Cheapest GPU Cloud for Indie AI Developers (2026 Prices)
For a bootstrapped AI founder, the cheapest sane GPU is usually a per-second rental on RunPod Community Cloud, a Vast.ai marketplace box or a Hyperstack VM, with Modal's $30 of free monthly compute for bursty work. Before renting anything, check whether a per-token API is cheaper: for a small open model it usually is, by a wide margin.
Key takeaways
- The lowest listed rates we found on 12 Nov 2026 were Hyperstack's NVIDIA A4000 at $0.15/hr and RunPod's RTX A5000 at $0.16/hr on Community Cloud, both from the providers' own pricing pages.
- For a 24GB consumer card, RunPod lists the RTX 4090 at $0.34/hr on Community Cloud and $0.89/hr on Secure Cloud, billed per second.
- Modal's free Starter plan includes $30 of compute every month, which on its per-second rates is roughly 12 hours of an A100 80GB.
- For a small open model, a per-token API can be hundreds of times cheaper than renting a GPU around the clock: DeepInfra lists Llama 3.1 8B Instruct Turbo at $0.02 input and $0.04 output per 1M tokens.
- Most startup GPU credit programs need VC money. RunPod's Starter tier ($1K credits, raised under $1M) and NVIDIA Inception (free, no funding requirement) are the realistic ones for a no-VC founder.
On this page 19
- TL;DR
- The 10 cheapest GPU clouds compared
- Popular GPUs side by side
- Serverless vs pods vs dedicated
- When an API is cheaper than renting a GPU
- How much VRAM you need
- Spot and interruptible: the cheapest, with a catch
- Community clouds and your data
- Storage and egress gotchas
- Startup GPU credits: which are realistic without VC
- Decision table
- Frequently asked questions
- What is the cheapest GPU cloud for indie developers?
- Is it cheaper to rent a GPU or use an LLM API?
- Is RunPod Community Cloud safe for customer data?
- Can a bootstrapped founder get free GPU credits?
- Should I use serverless GPUs or a rented pod?
- Ship it
- Sources
Short answer: The cheapest sane GPU for an indie AI developer in 2026 is a per-second rental on RunPod Community Cloud (RTX 4090 at $0.34/hr), a Vast.ai marketplace box, or a Hyperstack VM (A4000 at $0.15/hr). For short, bursty jobs, Modal's free $30 of compute each month can cost you nothing. Before you rent anything, run the API arithmetic below: for small open models, paying per token usually wins.
TL;DR#
- Lowest named hourly rates: Hyperstack NVIDIA A4000 at $0.15/hr (Hyperstack) and RunPod RTX A5000 at $0.16/hr on Community Cloud (RunPod).
- Best 24GB card for the money: RunPod RTX 4090, $0.34/hr Community or $0.89/hr Secure, billed per second.
- Best free tier: Modal Starter, "$30 / month free credits" with per-second billing (Modal).
- Often cheaper than all of them: a per-token API. Llama 3.1 8B Instruct Turbo is $0.02/$0.04 per 1M input/output tokens on DeepInfra.
Disclosure: ZeroRaised is a launch site for bootstrapped products. We have no commercial relationship with any provider below and earn nothing if you sign up. Every price comes from the provider's own page as checked on 12 Nov 2026, and advice marked "Our read" is ours. GPU prices change often, so recheck on the day you rent.
