For an AI startup that trains or fine-tunes every week, Nodus is the best value: the lowest listed NVIDIA H100 rate of the platforms here ($2.60/hr), plus automatic checkpoint restore when a machine is lost. Spend free AWS or Google Cloud credits before you pay anyone. Pick Modal or RunPod if you mostly serve low-latency endpoints, and a Lambda cluster only once 16 or more GPUs will stay busy for weeks.
We build Nodus. Every other company's price or program below links to its own page, the arithmetic is shown once, and one section says where each alternative is the better pick.
Which GPU platform fits each startup stage?
Pick by the work you run this quarter, not by the company you hope to be in two years.
| Your situation | Best fit | Why |
|---|---|---|
| You hold large AWS or Google Cloud credits | Spend them first, on that cloud | Free money beats any hourly rate |
| You fine-tune or train most weeks | Nodus | Lowest listed H100 rate here, checkpoint recovery, hard budgets |
| You mostly serve a model behind an API | Modal or RunPod Serverless | Built around autoscaling endpoints |
| You want a cheap pod to poke at by hand | RunPod Community Cloud | $2.69 H100 SXM, per-second billing |
| You will keep 16+ GPUs busy for weeks | Lambda clusters | Reserved clusters from 2 weeks to 1 year |
Sources: Nodus pricing, RunPod pricing, Modal pricing, Lambda pricing, AWS Activate, Google for Startups Cloud Program. Prices and programs checked October 30, 2026.
These companies sell different things. RunPod, Lambda and the big clouds hand you a machine to run and shut down, and Modal runs functions you write against its SDK. Nodus runs a command or training job to completion and handles placement, checkpoints and replacement machines.
What does one month of startup training cost on each?
We priced a realistic month for a small team that fine-tunes its own model: four runs of 8 x NVIDIA H100 for 10 hours each, or 320 GPU-hours. Each rate is the listed H100 price on the company's own pricing page.
| Platform | Listed H100 rate | 320 GPU-hours |
|---|---|---|
| Nodus | $2.60/hr | $832.00 |
| RunPod Community Cloud | $2.69/hr | $860.80 |
| Modal | $0.001097/s ($3.9492/hr) | $1,263.74 |
| RunPod Secure Cloud | $3.99/hr | $1,276.80 |
| Lambda, 8 x H100 SXM | $3.99/hr | $1,276.80 |
The math is rate x 320. On listed rates, Nodus is $431.74 a month below Modal and $444.80 below RunPod Secure Cloud or Lambda. Over a year of the same usage, that is $9,984.00 on Nodus against $15,321.60 at $3.99, a difference of $5,337.60.
At list prices, a month of four 8-GPU fine-tunes costs about a third less on Nodus than at the $3.99 rates.
RunPod Community Cloud comes within $28.80 of Nodus. The gap that matters there is not the rate but what happens when a run dies at hour 8, which the next section covers.
List price is also not the whole bill. Idle machines, failed runs and storage add to it, and we break those down in what an H100 or B200 hour really costs per finished job. Run nodus run --dry-run on your real job to see an estimate before you spend anything.
How far do startup credits actually go?
Credits are the biggest lever an early startup has, and their size varies by orders of magnitude. AWS Activate offers up to $200,000 to pre-Series B startups through an Activate Provider such as an accelerator or VC. Google offers up to $350,000 over two years to startups from pre-seed to Series A.
GPU-focused providers are smaller. RunPod's Startup Program gives a Starter tier $1,000 in credits, and a Growth tier $25,000 in bonus credits when you commit $50,000 upfront for 12 months. Modal lists $30 a month of free credit on its Starter plan and separate credit grants for early-stage startups.
A grant is worth what it buys. Here is what $1,000 of credit buys in H100 hours at each listed rate.
| Platform | H100 hours per $1,000 |
|---|---|
| Nodus | 384.6 |
| RunPod Community Cloud | 371.7 |
| Modal | 253.2 |
| RunPod Secure Cloud or Lambda | 250.6 |
The same $1,000 buys about 53% more H100 time at $2.60 than at $3.99.
The 53% comes from 384.6 / 250.6. Our rule: count a credit offer in GPU-hours you can actually use, and include the commitment it asks for. A $25,000 bonus that requires a $50,000 deposit is a discount on a contract, not free compute.
Can you use your AWS or Google Cloud credits and still get recovery?
Yes, with the Beta run-in-your-account option. The Nodus cloud accounts docs describe letting Nodus run your work in a connected AWS account, GCP project or Azure subscription, so Nodus spends your cloud credits before Nodus capacity.
