GPU budget for a startup: what the first months cost
A startup's GPU bill is GPU-hours times the rate, plus what the provider charges around them. The cheapest on-demand H100 on our index is $1.80/hr, and a 24 GB RTX 4090 starts at $0.35/hr, so the first decision is not the provider but how many GPU-hours the month really needs.
Count GPU-hours before choosing a card
Write down the jobs for the next three months: experiments, a fine-tune, an inference endpoint. An experiment that runs a few hours a day on one card is about 100 GPU-hours a month. A fine-tune on an 8-GPU node for two days is about 400. An endpoint that stays up all month on one card is 720. Then pick the smallest card whose memory holds the model (how much VRAM a model needs); a bigger card than the job needs raises every hour of the bill for no gain.
Plan for the month after the credit
Credits are a head start, not a price. A credit runs out or expires, and the bill after it is the one to plan for. Startup programmes ask for an application and often for a commitment, so read who can apply before counting on one.
Keep the bill to the GPU-hours you meant
Three habits keep a startup's bill close to the plan:
- Stop what you do not use. Some providers keep billing a stopped machine at its full rate; check what "stopped" costs on the fees page before you leave one overnight.
- Watch storage and data out. Model weights and checkpoints stay on disk after the GPU is gone, and moving data out can cost per GB. The storage and egress guide works through a budget.
- Use interruptible capacity for work that can restart. Spot and community rates are lower for the same card; they suit batch jobs that checkpoint (spot, on-demand or serverless).
The figures below come from the latest offers we checked and each provider's own published terms.
What a month of GPU-hours costs
The cheapest on-demand rate for four common cards, and the bill at three volumes: 100 GPU-hours is a month of experiments, 500 is a fine-tune or about three weeks of one card running day and night, 2,000 is about three GPUs running all month. Rates checked 5 November 2026.
| GPU | From | 100 GPU-hours | 500 GPU-hours | 2,000 GPU-hours | |
|---|---|---|---|---|---|
| H100 PCIeHyperAI | $1.80/hr | $180 | $900 | $3,600 | Rent → |
| A100 SXM 40GBJarvislabs | $0.89/hr | $89 | $445 | $1,780 | Rent → |
| L40SRuncrate | $0.54/hr | $54 | $270 | $1,080 | Rent → |
| RTX 4090TensorDock | $0.35/hr | $35 | $175 | $700 | Rent → |
None of these four rates' sources states a node size; each is quoted per GPU. GPU time only, at the rate shown: storage, data transfer and separately priced CPU or RAM are extra (fees by provider). A node of 8 GPUs bills 8 GPU-hours for every hour it runs.
What each of these providers asks before the first machine
| Provider | To start | Billed by | Minimum charge |
|---|---|---|---|
| HyperAI | a balance deposit before buying compute; containers stop when it runs out | not published | not published |
| Jarvislabs | $5 per its FAQ; its home page says $10 | per-minute | no minimum commitment |
| Runcrate | $20 minimum top-up; a launch needs $10 or an hour of the machine, whichever is more | per-second | no minimum |
| TensorDock | $5.00 minimum deposit | per-second | no minimum |
From each provider's own pricing and billing pages, as recorded on the fees page with the day each was read.
What a startup credit buys, and the month after it
RunPod Startup Program: $1,000 credits (Starter tier). Pre-Series A startups with a working product (venture backing strongly preferred), new to Runpod. Growth tier: a $50,000 deposit under a 12-month agreement brings $25,000 in bonus credits ($75,000 in all). Checked 31 October 2026 · source
At RunPod's cheapest on-demand H100 ($2.89/hr, checked 5 November 2026) that credit covers about 346 H100-hours. When it is spent, the same 500 H100-hours a month cost $1,445 at that rate: plan the month after the credit, not only the first one.
The other startup programmes, and whether the credit reaches a GPU
| Programme | Credit | Who can apply | GPU machines |
|---|---|---|---|
| Google for Startups Cloud Programchecked 31 October 2026 | $2,000 (Start tier) to $350,000 (Scale tiers) in Google Cloud credits over two years | Start: founded in the last 2 years; Scale: Pre-Seed to Series A funding | Not named; the credit covers Google's own pay-as-you-go products, and Compute Engine GPUs are billed that way |
| AWS Activatechecked 31 October 2026 | $1,000 to $5,000 (Founders) or up to $200,000 (Portfolio) in AWS credits | Pre-Series B, founded in the last 10 years | Not named; AWS points to EC2 for machine learning, and its credit terms exclude the upfront fee of a Savings Plan or Reserved Instance |
| Microsoft for Startupschecked 31 October 2026 | Up to $150,000 in Azure credits | Not published on its pages | Not named; Microsoft lists Azure Virtual Machines among the services credits cover |
| NVIDIA Inceptionchecked 31 October 2026 | Free membership with cloud credits from NVIDIA and partners; amounts not published | Incorporated, under 10 years old, any funding stage | Neither the amounts nor the partner clouds are published |
A big cloud's credit is spent at that cloud's own rates. At the latest checked rates, $10,000 of credit buys about 1,020 H100-hours at Google Cloud ($9.80/hr), 1,453 at AWS ($6.88/hr) and 1,432 at Microsoft Azure ($6.98/hr); the same amount paid at HyperAI's $1.80/hr buys 5,555. A large credit is still worth having; the month after it is billed at the same rates.
Welcome credits from the big clouds and every other programme, with the full conditions: free GPU credits.