The 10 cheapest GPU clouds compared#
Each rate is the provider's own published price for a named GPU. Vast.ai is a marketplace, so it has no fixed list price.
| Provider | Example GPUs | Cheapest listed rate (named GPU) | Billing | Serverless | Free credits | Best for |
|---|---|---|---|---|---|---|
| RunPod | A5000, 4090, L40S, A100, H100 | RTX A5000 $0.16/hr (Community) | Per second | Yes | Startup program | Cheap pods plus serverless in one account |
| Vast.ai | Host-listed | Set by hosts, varies live | Per second | Not covered here | Spend match | Lowest prices if you can shop around |
| Hyperstack | A4000, A6000, L40, A100, H100 | NVIDIA A4000 $0.15/hr | Per minute | Not on pricing page | None listed | Cheap VMs with free egress |
| Modal | T4, L4, A10, L40S, A100, H100 | T4 $0.000164/sec (about $0.59/hr) | Per second | Yes | $30/month on Starter | Bursty Python jobs and inference |
| Hugging Face | T4, L4, L40S, A10G, A100 | Spaces T4 small $0.40/hr | Per minute | Endpoints | PRO $9/mo includes $2 inference credits | Demos and hosted endpoints |
| Paperspace (DigitalOcean) | RTX4000, A4000, A6000, H100 | RTX4000 $0.56/hr | Not covered here | Not on pricing page | None listed | Simple GPU machines |
| Lambda | RTX 6000, A10, A6000, A100, H100, B200 | Quadro RTX 6000 $0.69/hr | Not stated | Not on pricing page | None listed | Plain on-demand instances |
| DigitalOcean | RTX 4000 Ada, L40S, H100, MI300X | RTX 4000 Ada $0.76/hr | Per second, 5-min minimum | Not on pricing page | None listed | Teams already on DigitalOcean |
| Replicate | T4, L40S, A100, H100, H200 | T4 $0.81/hr ($0.000225/sec) | Per second | Yes | None listed | Running public models by the second |
| Together AI | H100, H200, B200 | H100 cluster $1.99/GPU/hr (preemptible) | Not covered here | Per-token API | None listed | Big GPUs and per-token inference |
Notes on the table:
- Lambda prices above are the 1x instance rates that its page says take effect on 16 Nov 2026; until then it asks you to log in to see current pricing. Every Lambda row is "Plus applicable sales tax/VAT/GST."
- Modal's hourly figures in this post are our conversions of its per-second rates (x 3,600).
- Paperspace's page now says "Paperspace is now part of DigitalOcean!"
Popular GPUs side by side#
The same card can cost several times as much depending on where you rent it.
| GPU | RunPod Community | RunPod Secure | Others |
|---|---|---|---|
| RTX 4090 (24GB) | $0.34/hr | $0.89/hr | |
| L40S (48GB) | $0.79/hr | $1.09/hr | DigitalOcean $1.57/hr; Modal about $1.95/hr; HF Spaces $1.80/hr; Replicate $3.51/hr |
| A100 80GB | from $1.19/hr (PCIe) | $1.79/hr | Modal about $2.50/hr; Replicate $5.04/hr |
| H100 | from $1.99/hr (PCIe) | $2.89/hr (PCIe) | Hyperstack $2.50/hr (PCIe); Lambda $3.79/hr (PCIe); Modal about $3.95/hr (SXM); DigitalOcean $4.41/hr |
Sources: RunPod, DigitalOcean, Modal, Hugging Face Spaces GPUs, Replicate, Hyperstack, Lambda.
Our read: Replicate and Modal cost more per hour because you pay only while code runs. RunPod Community and Vast.ai are cheap because you rent third-party hardware with fewer guarantees.
Serverless vs pods vs dedicated#
There are three ways to pay for a GPU, and the cheapest one depends on how busy the GPU will be.
- Serverless (Modal, RunPod Serverless, Replicate). You pay per second while a request or job runs. RunPod lists a serverless RTX 4090 at $1.34/hr, about four times its Community pod rate. Watch Replicate's private models: they are billed for setup and idle time too (Replicate).
- Pods or VMs (RunPod, Vast.ai, Hyperstack, Lambda, DigitalOcean). You pay for every second the machine is on, busy or not.
- Dedicated endpoints (Together AI, Hugging Face). Capacity for steady production traffic. Together's dedicated H100 is $5.49/GPU/hr; a Hugging Face Endpoint starts at $0.50/hr for one T4 (HF Endpoints).
