Free GPU credits: the 3 that actually buy GPU time

Short versionDecision ledger

Rent	When	Example
Monthly serverless credit	You want recurring free GPU-hours: Modal includes $30 per month on a platform that meters GPUs per second. It is the only one left since Beam withdrew its credit.	Modal
GPU-exampled trial	You want a one-time budget whose GPU coverage the vendor illustrates itself: OVHcloud publishes GPU examples for its US$200 trial.	OVHcloud
Student credit	You have a US student email: Thunder Compute gives $20, usable on A6000 and L40 instances only.	Thunder Compute
Everything else	The famous credits buy zero GPU minutes: Daytona excludes GPU sandboxes, Novita is CPU-scoped, hyperscaler trials are quota-gated or upgrade-gated.	Daytona

Grant size and GPU access are uncorrelated. Read the exclusion line before the dollar amount.Checked 2026-08-21

Most free cloud credits cannot buy a single GPU minute. Across the 55 free credits and free tiers tracked on this site, exactly three have an official path to GPU hardware: Modal’s $30 / month, metered per second on a platform whose published pricing includes GPUs; OVHcloud’s US$200 trial, the only credit in the set whose official pages illustrate GPU usage; and Thunder Compute’s $20 for US students, scoped to A6000 and L40 instances. This was four until September 2026: Beam published $30 per month on the same terms as Modal, and withdrew it, which is covered in the null set below. The biggest sandbox credits, Daytona’s $200, E2B’s $100, and Novita’s $100, buy zero GPU time: Daytona’s billing docs exclude GPU sandboxes from the free balance, E2B and Vercel publish no GPU SKU at all, and Novita’s credit is officially scoped to Sandbox CPU and RAM. The hyperscaler trials are shaped the same way: Google Cloud blocks adding GPUs during the trial, Azure documents GPU SKUs at zero quota with no increases for trial subscriptions, AWS leaves GPU instances off the free plan’s service list, and Oracle’s Always Free tier has no GPU. Two general clouds, Civo at $250 and Scaleway at 100 euros, leave GPU eligibility open on paper without quantifying it, and the classic GPU clouds, RunPod, Vast.ai, Lambda, and TensorDock, offer no free credit at all.

This page is the analysis layer on top of the credits index. Every figure below has a dated per-provider entry with the official sources; this article answers the one question the entries answer only individually: which of these grants can actually pay for GPU hardware, and for how many hours. Sandbox comparison sites cover feature matrices well. What they do not cover is the intersection this page exists for: which sandboxes and platforms have GPUs at all, which credits are allowed to touch them, and what the official pages said on a named date. Nothing here is a benchmark, and nothing here is hands-on tested yet; this is documentation analysis with the receipts listed in the sources block.

The matrix

One row needs an immediate caveat. The Modal GPU-hour figures are ceiling arithmetic: the platform publishes per-second GPU rates and describes the $30 as credit against usage, and publishes no exclusion keeping the credit away from GPUs, the way Daytona explicitly does. But it does not itemize exactly which line items the credit covers, either. The per-provider sections below quote the official wording on both sides so you can judge the gap yourself.

Method

What counts as a source here: official first-party pages only. That means pricing pages, documentation, signup and registration pages, and legal terms published by the vendor. Blog posts, coupon roundups, YC threads, and Reddit do not count, no matter how often they are repeated. Every number in this article traces to a URL in the sources list, and each named provider has a dated entry on the credits index with the same sourcing.

Check date: every claim was verified against the official pages on 2026-08-21, with two exceptions. E2B’s figures carry its entry’s check date of 2026-08-20, and were not re-verified for this article. Beam was reread on 2026-09-09 after the daily source check found its published figure gone, and its rows and the headline count reflect that date; nothing else on this page was rechecked that day.

Multi-path verification: where a vendor states the same fact on more than one official page, I quote both, including when the two pages phrase it differently (Civo’s “31 calendar days” versus “a full month”, Google’s “90 days” versus “3 months”). Where a fact appears on exactly one official page, I say so explicitly rather than letting it read as well-attested.

Fail-closed rule: anything the official pages do not state is written as “Not stated officially”. Nothing in this article “should work” or “probably works”. Where I have not run the workload myself, the entry says “Not tested yet”, and that is currently true of every provider on this page.

