Runpod

A Pod is a container with a GPU attached that runs until you stop it. Runpod publishes list prices for every tier, which by itself puts it ahead of marketplaces where the answer is "it depends on the host". On 24 August 2026 the published rates started at $0.27/hr for an RTX A5000 and $0.44/hr for an A40, ran through $0.99/hr for an L40S or RTX 5090 and $1.59/hr for an A100 SXM 80GB, and topped out at $3.29/hr for an H100 SXM, $6.79/hr for a B200 and $7.89/hr for a B300. Deployment spans more than thirty regions.

Two details do more for the total bill than the hourly rate. First, Runpod charges nothing for data ingress or egress on Pods. Against providers that bill bandwidth per byte, that quietly erases a line item that can rival the GPU cost on dataset-heavy training. Second, savings plans let you prepay three or six months for a discount that reattaches automatically the next time you deploy the same GPU type — with the caveat, spelled out in Runpod's docs, that plans are non-refundable, have fixed expiry dates, and cover compute only.

Runpod's supply comes in two flavours. Secure Cloud is T3 and T4 datacentre capacity with high redundancy, positioned for production and sensitive data. Community Cloud is peer-to-peer supply from third-party providers at more competitive prices and variable reliability.

The 2026 fact worth knowing is that Runpod's documentation now states it is no longer accepting new hosts for Community Cloud. Existing capacity remains available, but the cheap tier is a legacy pool rather than a growing one. If your cost model assumes Community Cloud pricing will keep undercutting Secure Cloud indefinitely, that assumption has a shelf life. It also nudges Runpod's identity away from "marketplace" and towards "GPU cloud with a bargain bin", which is a different competitive position against Vast.ai than it held two years ago.

Serverless is the reason a lot of teams use Runpod at all. You package a handler into a container, deploy it as an endpoint, and Runpod runs workers only when requests arrive. Billing is per second from worker start to worker stop, rounded up, so an endpoint that serves a hundred requests a day costs roughly a hundred requests' worth of GPU time instead of twenty-four hours of it.

List rates on 24 August 2026 ran from $0.58/hr of worker time for 16GB-class GPUs and $0.69/hr for the 24GB tier, through $1.22/hr for A6000 and A40 workers, $2.72/hr for A100, $4.79/hr for H100, and $8.64 and $9.98/hr for B200 and B300. Note that the docs also publish a per-second table for the same tiers and the two do not always agree to the cent, so check the console figure at deploy time.

Endpoints come in two shapes. Queue-based endpoints put a managed queue in front of workers, with guaranteed execution and automatic retries — right for async and batch work. Load-balancing endpoints route straight to workers, bypassing the queue, for low-latency APIs and real-time inference. Autoscaling runs on either queue delay (add workers when requests wait longer than a threshold, four seconds by default) or request count, which the docs recommend for LLM workloads and frequent short requests.

Because you pay from the moment a worker starts, the time spent pulling a container and loading weights into VRAM lands on your invoice. Runpod's answers are FlashBoot, which retains worker state after spin-down so a revival is faster than a fresh boot and is on by default, and model caching, which schedules workers onto machines that already hold the model files. Both help most when traffic is steady enough that workers cycle rather than starting from nothing.

If you need zero cold start, you set active workers to one or more. Those stay warm permanently and bill continuously, which is effectively a Pod wearing a Serverless interface — perfectly reasonable for consistent traffic, and worth pricing against a plain Pod before you commit.

Two lifecycle defaults catch people out. The idle timeout is five seconds, and you are billed through it, so a chatty endpoint with short requests spends a meaningful share of its budget waiting. And endpoints that go three days without a request drop to two max workers; after seven days, max workers is set to zero and stays there until you raise it manually.

Container disk is temporary and vanishes when the Pod stops. Volume disk persists across a stop but not a delete, and costs $0.10/GB/month running and $0.20/GB/month stopped — the inversion is deliberate, and it means a parked Pod with a large volume is more expensive per gigabyte than a working one. Network volumes are the durable option at $0.07/GB/month under a terabyte and $0.05 above, with a high-performance tier at $0.14.

Network volumes also solve a specific failure mode Runpod documents plainly: a stopped Pod stays bound to its original physical machine, and if someone else rents that GPU in the meantime, your Pod can only restart with zero GPUs for data recovery. Keeping working data on a network volume means you can simply terminate and redeploy elsewhere. The trade-off is added network latency and being pinned to that volume's data centre, which shrinks the GPU pool an endpoint can draw from.

Against Vast.ai, Runpod is the more predictable of the two. Vast will usually beat it on the raw GPU line — cheapest single-GPU H100 SXM listings were under half Runpod's $3.29/hr H100 rate on 24 August 2026 — but Vast bills bandwidth and per-host storage separately, where Runpod charges no transfer fees at all. For jobs that move a lot of data, the gap narrows sharply.

Against Modal, Runpod is cheaper per GPU-second and less opinionated; Modal's Python-native developer experience is the nicer one if you want to think in functions rather than containers. Against Replicate or Together, Runpod is the choice when you want to run your own container rather than a hosted model behind someone else's API. And against SageMaker, Vertex or Anyscale, Runpod is deliberately not competing: it sells compute and endpoints, not pipelines, experiment tracking or distributed training frameworks.

Runpod also now ships a Hub of deployable templates, multi-node Clusters scaling to 64 GPUs with reserved options, and pre-deployed public endpoints for common open models billed per request — useful if you want inference without maintaining a worker at all.