Connecting starts read-only. For AWS, Nodus assumes a role whose policy only lists instances, reads daily cost, reads credit balances and reads GPU quotas. Running work is a separate step with a monthly limit you set.
The payoff is simple. Your jobs land in your credited account first, up to the limit you set, and use Nodus capacity after that.
What protects a small team from runaway GPU spend?
A limit the platform enforces, not a dashboard you check on Monday. A forgotten 8 x H100 machine at $3.99 per GPU-hour costs $31.92 every hour it sits idle, which is $766.08 over a day.
Nodus has two enforced limits, both checked before work starts and every few minutes while it runs. A Budget caps an org or project per month and refuses new work once it is used up. A maxCostUSD on a single Job checkpoints and suspends it when it reaches the cap. This Budget is copied from the budgets guide:
apiVersion: nodus.dev/v1
kind: Budget
metadata:
name: research-monthly
spec:
limitUSD: "500.00"
period: Monthly # a UTC calendar month; Total counts from creation
scope:
project: research # leave out for the whole org
selector:
matchLabels: {team: nlp}
action: Block # Notify only sends notices
thresholds: [50, 80, 100]
notify:
emails: [research-leads@example.com]
With action: Block, new work over the limit is refused with a 402 BudgetExceeded error that names the budget and what is left. For a founder, that turns "how much did we burn on GPUs" into a number you chose in advance. It also makes it safe to let an AI coding agent launch jobs, which we cover in how to cap the GPU spending of an AI coding agent.
What happens when a GPU machine disappears mid-run?
On a plain machine, you rebuild the run yourself. Cheap capacity gets reclaimed, and the startup that loses an 8 x H100 run at hour 8 pays for 8 hours of nothing plus someone's evening.
On Nodus, when a provider reclaims interruptible capacity, Nodus asks your attempt for an urgent checkpoint and prepares a replacement machine at the same time. The new attempt restores your latest committed checkpoint, so you lose only the work since your last save. Your script saves state under /nodus/state and loads it on start, as the checkpoints guide shows.
Our decision rule for startups: if a lost run would cost you more than a day of progress, pick a platform that restores checkpoints for you. If your jobs are short, or a person watches every one, the cheapest pod is fine.
Where are RunPod, Modal, Lambda and the big clouds the better choice?
Each alternative wins a specific kind of startup, and we would tell you to use it.
- Modal is the better choice if your product is an inference API. It is built around Python functions that autoscale, and its startup page pairs credits with direct engineering support. Our best serverless GPU platforms guide compares it with others.
- RunPod is the better choice for hands-on work on cheap pods, or serverless endpoints, and its Community Cloud H100 is close to Nodus on price. Our RunPod alternatives guide covers the tradeoffs.
- Lambda is the better choice once you need a reserved cluster. Its pricing page lists clusters of 16 to 2,000+ H100 or B200 GPUs on terms from 2 weeks to 1 year.
- AWS and Google Cloud are the better choice while large credits last, or when your customers require your stack to run there. A $200,000 or $350,000 grant changes every comparison on this page.
Nodus is the better choice for the common middle case: a small team that trains or fine-tunes often, pays its own bills and cannot afford a lost night or a surprise invoice.
How do you run your first training job on Nodus?
Install the CLI, sign in and submit the command you already run, from the directory that holds your script:
pip install nodus-compute
nodus login
nodus run --gpu H100:8 --max-cost 250 -d -- torchrun train.py
The --gpu H100:8 form, the --max-cost cap and -d to return once the Job exists come from the nodus run reference. Add --dry-run first to see the estimate. Then follow the run with nodus logs -f and copy outputs back with nodus cp, as the quickstart shows. For a sense of the bigger numbers ahead, see how much it costs to train an LLM from scratch.
FAQ
Does Nodus have free credits for startups?
New orgs get a starter credit: $2.00 usable right away after email and card verification, and another $28.00 after a paid top-up or support approval, both expiring 30 days after the grant. GPUs need a paid top-up or review, so the starter credit alone covers CPU trials, as the credits docs explain.
Is RunPod or Nodus cheaper for a startup?
On listed H100 rates they are close: $2.60 on Nodus against $2.69 on RunPod Community Cloud and $3.99 on Secure Cloud. The bigger difference shows up in failed runs, idle pods and spend caps, so price your real job with a dry run.
Should an early startup sign a reserved GPU contract?
Only when you can show weeks of steady usage. Reserved clusters and committed-spend contracts lower the rate but charge you whether or not your GPUs are busy. If your training comes in bursts, pay per run until it does not.
Can one startup use several of these at once?
Yes, and many should. A common split is inference on a serverless platform, training on Nodus, and hyperscaler credits spent first through a connected account.
Originally published on the Nodus blog.
Top comments (0)