The math, so you can recompute it: serverless platforms publish per-second rates, so the hourly rate is the per-second rate multiplied by 3600, and the free GPU-hours are the credit amount divided by that hourly rate. Example: Modal lists a T4 at $0.000164 per second; $0.000164 x 3600 = $0.5904 per hour; $30 / $0.5904 = 50.8 hours. Figures are rounded to one decimal. Where a platform bills per hour directly, the division is straight: $20 / $0.35 = 57.1 hours. These are ceilings for the GPU line item alone: real workloads also bill CPU, memory, and storage, so treat every hours figure as the maximum the credit could buy if it went to the GPU meter and nothing else.

The taxonomy: three classes of credit

Every grant in the dataset falls into one of three classes, and the class matters more than the dollar amount:

1. Credits with an official path to GPU time. Three qualify: Modal, OVHcloud, Thunder Compute. Two more, Civo and Scaleway, are officially not excluded but also not confirmed. Beam qualified until September 2026 and no longer does.
2. Credits on platforms that have GPUs the credit cannot touch. The most misleading class, because the vendor’s GPU pricing sits right next to the credit offer: Daytona, Novita, Northflank, Crusoe, and all four hyperscaler trials.
3. Credits on platforms with no GPU to buy. The credit may be excellent for CPU work, but there is no GPU SKU on the official pricing at all: E2B, Vercel Sandbox, Railway, Render, Fly.io.

There is also a null set worth naming, because it is where most “free GPU credit” blog claims point: GPU platforms with no free credit whatsoever. RunPod, Vast.ai, Lambda, TensorDock, Salad, Cerebrium, Nebius, Prime Intellect, and Shadeform are covered at the end.

Class 1: credits that can buy GPU time

Modal: $30 per month, the deepest GPU menu

Modal’s Starter plan costs $0 plus compute and includes $30 per month of credit. That figure is unusually well attested: the pricing page lists it as included compute on the Starter plan, the signup page says “$30 free monthly compute!”, and Modal’s own llms.txt says “sign up for free and get $30/month of credits.” Three official statements, one number. The billing docs add the access condition, in the vendor’s own words: “you must have a payment method on file in order to use Modal.” The official word is payment method, not credit card.

What the $30 buys, at the official per-second rates on the pricing page:

Now the honesty caveat, which the Modal entry spells out: Modal’s pages call the $30 “free credits” and “included compute”, and the pricing page lists GPU, CPU, memory, and storage as separate line items, but no official page itemizes which line items the credit applies to. So “$30 of free GPU” is not an official sentence. What is official: the credit exists, the GPU rates exist on the same page, and no exclusion is published. That is materially different from Daytona below, where the exclusion is published. Two more Starter facts that shape real usage: concurrency caps at 10 GPUs, and whether unused credit rolls over is not stated officially.

OVHcloud: US$200 with the only official GPU examples

OVHcloud’s Public Cloud free trial is US$200, valid one month from the creation of your first Public Cloud project, with a valid payment method required before the trial starts. It is the only credit in this dataset whose official pages state GPU examples of what the credit covers: 35 hours of an AI1-1-GPU notebook, or 5 hours of AI Training on 4 nodes. That sentence does the work every other vendor leaves to inference: the vendor itself says the credit buys GPU time, and quantifies it. Quotas still apply, so larger configurations may need a quota request.

The catch is the exit: usage past the US$200 is billed automatically to your payment method. Most trials in this set end in suspension; OVHcloud’s ends in a bill. If you want a hard stop at zero, you delete your resources before the credit runs out. Details on the OVHcloud entry.

Thunder Compute: $20, US students only

Thunder Compute’s students page offers “$20 of free Thunder Compute credit” to signups with a US student email address, verified automatically by domain. The scope is explicit: usable on A6000 and L40 instances only, not on A100 at $1.09 per hour or H100 PCIe at $2.19 per hour. At the official rates, $20 / $0.35 = ~57 A6000 hours, and $20 / $0.79 = ~25 L40 hours. For eligible students that is the largest verified pile of free GPU-hours in this article.

Expiry is genuinely ambiguous, and the ambiguity is visible in the official docs themselves. The billing docs make two separate statements: “Preloaded credit does not expire” and “Promotional credit can be revoked at our discretion.” Which bucket the student $20 falls into is not stated officially, so the Thunder Compute entry reports the expiry as not stated rather than picking a side.

