RunPod

RunPod sells four fairly different things under one account, and the pricing logic of each is different.

Pods are rented GPU containers you SSH into or reach through JupyterLab, VS Code or Cursor. You pick a template or bring any Docker image from Docker Hub, GHCR or ECR. This is the core product and where the price advantage lives.

Serverless is a request-driven autoscaler: you package a handler in a container, RunPod queues requests and spins workers up and down. Flash, generally available since 2026, is a newer front door to the same infrastructure — a Python SDK where you decorate a local function with @Endpoint(gpu=GpuType.NVIDIA_GEFORCE_RTX_4090, dependencies=["torch"]) and it provisions and scales remotely. That is a Modal-shaped developer experience, and it is a real differentiator against pure GPU-rental competitors. The SDK is MIT-licensed on GitHub; the platform behind it is proprietary and hosted.

Instant Clusters are multi-node jobs — RunPod says more than 200 simultaneous GPUs over InfiniBand, which we have not independently verified.

Public Endpoints are pre-deployed models (Flux, WAN, Kling, SORA 2, Qwen3, Kimi) behind a per-unit API, with no infrastructure to manage.

If you came for cheap GPUs, you want Pods. The rest is worth evaluating separately and on its own numbers.

Verified against RunPod's own pricing page on 15 August 2026. On-demand, per GPU, per hour, Community Cloud / Secure Cloud:

For a like-for-like reference point, Lambda lists an H100 PCIe 80GB at $3.29/hr on-demand today. RunPod's Secure Cloud undercuts that by 12% and its Community Cloud by 40%.

Two things about the billing model matter as much as the headline rate. Pods bill per second with no ingress or egress fees — worth real money if you are shuttling model weights or datasets, where hyperscalers charge for every gigabyte out. And 3- or 6-month prepaid savings plans exist for steady workloads, though the discount is not published; you see it at checkout.

One inconsistency to be aware of: the Pods overview page says pods are "billed by the minute" while the Pods pricing page says "billed by the second." We have not established which is authoritative.

This choice moves your bill by 20-55% and it changes what you are guaranteed. It is the most consequential thing on the deploy screen and the easiest to click past.

Secure Cloud runs in T3/T4 data centres. RunPod says its Secure Cloud infrastructure partners meet SOC 2, ISO 27001 and PCI DSS. RunPod itself has completed SOC 2 Type 2, and its Trust Center lists SOC 2 Type 2, SOC 3, HIPAA, GDPR and a 2026 bridge letter — behind an access request, so we could not read them.

Community Cloud connects individual compute providers to users through what RunPod describes as a vetted, secure peer-to-peer system. Isolation is containerised and multi-tenant. Protection against the host itself is a contractual control: RunPod's terms of service prohibit hosts from inspecting your data or analysing usage patterns, with removal from the platform for violation.

That is a legitimate trade to make, and for training runs, experimentation and anything you can checkpoint it is usually the right one. It is a different trade if you are processing customer data under a DPA, because the certifications RunPod names are attached to Secure Cloud's partners, and the compliance page itself says coverage "can vary by workload, region, provider, and deployment model." If a compliance question is going to be asked of you later, get the answer in writing during evaluation rather than inferring it from the marketing page.

Serverless is billed per second at a higher hourly rate than a dedicated pod, which means it saves money only below a certain duty cycle. RunPod publishes both numbers, so the break-even is straightforward arithmetic. Flex-worker rates verified 15 August 2026:

Against Community Cloud pods the crossover comes much sooner — roughly 42% for H100 and 31% for the 4090.

Above those lines, a dedicated pod that you simply leave running is cheaper. Below them, Serverless wins, and by a lot at low traffic. Three details shift the real number against you: you are billed for start time (container init and model load) as well as execution, for the idle timeout after each request (default 5 seconds), and container disk is charged at roughly $0.10/GB/month on 5-minute intervals. Bursty, spiky traffic with big models pays the start-time tax repeatedly.

"Active workers" — always-on, no scale-to-zero — have discounts available only through a sales conversation, so you cannot model that tier from public information.

RunPod claims Serverless saves 25% over other serverless cloud providers on flex workers alone, and claims sub-200ms cold starts via FlashBoot. Both are vendor figures we did not verify.

There are three storage types and they behave differently when a pod is not running.

Volume disk costs twice as much while your pod is stopped as while it is running. A 500GB volume disk on a pod you parked for a month is $100; the same 500GB on a network volume is $35 whether the pod runs or not. RunPod documents this openly — it is not hidden — but it inverts the intuition that stopping something makes it cheaper, and it is the most common way to be surprised by a RunPod invoice.