Our rule of thumb: if the GPU will be busy for less than about a quarter of each hour, serverless is usually cheaper. Above that, rent a pod and shut it down when you are done.
When an API is cheaper than renting a GPU#
This is the question to answer first. Our guide on how to reduce LLM costs covers the API side in depth; here is the GPU comparison.
Worked example (our arithmetic, list prices from 12 Nov 2026). Your app sends 50M input tokens and 10M output tokens a month to Llama 3.1 8B.
- API: DeepInfra lists Llama 3.1 8B Instruct Turbo at $0.02 per 1M input and $0.04 per 1M output tokens. That is 50 x $0.02 + 10 x $0.04 = $1.40 a month.
- Rented GPU, always on: a RunPod Community RTX 4090 at $0.34/hr for a 730-hour month is $248.20 a month, before storage.
- Break-even: at that 5:1 input-to-output mix, the blended API price is about $0.023 per 1M tokens, so you would need roughly 10.6 billion tokens a month before the always-on GPU wins on price alone.
The gap narrows for bigger models (DeepInfra lists Llama 3.3 70B Instruct Turbo at $0.10/$0.32 per 1M tokens). Renting makes sense when you serve a fine-tuned or custom model no API offers, run large batch jobs that keep a pod busy for hours, cannot send data to a third party, or are training.
If you only want to experiment, you may not need the cloud at all: see how to run an LLM locally.
How much VRAM you need#
The model has to fit in GPU memory. Hugging Face's model memory estimator shows total size halving as you move from float32 to float16 to int8 to int4 (418.18 MB, 206.59 MB, 103.29 MB and 51.65 MB for bert-base-cased). It warns that it measures memory "to purely load the model in, not to perform inference", and that "When performing inference you can expect to add up to an additional 20%".
Our arithmetic: 16-bit weights take 2 bytes per parameter, so an 8B model needs about 16GB to load, plus up to 20% for inference, which is about 19GB. That fits on a 24GB RTX 4090. A 70B model at 16-bit is about 140GB, so you need two 80GB cards or a 4-bit quantised version.
Spot and interruptible: the cheapest, with a catch#
Interruptible capacity is the deepest discount, and the catch is the same everywhere: you can lose the machine.
- Vast.ai: interruptible instances are "Lowest cost (often 50%+ cheaper than on-demand)" but "May be paused if outbid or if on-demand requested" (Vast.ai docs).
- Hyperstack: spot H100 PCIe is $2.00/hr against $2.50 on-demand.
- Together AI: preemptible H100 cluster capacity is $1.99/GPU/hr against $3.99 on-demand.
Our read: use interruptible capacity for training and batch work that checkpoints often. Never put a live user-facing endpoint on it.
Community clouds and your data#
Cheap marketplaces rent you hardware owned by third parties. Here is what the providers say about it:
- RunPod: "Community Cloud: Connects individual compute providers to users through a vetted, secure peer-to-peer system", while "Secure Cloud: Operates in T3/T4 data centers, providing high reliability and security for enterprise and production workloads" (RunPod docs).
- Vast.ai: it "connects compute providers, from hobbyists to Tier-4 datacenters, with users who need GPUs" and offers a Secure Cloud filter for data-center hosts (Vast.ai docs).
Our read: experiment on community hosts. Customer data, API keys and proprietary weights belong on Secure Cloud or a data-center provider.
Storage and egress gotchas#
The GPU rate is not the whole bill. From the providers' own pages:
- RunPod: volume disk is $0.10/GB/mo while running but $0.20/GB/mo while idle. RunPod says "No fees for ingress/egress."
- Vast.ai: storage is charged "continuous while the instance exists, including when it is stopped", and bandwidth is charged "per byte for both ingress and egress" at host-set rates (Vast.ai pricing docs).
- Modal: network egress is $0.04/GiB after 1 TiB free each month on Starter.
