTX 4090 cover inference serving, rendering, and development.

The consumer-class cards are the cheapest entry point. An RTX 4090 carries 24 GB GDDR6X and suits development, prototyping, and small-model inference up to roughly 13B parameters in FP16. For 70B inference you need considerably more VRAM, which means an A100 80GB at $1.48/hr or above. Every rate on this page is per-minute billed with no minimum commitment, and each GPU page lists both its dedicated and spot rates so you can compare the two directly.

Dedicated instances carry a 99.99% SLA and cannot be reclaimed by the provider: you pay a fixed hourly rate and the instance runs until you stop it. Spot instances run on spare capacity and are interruptible when that capacity is reclaimed. Use spot for checkpointable training runs, batch inference, and hyperparameter sweeps. Use dedicated for production serving and anything where an interruption is expensive. The two tiers track different capacity pools and are priced independently, so check both rates on the GPU page rather than assuming one is always lower.

Yes. Single-node configurations scale to 8x with NVLink on H100, H200, B200, B300, and A100. Beyond one node, Spheron provisions InfiniBand clusters with 400 Gb/s NDR fabric, RDMA, and NCCL-tuned topology, which is what multi-node training needs to keep all-reduce from dominating step time. Reserved cluster pricing is quoted separately and is discounted against the on-demand rates listed here.

No. Spheron bills per minute with no minimum term and no long-term contract. Stop an instance at any time and billing stops immediately. Reserved capacity for large clusters over weeks or months is available at discounted rates, but none of the standard rental tiers require a commit.

Checked on 8 Sep 2026, for a single H100 on-demand, Runpod listed $2.89/hr, Lambda listed $3.29/hr, and CoreWeave listed an 8x HGX H100 node that works out at $6.16/hr per GPU, all on their own public pricing pages. The live Spheron H100 on-demand rate is $2.65/hr. For H200, Runpod listed $4.59/hr and CoreWeave $6.31/hr against Spheron at $4.96/hr; Lambda does not list a self-serve H200 instance. Third-party rates move without notice, so treat those figures as a dated snapshot and check each provider's page before you commit. The comparison table on this page carries the same caveat.