Two related things worth setting up on day one. If your account balance reaches $0, pods are stopped automatically — but pods without a network volume are terminated and that data cannot be recovered. And there is a default account-wide spend limit of $80/hour, which is a useful guardrail and also a ceiling you must contact support to raise before a large cluster run.

The practical upshot across both this and the section below: use network volumes, not volume disks, for anything you care about.

A pod is bound to a specific physical machine. When you stop it, you release that GPU back to the pool and someone else can rent it — but the pod stays tied to that machine. If the GPU is taken when you come back, RunPod offers to restart the pod with zero GPUs, which is a data-recovery mode: you can reach the volume disk, but you cannot compute. Your options are to wait with no guaranteed timeline, or terminate and redeploy elsewhere.

This is inherent to a marketplace model rather than a defect, and RunPod documents it plainly and recommends the fix: network volumes decouple your data from the machine, so you can redeploy anywhere and reattach. But it means "stop the pod overnight to save money" is not the safe habit it is on a hyperscaler, and it is a genuine argument for architecting around network volumes and checkpointing from the start.

RunPod's own maintenance guidance is consistent about this: it recommends checkpointing every one to a few hours for long jobs, tells you container-disk-only data is unrecoverable after an interruption, and states it does not charge for planned maintenance windows.

RunPod's hosted-model API charges per unit of output: $0.02 per megapixel for Flux Dev, $0.0024/MP for Flux Schnell, $0.50 per 5 seconds of WAN 2.5 video, $0.05 per 1,000 characters for Minimax Speech, and $10.00 per 1M tokens for Qwen3 32B.

The image and video rates are reasonable for what they are — a 1024×1024 Flux Schnell image works out to about $0.0025, and access to Kling, Seedance and SORA 2 behind one API key with no deployment is genuinely convenient.

Text is a different story. $10.00 per 1M tokens for a 32B dense model is not competitive with dedicated inference providers: Together lists Qwen3 235B A22B Instruct — a substantially larger model — at $0.20 in / $0.60 out per 1M tokens today. If you are buying tokens at any volume, buy them somewhere else and use RunPod for the GPU rental it is actually good at.

RunPod raised $100M led by Summit Partners in June 2026 at a $1B valuation, its first major round since a $20M seed led by Intel Capital and Dell Technologies Capital in May 2024. Coverage of the round reports the company turned down $500M buyout offers. Company-stated figures in and around that announcement — more than one million developers, over 20 billion inference requests processed, ARR growth from $120M to $240M — are RunPod's own numbers and should be read as claims.

What we could verify independently is more mundane and more useful. RunPod runs a public status page at uptime.runpod.io with 90-day per-datacenter uptime, which lists 32 named regions — corroborating the "31 global regions" claim on the homepage. On 15 August 2026 most regions showed 100% over 90 days, with the weakest at 99.345% (EU-SE-1) and 99.515% (US-TX-3), and an open incident: a power failure at US-IL-1. Publishing per-region history at that granularity, including live bad news, is not something an operator does casually.

On GitHub, the tooling is real and actively maintained: runpod-python (306 stars, MIT), runpodctl (433 stars, GPL-3.0) and flash (182 stars, MIT) all had pushes within the last week, alongside a Terraform provider, Go and JS SDKs and an MCP server. Note that these are clients for a proprietary hosted platform — none of them makes RunPod an open-source product, and none can be self-hosted.

Customer sentiment is mixed and worth reading yourself. RunPod is rated 3.7 out of 5 on Trustpilot as of 15 August 2026 (total review count not established). Positive reviews cite ease of setup and price; recurring complaints cluster on GPU availability — waiting for capacity on serverless, broken or CUDA-incomplete templates on newer cards — and on refund handling. RunPod responds to reviews on both sides, including the critical ones.

Documented constraints that will not appear in a pricing comparison but will break a plan:

• No Docker Compose. RunPod runs Docker for you; you cannot run your own Docker daemon inside a pod. Multi-container local stacks need rearchitecting.
• No UDP. Pods support TCP and HTTP only.
• No Windows.
• Flash is macOS and Linux native; Windows requires WSL2. Flash CPU endpoints are restricted to the EU-RO-1 datacenter.
• Default $80/hour account spend limit, raised only by contacting support.

None of these is unusual for a container-based GPU cloud, but the Docker Compose and UDP restrictions in particular have sent teams back to the drawing board after they had already migrated.