RunPod

What RunPod actually is

RunPod sells GPU hours and very little else. In another category that would be a limitation; here it is the design. The product surface is GPU pods (Docker containers pinned to an accelerator), serverless endpoints that scale to zero, multi-node clusters for distributed training, and network volumes that outlive the compute attached to them. If the job is "put this container on an NVIDIA GPU and bill me by the second", RunPod is purpose-built for it. If you want managed Postgres, a CDN, or cheap shared web hosting, this platform has nothing for you.

The fleet splits into two explicit tiers. Community Cloud runs on vetted third-party hosts and prices lower. Secure Cloud runs in tier-3/4 datacenters, carries a SOC 2 Type II certification, and attaches a 99.9% uptime SLA. The pricing page advertises thousands of GPUs across 30+ regions. RunPod launched in March 2022, co-founded by Zhen Lu and Pardeep Singh, and lists a San Francisco address (329 Bryant St) in the footer of its pricing page. We think the two-tier split is more honest than most GPU marketplaces manage: the discount is stated openly, and the reliability trade-off is yours to accept rather than hidden behind a single blended price.

We file RunPod under GPU cloud, not general VPS hosting, because almost none of it transfers to running websites. There is no budget shared tier and no control panel for WordPress. There is a very large GPU list, per-second billing, and a template library that gets a container running in minutes.

Pricing: what you really pay

Billing is per second, on demand, with no commitment, and the rates are published — three things we weigh heavily. On the Secure Cloud rates RunPod itself verified in August 2026, an RTX A5000 24GB starts at $0.27/hr, the RTX 4090 runs $0.89/hr, A40 and RTX A6000 sit at $0.59/hr, L40S at $1.09/hr, A100 80GB at $1.79/hr, H100 PCIe at $2.89/hr, H100 SXM at $3.99/hr, H200 at $5.29/hr, and a 288GB B300 tops the list at $8.99/hr. Community Cloud undercuts most of that: the current pricing page shows the B300 at $7.89/hr and the RTX 4090 at $0.74/hr on community hardware. Independent trackers list community cards as cheap as $0.12/hr for an RTX A2000, with the obvious caveat that stock on the cheapest parts comes and goes.

Serverless costs more per busy hour — $2.72/hr for A100 workers, $4.79/hr for H100, $1.75/hr for the 48GB tier — but idle time bills at zero, so for spiky or low-volume inference it usually lands cheaper than a pod left running. Multi-node clusters price on their own: H200 SXM at $4.31/hr per GPU, A100 SXM at $1.79/hr per GPU, per RunPod's own cost guide.

Storage is where the arithmetic needs care. Container disk is about $0.10/GB/month; volume disk is $0.10/GB/month while the pod runs and $0.20/GB/month while it is stopped — double rate for data at rest, which is the most common bill complaint we see. Network volumes are the sensible default: $0.07/GB/month for the first terabyte, $0.05/GB/month above that, $0.14/GB/month for the high-performance tier. Egress is free, which quietly removes a whole category of hyperscaler surprise. A 500GB network volume holding model weights costs roughly $35/month whether or not anything trains on it, so include that line in any comparison. There is no free tier — sign-up credits arrive through referral schemes only.

Two storage rules from the docs are worth reading before you trust the platform with irreplaceable data. If your balance reaches zero, storage charges keep accruing on stopped pods, and an uncovered network volume can eventually be terminated with its data unrecoverable. And volumes can grow but never shrink.

Performance & infrastructure

We have not benchmarked RunPod ourselves — our dated sysbench, fio and iperf3 runs cover Hetzner, DigitalOcean, UpCloud, Vultr and Linode instances, not GPU pods — so everything in this section is editorial judgment rather than our own test data. Read our benchmarks for the CPU-side numbers we do stand behind.

What the hardware lists show is breadth and currency: A40 and A6000 class cards at the bottom, L40S and RTX 6000 Ada in the middle, A100, H100, H200 and Blackwell B200/B300 parts at the top, with real system memory and vCPUs bundled per pod (6 vCPUs and 41GB RAM on the RTX 4090 tier, 20 vCPUs and 125GB on H100 SXM). Nobody renting inference capacity should assume consumer cards behave like datacenter parts — that is the Community Cloud trade — but for throughput per euro the consumer tier is hard to argue with.

