They’re worth roughly 15–30 T4 GPU hours per week (dynamic) plus up to 12-hour sessions with about 12–13 GB RAM and 16 GB VRAM when a GPU is available. In practice, that’s enough for quantized 7B LLM inference, LoRA/QLoRA fine-tuning on smaller models, Stable Diffusion image generation, and Whisper transcription. The exact “value” changes with availability because GPU access is not guaranteed at peak times. If you treat it like a disposable GPU sandbox for experiments and demos, it’s one of the better free deals out there.

No. Colab’s free tier does not require a credit card or any payment method.

A Colab free-tier session can run up to about 12 hours, and GPU availability is governed by a dynamic weekly quota rather than a fixed published limit.

Yes. If you have Google credits you won’t use before they expire, you can list them on AI Credit Mart and sell them at up to 70% of face value. Companies regularly list surplus credits from startup programs and enterprise agreements.

AI Credit Mart has discounted Google credits available from companies with surplus allocations. Prices are typically 30-70% below retail.

When your free GPU quota is exhausted, Colab typically restricts you to CPU-only runtimes, and if a session ends or disconnects, the VM is reclaimed and local files are wiped.

Because the idle timeout is based on interaction with the browser tab (clicking, typing, scrolling), not on whether code is running. If you’re inactive for about 90 minutes, Colab disconnects and reclaims the VM. Plan for it. Save checkpoints to Google Drive, and don’t assume a long training job will finish unattended on free tier.

No. The VM disk is ephemeral, so you need to mount Google Drive or download files to persist them.

They’re worth 30 hours/week of NVIDIA GPU time plus about 20–30 hours/week of TPU v3-8 time, which is often enough for a full week of experiments or a few serious training runs. In terms of capability, you’re getting access to a P100 (16 GB) or dual T4s (32 GB total) for PyTorch/TensorFlow training, and a TPU v3-8 (128 GB HBM) for TensorFlow/JAX workloads. The real “value” comes from the environment too: pre-installed ML libraries, easy dataset attachment, and background execution for long runs. If you checkpoint well, you can push surprisingly far without paying anything.

No.

For Kaggle notebooks, GPU and TPU access runs on a rolling weekly window that resets over time rather than expiring on a single date. CPU notebooks don’t have a weekly cap, only a per-session limit.

Yes. If you have Google credits you won’t use before they expire, you can list them on AI Credit Mart and sell them at up to 70% of face value. Companies regularly list surplus credits from startup programs and enterprise agreements.

AI Credit Mart has discounted Google credits available from companies with surplus allocations. Prices are typically 30-70% below retail.

On Kaggle, you don’t get charged when your GPU/TPU quota runs out; you simply lose accelerator access until the rolling weekly quota window refreshes.

Kaggle offers NVIDIA Tesla P100 (16 GB), NVIDIA T4 x2 in beta (2 × 16 GB), and TPU v3-8 (128 GB HBM across 8 cores) as notebook accelerators once your account is phone-verified.

Use “Save Version” and choose “Save & Run All (Commit)” so the notebook runs in the background. Interactive sessions can prompt “Are you still there?” after inactivity and may terminate if you don’t confirm, which is brutal if you walked away for lunch. Commit runs still obey the session caps (9 hours for GPU/TPU and 12 hours for CPU), so save checkpoints to /kaggle/working/ and plan to resume in a new session if needed.

They’re worth up to $5,000 of Lambda Cloud usage on on-demand GPU instances. At posted on-demand rates, that’s roughly about 10,000 hours on a Quadro RTX 6000, about 6,000 hours on an A6000, around 3,800 hours on an A100 PCIe 40 GB, about 2,000 hours on an H100 PCIe, or roughly 945 hours on a B200 SXM6. In practice, your “real” value depends on whether the GPU you want is in stock and whether you can keep instances utilized. If you’re doing research training runs, it’s a meaningful budget.

Yes. Lambda Cloud requires a valid payment method even if you have grant credits.

Lambda doesn’t publish a fixed expiration window for the research grant credits. What they do state is that grant credits are applied as service credits and decrease at the end of each weekly billing cycle, so you should monitor usage weekly.

Yes. If you have Lambda Labs credits you won’t use before they expire, you can list them on AI Credit Mart and sell them at up to 70% of face value. Companies regularly list surplus credits from startup programs and enterprise agreements.

AI Credit Mart has discounted Lambda Labs credits available from companies with surplus allocations. Prices are typically 30-70% below retail.

If you don’t have credits left to cover usage, Lambda will bill your registered payment method for any ongoing resources.

No. Lambda states there is no general-purpose free trial or free tier; the research grant is the only path to free compute they describe.