- DigitalOcean: billed per second "with a 5-minute minimum billing period", and powered-off Droplets are still billed.
- Hugging Face Spaces: upgraded Spaces "run indefinitely by default, even if there is no usage" unless you set a sleep time.
- Hyperstack: egress and ingress traffic are listed as free, and its FAQ says "there are no additional charges for bandwidth or data transfer."
Our read: the most common surprise bill is a forgotten idle pod or volume. Set a shutdown script or calendar reminder the day you create one.
Startup GPU credits: which are realistic without VC#
| Program | What you get | Eligibility | Realistic without VC? |
|---|---|---|---|
| Modal Starter plan | $30 of compute per month | Anyone, no monthly fee | Yes |
| RunPod Starter | $1K in credits, expire after 1 year | "Raised under $1M"; "a real AI workload" | Yes, likely, but RunPod says it may verify funding, and existing customers are not eligible |
| NVIDIA Inception | Partner cloud credits (amounts not stated), hardware pricing, training | Free; incorporated, working website, under 10 years old; no funding or revenue requirement | Yes, if incorporated |
| Vast.ai Startup Program | Credits matching your spend "dollar for dollar" up to an approved amount | Active training or inference workload; not for "free compute with no committed spend" | Only if you are already spending |
| RunPod Venture | $5K in credits | "Backed by a verified VC" | No |
| Modal startups | Credits, amount not stated | VC funding from a partner fund, or over $1M from any fund | No |
For non-GPU perks such as hosting and software deals, see our list of free startup perks without VC.
Decision table#
This is our editorial advice based on the prices above.
| Your situation | Our pick |
|---|---|
| Calling an open model from your app | Per-token API first (e.g. DeepInfra, Together AI) |
| Occasional experiments, a few hours a month | Modal Starter ($30 free each month) |
| Fine-tuning a 7-8B model on public data | RunPod Community RTX 4090 or Vast.ai interruptible |
| Batch jobs that checkpoint | Vast.ai interruptible or Hyperstack spot |
| Spiky production inference for a custom model | Modal or RunPod Serverless |
| Steady production inference with customer data | RunPod Secure Cloud or Hyperstack |
| Large egress (serving files or media) | Hyperstack or RunPod (no egress fees listed) |
| Short H100 training run | RunPod Community from $1.99/hr or Together preemptible at $1.99/GPU/hr |
Frequently asked questions#
What is the cheapest GPU cloud for indie developers?#
On list prices checked on 12 Nov 2026, Hyperstack's NVIDIA A4000 at $0.15/hr and RunPod's RTX A5000 at $0.16/hr on Community Cloud were the lowest named rates. Vast.ai can be cheaper still because hosts set their own prices, but its rates move with supply and demand. For bursty jobs, Modal's free $30 of compute each month may cost you nothing at all.
Is it cheaper to rent a GPU or use an LLM API?#
For most indie apps, the API is cheaper. In our worked example, 50M input and 10M output tokens a month on Llama 3.1 8B at DeepInfra's listed rates costs about $1.40, while a RunPod Community RTX 4090 left on all month costs about $248. Renting wins when you have steady, heavy traffic, a custom or fine-tuned model no API serves, or data you cannot send to a third party.
Is RunPod Community Cloud safe for customer data?#
RunPod describes Community Cloud as connecting individual compute providers to users through a vetted, secure peer-to-peer system, and Secure Cloud as running in T3/T4 data centers for enterprise and production workloads. Our read: experiment on Community Cloud, but put production workloads that touch customer data on Secure Cloud or a data-center provider.
Can a bootstrapped founder get free GPU credits?#
Some. Modal's Starter plan includes $30 of compute a month with no fee. RunPod's Starter tier gives $1K in credits to startups that have raised under $1M. NVIDIA Inception is free and does not require funding or revenue. Modal's startup credits and RunPod's Venture tier require VC funding, and Vast.ai's program matches spend rather than giving free compute.