The gray zone: Civo and Scaleway

Two credits sit between classes, and the fail-closed rule keeps them out of the headline four. Civo gives new accounts $250; the legal terms say 31 calendar days, the marketing says “a full month”, and both are official so I quote both. Civo lists L40S, A100, H100, H200, and B200 GPU instances on public pricing, and the trial text does not exclude GPU from credit eligibility. Not excluded is not the same as confirmed, and no official page states GPU hours for the credit, so the Civo entry prints no number and neither will I. The same logic applies to Scaleway’s 100 euro voucher for new Professional accounts: GPU instances are sold at standard billing, the trial FAQ does not exclude them, organization quotas apply, and nothing quantifies GPU-on-credit officially. Both providers require a valid payment method, and both reject virtual or prepaid cards. If a reader burns either credit on GPU instances and documents it, that would upgrade these rows from “not excluded” to “tested”; as of 2026-08-21, not tested yet.

An aside: the free GPU that involves no credit

One provider hands out GPU access with no dollar balance attached. Paperspace’s free plan includes a Free GPU machine, an M4000, “for Notebooks in private workspaces only”, with 5 GB of storage, per the DigitalOcean-hosted pricing docs. There is no credit to spend down and no expiry date; session-hour caps are not stated officially. An M4000 is old hardware, but it is a real CUDA device at $0, which is more than most $200 credits in this article deliver. Details on the Paperspace entry.

Class 2: platforms with GPUs the credit cannot touch

This is the class that generates the misleading blog posts, because the vendor’s GPU price list sits on the same site as the credit offer, and nothing stops a roundup from multiplying one by the other.

Daytona: $200 that buys zero GPU minutes

Daytona’s pricing page says “$200 in free compute included” and “no credit card required”, and the same site publishes GPU sandbox rates: H200 at $2.61 per hour, H100 at $2.27, RTX PRO 6000 at $1.74, RTX 5090 at $0.74, RTX 4090 at $0.57. Divide naively and you would print “$200 = 88 free H100 hours”. The billing docs close that door in five words: the free credit balance is “not available for GPU sandboxes.” So the honest multiplication is $200 x 0 GPU eligibility = 0 GPU hours. What the credit does buy is substantial CPU sandbox time, at $0.0504 per vCPU hour and $0.0162 per GiB hour of memory, billed per second. Two operational notes from the billing docs: GPU sandboxes are auto-deleted when stopped, and stopped non-GPU sandboxes still bill for disk. Full breakdown on the Daytona entry.

This is the clearest example in the dataset of why the exclusion line matters more than the amount. Daytona’s $200 is one of the largest no-card credits anywhere, genuinely useful for agent workloads, and worth exactly nothing for the GPU rates published one page away.

Novita AI: $100 scoped to Sandbox CPU and RAM

Novita’s registration page words the offer as “Get $100 in Sandbox credits, valid for 90 days. No credit card required.” Novita is known as a GPU cloud and inference provider, which is exactly why the scoping matters: the credits are tied to the Sandbox product, which meters $0.0000098 per vCPU-second and $0.0000032 per GiB-second, with free-tier caps of 5 concurrent sandboxes, 1-hour max sessions, and 2 vCPU with 4 GB per sandbox. I found no official statement that the $100 spends on GPU instances or the model API, so the Novita entry treats GPU coverage as officially absent. A second, vaguer offer exists: the quickstart doc mentions a “voucher with some credit” with no figure, single-source and unquantified, so it stays out of the math entirely.

Northflank: GPUs behind a $50 purchase, no grant at all

Northflank has no signup cash credit; its free offering is the Developer Sandbox plan with 2 services, 2 jobs, and 1 addon. The platform publishes GPU rates, L4 at $0.80 per hour up to H100 at $2.74 and RTX PRO 6000 96GB at $3.00, but the credits docs state that GPU workloads require a $50 or greater credit purchase to unlock access, and that on Northflank credits are something you buy, minimum $50, maximum $10,000. Add the account-wide rule that all users must add a payment method to create resources, regardless of plan, and the free GPU math is simple: the cheapest possible Northflank GPU minute costs $50 up front. GPU services also stop immediately when credits run out. Details on the Northflank entry.

Crusoe: $5 for serverless tokens, not GPU VMs

Crusoe’s docs banner offers “$5 in free credits to explore Serverless Inference and Serverless Fine-Tuning”. Those products bill per token, not per GPU-hour, so there is no honest GPU-hour conversion and I will not print one. The GPU VM rates on Crusoe’s pricing pages, H100 at $5.50 managed or $3.90 on-demand infrastructure, are context for what the $5 is not for. Account creation requires a “valid, non-prepaid credit card”, official wording that is stricter than the generic payment method phrasing elsewhere. The $300 Crusoe credit that circulates in roundups is not on any official Crusoe page as of 2026-08-21; the Crusoe entry documents that correction.