You’ll create a Lambda Cloud account, complete phone verification, add a credit card, and upload an SSH key (OpenSSH, RFC4716, PKCS8, or PEM formats). From the dashboard you can launch an instance by picking a GPU type and region, and attaching persistent storage. Don’t skip the storage step if you care about checkpoints, because data on terminated instances is permanently lost unless it’s saved to persistent storage at /lambda/nfs/<filesystem>. Also keep in mind capacity is first-come, first-served, so you may have to check back for popular GPUs.

There’s no lump-sum dollar credit; the value is the $0/month Basic plan plus included allowances (1 GB storage and 10 GB/month egress) and the ability to use serverless endpoints and run up to 48 CPUs and 2 GPUs concurrently in one workspace.

No upfront payment is required; it’s usage-based billing, so you only pay if you consume paid resources.

The Basic plan is ongoing (no stated expiration), but the platform is in transition post-acquisition, so plan structures and pricing may change over time.

Yes. If you have Lepton AI credits you won’t use before they expire, you can list them on AI Credit Mart and sell them at up to 70% of face value. Companies regularly list surplus credits from startup programs and enterprise agreements.

AI Credit Mart has discounted Lepton AI credits available from companies with surplus allocations. Prices are typically 30-70% below retail.

On the Basic plan, there is no credit balance to expire; instead, free allowances reset by policy (like monthly egress) and anything beyond the included amounts is billed pay-as-you-go.

10 requests per minute.

Dedicated GPU compute is billed by the minute with no minimum commitment, and the Basic plan allows up to 2 GPUs concurrently. Listed options include NVIDIA A10, RTX A6000, H100 80GB, and A100 80GB (with A100/H100 able to scale to 1, 2, 4, or 8 GPUs on supported plans). Model availability and partner routing can affect what you can access at any given time, so you will want to confirm current options in the dashboard before building around a specific SKU. If you stay idle and don’t spin resources up, you pay nothing.

You get 15 Lightning credits per month, and each credit is worth about $1. On a T4, that works out to roughly 22 GPU hours per month; on faster GPUs like an A10G or L40S, the same credits buy fewer hours because the hourly credit burn is higher.

No.

They renew monthly, and unused credits do not roll over.

Yes. If you have Lightning AI credits you won’t use before they expire, you can list them on AI Credit Mart and sell them at up to 70% of face value. Companies regularly list surplus credits from startup programs and enterprise agreements.

AI Credit Mart has discounted Lightning AI credits available from companies with surplus allocations. Prices are typically 30-70% below retail.

Unused credits don’t carry forward to the next month, so they effectively disappear at renewal. Your Studio storage is still persistent across restarts and sessions, but you won’t have more GPU time until the next month’s credits arrive or you move to a paid plan. If you’re mid-project, switch to CPU for everything you can, then schedule GPU runs in focused chunks. Honestly, it’s the difference between getting a demo shipped and staring at an “insufficient credits” screen. If you need continuous uptime, the free plan’s 4-hour restarts also push you toward checkpointing.

The GPU you pick changes your credit burn rate: a T4 is about 0.68 credits/hour (around 22 hours for 15 credits), while an A10G is about 1.80 credits/hour (around 8 hours). L4 is close to T4, and L40S is around 2 credits/hour (estimated), which means roughly 7–8 hours for the month.

Yes, if you want the initial bonus of 7 free GPU hours, phone verification is required.

They’re worth $30 of Modal compute usage every month on the Starter plan. Based on Modal’s published GPU prices, that’s roughly 50 hours on a T4, around 15 hours on an L40S, about a dozen hours on an A100-class GPU, or roughly 5 hours on a B200 (give or take). For CPU workloads, the value stretches further because you pay per core-second and there are no idle charges. The best use is bursty work: batch inference, scheduled jobs, or short fine-tunes that can scale up and then drop back to zero.

No.

They reset every month, and unused credits do not roll over.

Yes. If you have Modal credits you won’t use before they expire, you can list them on AI Credit Mart and sell them at up to 70% of face value. Companies regularly list surplus credits from startup programs and enterprise agreements.

AI Credit Mart has discounted Modal credits available from companies with surplus allocations. Prices are typically 30-70% below retail.

Unused Starter credits don’t carry forward, and the monthly allocation resets.

Starter includes 10 GPU concurrency, up to 100 concurrent containers, 5 deployed cron jobs, and 8 deployed web endpoints. You also get 1-day log retention, up to 3 workspace seats, and only 3 rollback versions per deployment. Custom domains aren’t available on Starter.

Not unless you add billing. Modal notes there’s no billing surprise risk on Starter: if you exceed $30 without a payment method, workloads stop rather than incurring charges, and you must explicitly add billing to go past the free tier.