Network volumes are NVMe-backed, and the high-performance tier is documented at up to 3x throughput and 4x IOPS over standard — meaningful for training runs and checkpoint-heavy fine-tuning. RunPod also exposes an S3-compatible API over network volumes with a per-datacenter endpoint (the docs reference EUR-NO-1, US-GA-2 and CA-2), so datasets can be moved without booting a pod. Those codes confirm EU, US and Canadian presence; one third-party profile mentions Oceania too, which we could not verify on RunPod's own pages.

On reliability, the 99.9% SLA covers Secure Cloud and is paired with a SOC 2 Type II report. Community Cloud carries no such guarantee by construction — you are renting from a vetted third party, and stock trackers show some community listings unavailable for long stretches. Per-second billing caps the cost of a crashed job at the seconds it ran.

Developer tooling is better than the bare-bones console suggests. There is a REST API covering pods, volumes and endpoints, a CLI, and a large library of community templates that preinstall frameworks, so the gap between signing up and a working training environment is usually a single command. SSH and Jupyter access into pods is standard, which matters more than it sounds: much of the GPU market still forces you through clunky web terminals. For teams automating pipelines, the API-first surface means pods can be created and torn down from CI without human hands. The console UI itself draws regular criticism, and fairly — it is functional rather than pleasant — but almost everything it does is also scriptable, which is the part that ages well.

Where it falls short

Support is the complaint with the longest tail. RunPod is a lean company selling cheap GPU seconds, and ticket response shows it; teams that need a named engineer and a phone number should look at the enterprise tier or elsewhere. Community Cloud reliability is a feature of the pricing, not a bug you can support-ticket away: hosts can vanish and cheap GPUs sell out.

The storage billing model punishes inattention. Idle volume disk at double the running rate, network volumes that keep billing after compute stops, and eventual deletion when a zero balance can't cover charges — each is documented, and each still catches people. Volumes cannot shrink, and concurrent writes to one network volume from multiple workers can corrupt data; the docs place that burden on your application code. That is fine for engineers, less fine for everyone else.

Then there is the transparency gap on infrastructure details. We found no published IPv6 support statement and no clear DDoS protection story for pods, and while "30+ regions" is advertised, the exact location list is scattered through documentation codes rather than presented as a clean table. For a company whose buyers are exactly the people who read spec tables, that is an odd omission.

Who should pick this

Our pick for solo ML developers, small research teams, and anyone running bursty inference who wants per-second billing without a sales call. Fine-tuning jobs, batch embedding pipelines, ComfyUI and Stable Diffusion workloads, and prototypes that would cost ten times more on hyperscaler GPUs all fit the model well. The serverless endpoints are a genuinely good fit for a model that gets traffic in unpredictable bursts.

Look elsewhere if you need contractual reliability for production inference with an escalation path, if you want your data in a specific jurisdiction with paperwork to prove it, or if your workload is not GPU-shaped at all. EU buyers who like the Norway and EU region footprint should still check the Secure Cloud location list before assuming residency.

Key takeaways

• Per-second on-demand billing with published rates: RTX A5000 at $0.27/hr up to B300 at $8.99/hr on Secure Cloud; Community Cloud runs cheaper (RTX 4090 at $0.74/hr).
• Two explicit tiers: cheaper Community Cloud on vetted hosts versus Secure Cloud in tier-3/4 datacenters with SOC 2 Type II and a 99.9% SLA.
• Network volumes at $0.07/GB/month (first TB) with free egress are the cost-effective storage path; idle volume disk bills at double rate.
• Serverless endpoints scale to zero, which suits spiky inference better than pods left running.
• No free tier, no general-purpose cloud services, and no published IPv6 or DDoS position for pods.
• Best fit is GPU-shaped work on a budget; teams needing enterprise support and firm guarantees should look upmarket.

Sources

• https://www.runpod.io/pricing-page
• https://runpod.io/articles/guides/ai-server-cost
• https://docs.runpod.io/storage/network-volumes
• https://docs.runpod.io/serverless/pricing
• https://docs.runpod.io/serverless/storage/s3-api
• https://checkthat.ai/brands/runpod
• https://gpufinder.dev/providers/runpod