The hyperscalers: four trials, four different GPU locks

The four big-cloud trials all have GPUs on the platform and all keep the trial money away from them, each by a different mechanism. The mechanisms are worth distinguishing because they fail differently.

Google Cloud, $300 over 90 days: an explicit action block. The free trial docs state that during the trial you cannot add GPUs to your VM instances and cannot request a quota increase. The documented path is upgrading to a paid billing account first; remaining credit can still be spent after the upgrade, subject to GPU quota. Note the two official time phrasings: the docs and marketing say 90 days, the terms say 3 months. Both quoted on the Google Cloud entry.

Azure, $200 over 30 days: a quota gate at zero. Microsoft’s NC-family GPU docs say some GPU SKUs have a vCPU quota that is initially 0 and that quota increases “aren’t available to trial subscriptions.” For those documented SKUs the trial cannot obtain GPU quota, so the $200 cannot buy that GPU time. Whether every GPU SKU is quota-0 on trial is not stated officially in one place, so the claim stays scoped to the documented families. The Azure credit also cannot buy Spot VMs at all. Details on the Azure entry.

AWS, up to $200 over 6 months: a service list that omits GPUs. The new AWS Free Tier grants $100 at signup plus up to $100 more for exploring key services, on a free plan limited to select services. GPU and accelerated EC2 instances are not listed among them; full access requires the Paid plan, where GPU quotas are handled separately. The AWS entry covers the sharper edge: the free plan account is closed if you do not upgrade within 6 months.

Oracle, US$300 over 30 days: silence plus a GPU-free Always Free tier. Oracle’s Always Free tier has no GPU, officially. For the $300 trial, the FAQ says the credit applies to “any eligible OCI service”; GPU is neither listed as included nor excluded, and capacity limits apply. Single-source and unresolved, so the Oracle entry prints no GPU promise either way. IBM’s $200 has the same shape, unlocking only after a Pay-as-you-go upgrade, with GPU coverage not stated; and Alibaba Cloud’s per-product trials have no GPU card in the free-trial catalog at all. UpCloud splits cleanly in two: the standard-trial docs table says “GPU Servers, Not available”, while the $500 / €500, 30-day promotional trial its signup flow advertises states nothing about GPU either way (rechecked 2026-08-24; the rest of this page reflects the 2026-08-21 check), so no GPU promise and no GPU-hours math survives the sourcing rule; both layers are documented on the UpCloud entry. Vultr sells Cloud GPU but states no trial dollar amount and no GPU coverage, so there is nothing to compute.

Class 3: credits on platforms with no GPU to buy

These grants are not defective, they are just answering a different question. If your agent workload is CPU-bound, several of these are the best money in the dataset. For GPU purposes they are zeros, and the zero is structural: there is no GPU SKU on the official pricing to spend anything on.

E2B’s Hobby plan includes a one-time $100 usage credit, no credit card required to start, billed per second while sandboxes run. The pricing and billing pages cited on its entry list no GPU SKU (checked 2026-08-20, the one date on this page older than the rest). When the credit runs out, the FAQ is blunt: your account is blocked until you add a payment method.

Vercel Sandbox on the Hobby plan includes 5 hours of active CPU per month, 420 GB-hours of provisioned memory, and 5,000 sandbox creations, at up to 4 vCPUs and 8 GB per sandbox, sessions capped at 45 minutes. No GPU SKU appears on the Sandbox pricing page. Hobby is also restricted to non-commercial personal use, which rules out most agent products anyway.

Railway states it plainly: “Railway doesn’t currently offer GPU instances.” The trial is $5 for 30 days, then a $1 monthly allowance. Render’s free instance is 512 MB RAM and 0.1 CPU with 750 instance hours per month; GPU does not appear on its official pricing, and its GPU doc URLs returned 404 on the check date. Fly.io’s trial is 2 hours of machine runtime or 7 days, whichever comes first; its GPU documentation URLs also returned 404 on 2026-08-21, and neither the pricing nor trial page advertises GPU machines, so no GPU claim survives the sourcing rule. Fly.io has the strangest card rule in the dataset: adding a card ends the trial immediately.