Should I use serverless GPUs or a rented pod?#
Use serverless when traffic is spiky or low, because you pay only while a request runs. Use a pod or VM when the GPU is busy most of the hour, because the hourly rate is lower. On RunPod, a serverless RTX 4090 is listed at $1.34/hr against $0.34/hr for a Community Cloud pod.
Ship it#
Once your model is running at a price you can live with, read our guide on how to launch an AI app, then submit your product to ZeroRaised for free. For more cost breakdowns for founders building without investors, join the ZeroRaised newsletter.
Sources#
- RunPod pricing, last checked 12 Nov 2026
- RunPod Startup Program, last checked 12 Nov 2026
- RunPod Pods overview (docs), last checked 12 Nov 2026
- Vast.ai pricing, last checked 12 Nov 2026
- Vast.ai pricing docs, last checked 12 Nov 2026
- Vast.ai instance types docs, last checked 12 Nov 2026
- Vast.ai docs, last checked 12 Nov 2026
- Vast.ai Startup Program, last checked 12 Nov 2026
- Lambda pricing, last checked 12 Nov 2026
- Modal pricing, last checked 12 Nov 2026
- Modal for startups, last checked 12 Nov 2026
- Replicate pricing, last checked 12 Nov 2026
- Together AI pricing, last checked 12 Nov 2026
- DigitalOcean GPU Droplets pricing, last checked 12 Nov 2026
- Hyperstack GPU pricing, last checked 12 Nov 2026
- Paperspace pricing, last checked 12 Nov 2026
- Hugging Face pricing, last checked 12 Nov 2026
- Hugging Face Spaces GPU docs, last checked 12 Nov 2026
- Hugging Face Inference Endpoints pricing, last checked 12 Nov 2026
- Hugging Face model memory estimator, last checked 12 Nov 2026
- DeepInfra pricing, last checked 12 Nov 2026
- NVIDIA Inception, last checked 12 Nov 2026
Frequently asked questions
What is the cheapest GPU cloud for indie developers?
On list prices checked on 12 Nov 2026, Hyperstack's NVIDIA A4000 at $0.15/hr and RunPod's RTX A5000 at $0.16/hr on Community Cloud were the lowest named rates. Vast.ai can be cheaper still because hosts set their own prices, but its rates move with supply and demand. For bursty jobs, Modal's free $30 of compute each month may cost you nothing at all.
Is it cheaper to rent a GPU or use an LLM API?
For most indie apps, the API is cheaper. In our worked example, 50M input and 10M output tokens a month on Llama 3.1 8B at DeepInfra's listed rates costs about $1.40, while a RunPod Community RTX 4090 left on all month costs about $248. Renting wins when you have steady, heavy traffic, a custom or fine-tuned model no API serves, or data you cannot send to a third party.
Is RunPod Community Cloud safe for customer data?
RunPod describes Community Cloud as connecting individual compute providers to users through a vetted, secure peer-to-peer system, and Secure Cloud as running in T3/T4 data centers for enterprise and production workloads. Our read: experiment on Community Cloud, but put production workloads that touch customer data on Secure Cloud or a data-center provider.
Can a bootstrapped founder get free GPU credits?
Some. Modal's Starter plan includes $30 of compute a month with no fee. RunPod's Starter tier gives $1K in credits to startups that have raised under $1M. NVIDIA Inception is free and does not require funding or revenue. Modal's startup credits and RunPod's Venture tier require VC funding, and Vast.ai's program matches spend rather than giving free compute.
Should I use serverless GPUs or a rented pod?
Use serverless when traffic is spiky or low, because you pay only while a request runs. Use a pod or VM when the GPU is busy most of the hour, because the hourly rate is lower. On RunPod, a serverless RTX 4090 is listed at $1.34/hr against $0.34/hr for a Community Cloud pod.