/blog / review / runpod

Runpod review: bare-metal H100s without the enterprise tax

Six weeks on Runpod across Community Cloud, Secure Cloud, and Serverless. The benchmarks, the bills, and where it falls short of AWS and Lambda.

Tobias 14 min read 8.7
  • gpu
  • review
  • runpod
  • h100
  • serverless

We rented Runpod for six weeks across all three of their tiers — Community Cloud, Secure Cloud, and Serverless — and ran the same benchmark suite we use for every GPU provider. Below is what we found, what to watch out for, and where the marketing pages tell a slightly nicer story than the reality.

If you want to skip ahead and try them, our referral is runpod.io/?ref=8vbo5oc9 — you get a random credit bonus between $5 and $500 when you spend your first $10. Yes, $500 is real, just unlikely.

Why Runpod, why now

The hyperscalers price H100s like enterprise software. You pay for the GPU, the node, the networking, the egress, and the privilege. Runpod’s pitch is that for 80% of workloads — fine-tuning, batch inference, prototyping — you don’t need any of that. You just need the GPU.

Community Cloud is the cheapest tier, where individual hosts list spare capacity. Secure Cloud is Runpod’s own datacenter footprint with stronger networking and SLAs. Serverless is exactly what it sounds like — pay-per-second, scale-to-zero containers.

The numbers

We ran a fixed harness against four providers. Higher is better unless noted.

Metric Runpod Lambda Vast.ai Paperspace
  • $/hr H100 80GB (PCIe) $
    1.99
    2.49
    1.79
    3.09
  • Llama-3 70B tok/s (4-bit)
    142
    138
    121
    109
  • SDXL imgs/min
    38
    35
    31
    28
  • fio rand-read MB/s
    3850
    4100
    2200
    2900
  • Cold start (s) s
    2.5
    6.1
    11.4
    8.7
  • Egress $/GB $
    0
    0
    0.01
    0.02

A few things stand out. Vast.ai undercuts Runpod on raw $/hr — but the Vast box was a Community-tier listing and the throughput numbers reflect that. Runpod’s Community Cloud was more consistent across our 7-day window than Vast’s, which is the reverse of what you’d expect.

Cold start is the headline metric. Runpod Serverless reliably hit our 2.5-second target across 1,200 invocations. Lambda’s slowest tier was nearly 3× that. If your workload is spiky, this is the difference between affording GPU inference and not.

Pricing in real terms

These are the plans we benchmarked. The setup column is “time from clicking deploy to running your first command.” Egress is per-GB outbound.

Plan Price / hr Setup Egress
  • Community Cloud H100 80GB $1.99 ~90s $0
  • Secure Cloud H100 80GB $2.79 ~120s $0
  • Serverless H100 (active) $0.00074/s ~2.5s cold $0
  • RTX 4090 (Community) $0.34 ~60s $0

A 24-hour H100 fine-tune on Community Cloud is about $48. The same job on AWS p5.48xlarge (prorated) is north of $300 once you factor in storage and egress. Even comparing apples to apples — Lambda 1xH100 on-demand at ~$2.49/hr — Runpod is meaningfully cheaper at the going rate, with the caveat that “going rate” on Community Cloud fluctuates.

What’s actually good

The web console is fast. Spawning a pod genuinely takes ~90 seconds from first click. SSH keys propagate in seconds. The Jupyter integration just works. Volumes are stable across restarts. The CLI (runpodctl) is a thin but coherent wrapper — nothing fancy, but it doesn’t fight you.

Serverless is the killer feature. We routed an LLM proxy through Runpod Serverless for a week and the p99 cold start was 2.8 seconds. The p50 with a warm worker was 110ms.

What’s not

Documentation is patchy. Community Cloud capacity disclosures are vague — you’ll hit “no available hosts” warnings on busy days, especially for H200s. Networking between pods is fine, not great; you wouldn’t run a multi-node training job here without a hard look at NCCL throughput.

Their support channel is Discord-first. That’s fine if you live in Discord. Less fine if you’re trying to file a billing question on a Tuesday.

Verdict

Verdict

8.7

/ 10

Pros

  • + Best $/hr H100 across our tested providers
  • + Genuinely zero egress fees
  • + Serverless cold start is real (~2.5s)
  • + Fast pod spin-up, clean console
  • + Works for hobbyists and production both

Cons

  • − Community Cloud variance is real (mostly fine, occasionally not)
  • − Docs lag behind product
  • − Multi-node networking isn't a strength
  • − Discord-first support

Buy it. If you’re choosing your first GPU rental, choose Runpod. If you’re already on Lambda or Paperspace and your monthly bill is under $1k, the move probably isn’t worth it. Above that, run our benchmark suite on Runpod yourself — that 20-30% bill reduction materialises faster than you’d think.

Sign up via our referral if you want the random $5–$500 credit bonus on your first $10 spend. We’ll get a small kickback. Thanks if you do.

Keep reading