The null set: GPU platforms with no free credit

Beam, moved here on 2026-09-09. This is the reason the headline count went from four to three. Beam’s pricing page carried “$30 Free Credit, Every Month” and the wording “refreshed monthly”, verified here on 2026-08-21, 2026-08-29 and 2026-09-02. On 2026-09-09 the daily source check could not find the figure on any page this site cites, and a manual pass over the pricing page, both docs pages, the homepage and the docs llms.txt index found no free credit wording anywhere on any of them. The pricing page had been rebuilt around per-second rates with no plan tiers at all, so the $0 base plan and the $89 Team plan went with it. Storage stays free to 1TB and that is the whole free component now. Whether the allowance survives inside the signup flow, or for accounts that already had it, is not knowable from public pages, and this site reports what public pages state. Full detail on the Beam entry.

The largest class of “free GPU credit” claims on the wider internet points at platforms that officially give nothing. Naming them matters, because each has a number circulating that its own pages do not support as of 2026-08-21.

RunPod: no signup credit; the circulating $10 is the official minimum deposit, “start with as little as $10”, your money. Vast.ai: the docs state “The minimum deposit amount on Vast.ai is $5”; the circulating $5 free is that deposit. Lambda: no signup credit; the one official free path is a research grant of “up to $5,000 in cloud credits” for qualifying researchers, by application. At Lambda’s official rates that would be roughly 1,253 hours on an 8x H100 SXM node at $3.99 per GPU-hour, but only if awarded. TensorDock: “start with just $5” is a deposit; circulated $1, $5, and $15 promo credits are not on official pages. Salad: the pricing page says Container Engine does not have a free trial; the only free item is a 5-audio-hour transcription trial for a separate product. Cerebrium: the Hobby plan’s “Free + compute” means the plan fee is free and every compute second is billed; the circulating $10 and $30 signup credits are not on official pages. Nebius: the free trial program was suspended as of July 13, 2026, new sign-ups not accepted. Prime Intellect and Shadeform: promo-code fields and wallet top-ups respectively, with circulated credits ($50 ICLR, $300 Shadeform) not found on live official pages. Koyeb: GPU Instances are sold on the platform and the free Instance is not one of them; the monthly credit that older lists still cite is discontinued, per its own pricing FAQ.

The pattern across the null set: a deposit minimum gets retold as a gift, or a dead promotion outlives its official page. If a number is not on the vendor’s own site on a date you can name, treat it as expired.

Limitations: what this article could not verify

Fail-closed applies to the article itself, so here is what is missing and why.

Baseten. Official pages say “New workspaces receive credits for testing and deployment”, with no dollar amount anywhere on the pricing or billing pages. Baseten runs GPU inference, so an unquantified credit on a GPU platform is exactly the intersection this article covers, and I cannot put a number on it. The circulating $30 figure is not official. Amount: not stated officially.

Lightning AI. The docs say eligible free plan users receive complimentary credits, no credit card required, with one free 4-CPU Studio; the dollar figure and the GPU-hour allowance circulating for the free tier ($15, 80 GPU-hours) are not on a crawlable official page as of the check date, because the pricing page renders as a JavaScript shell. A platform that likely belongs in Class 1 stays out of the matrix for lack of a quotable official number.

Google AI Studio and the Gemini API free tier. The official rate-limit and billing pages redirected to authentication during research, so nothing could be verified first-party. Third-party numbers conflict, so none are printed. This gap is about token-based API access rather than GPU-hours, but readers keep asking about it in the same breath, so the non-answer is recorded here.

Oracle trial GPU, Civo GPU hours, Scaleway GPU hours, Vultr everything. Each is an open question the vendor could close with one sentence on an official page. Until then: not stated officially.

And the standing limitation: every “what you can do” judgment on this page derives from official documents, not from spending the credits. Not tested yet, on every row. When a credit gets burned down for real, the per-provider entry gains a tested section and this article inherits it.

Verdict

And the general rules, which will outlive the specific numbers. Choose a monthly credit over a one-time credit when you expect to iterate: $30 every month beats $200 once by month seven. Choose the credit whose GPU eligibility is written down over the bigger credit with silence attached: OVHcloud’s exampled US$200 is worth more GPU-wise than Oracle’s unstated US$300. And never multiply a credit by a GPU rate until you have searched the vendor’s billing docs for the exclusion sentence; Daytona proves the two numbers can live one click apart and never touch.

For the structural question underneath this one, which category of platform fits your workload at all, read how to choose between a VPS, GPU cloud, serverless, or a sandbox. For the sandbox-only slice of this question, including Cloudflare Sandboxes and Blaxel, which have no credits entries yet, read AI sandboxes with GPUs. Every grant named here has a dated entry with official sources on the credits index, and head-to-head pairs live under /compare/.