If you’re awarded the maximum package, it’s up to 30,000 H100 80GB GPU hours delivered as cloud compute time. The real “worth” depends on whether you can use the allocation inside the roughly 6-month availability window. In practice, that’s enough capacity to run multiple training and experimentation cycles for a lab, especially if you’re disciplined about shutting down idle resources (Saturn Cloud auto-shuts off after 1 hour of inactivity). Some focus areas also include optional physical hardware, which can be a big deal for ongoing work after the cloud window ends. NVIDIA also notes it claims no IP developed under the program, which is valuable in its own way.

No.

GPU hours are available for about 6 months after the award decision, and the program runs on quarterly submission and decision cycles.

Yes. If you have NVIDIA credits you won’t use before they expire, you can list them on AI Credit Mart and sell them at up to 70% of face value. Companies regularly list surplus credits from startup programs and enterprise agreements.

AI Credit Mart has discounted NVIDIA credits available from companies with surplus allocations. Prices are typically 30-70% below retail.

The program frames the GPU hours as time-bound: after the availability window (about 6 months post-decision), you should expect the grant allocation to end.

No. Students and postdocs can participate on a funded project, but a full-time faculty PI must submit the application.

NVIDIA delivers cloud GPU hours through Saturn Cloud. After your grant is approved, you receive login credentials via email, then you can use JupyterLab or connect over SSH with tools like PyCharm or VSCode. You create a Python server resource and select a GPU configuration, and you can also create Dask clusters for distributed work. One practical detail: resources auto-shut off after 1 hour of inactivity to conserve your hours.

They’re not a fixed dollar balance; the value is free access to shared notebook machines, most reliably the Free-GPU M4000 (8GB VRAM), plus 5GB persistent storage. In practice, you can run unlimited 6-hour notebook sessions with restarts, which is plenty for coursework, quick model experiments, and iterative prototyping. The “worth” depends on what you would otherwise pay for GPU time, but the big win is no weekly GPU hour cap. Treat any free Ampere GPU access as a bonus, not the baseline.

Yes. A valid credit card is required for account verification, even on the free tier.

For this program, the key limit is the 6-hour auto-shutdown per notebook session, and you can restart sessions without a weekly GPU-hour cap.

Yes. If you have DigitalOcean credits you won’t use before they expire, you can list them on AI Credit Mart and sell them at up to 70% of face value. Companies regularly list surplus credits from startup programs and enterprise agreements.

AI Credit Mart has discounted DigitalOcean credits available from companies with surplus allocations. Prices are typically 30-70% below retail.

Gradient’s free tier isn’t a spendable credit balance, so there’s nothing to “expire” like a coupon. What you will hit are usage limits: your notebook shuts down after 6 hours, and you may be unable to start a free GPU when capacity is tight. If you move to paid on-demand GPUs, that usage is billed hourly (for example, A100 instances are priced around $3/hour). So the practical outcome is: free access can be unavailable or time-limited, and paid usage is billed normally.

Some Ampere free instances exist but are frequently unavailable, and the A100 is not free (it requires a paid plan plus hourly charges).

Not really. Free tier notebooks are public, so you should not put API keys, credentials, or proprietary code in them.

They’re worth a random bonus between $5 and $500 in RunPod account credits, awarded after you spend your first $10. You can apply them to Pods, Serverless, and storage, which means the value depends on what you run (an RTX 4090 in Community Cloud is roughly $0.34–$0.39 per hour, while higher-end GPUs cost more). Practically, the credits are most valuable when you’re experimenting and can choose cheaper instances, Spot capacity, or shorter runs. And no, you can’t cash them out.

Yes (or a crypto deposit), because RunPod requires a payment method before you can deploy any GPU.

RunPod doesn’t state an expiration window for this bonus in the program details.

Yes. If you have RunPod credits you won’t use before they expire, you can list them on AI Credit Mart and sell them at up to 70% of face value. Companies regularly list surplus credits from startup programs and enterprise agreements.

AI Credit Mart has discounted RunPod credits available from companies with surplus allocations. Prices are typically 30-70% below retail.

The program details don’t describe a specific expiration behavior; what RunPod is explicit about is that bonus credits can’t be withdrawn and are only usable on RunPod services.

You have to spend it.

Yes, but it’s through referrals, not by repeating the new-user bonus. RunPod’s referral program awards both the referrer and the referred user a $5–$500 random bonus after the referred user spends $10. The referrer can also earn commission for six months: 3% on Pod spend and 5% on Serverless spend. If you’re already planning to use RunPod for a while, this can stack up into meaningful savings.

They aren’t dollar credits; you get about 4 GPU hours/day (NVIDIA T4 16 GB) plus about 8 CPU hours/day (T3.xlarge) and 15 GB persistent storage. In practice, that’s enough for repeated small training runs, short fine-tunes, and GPU inference experiments without paying anything. The real “value” comes from persistence: your conda envs and installed packages stick around, so you don’t burn time rebuilding your setup each session.

No.

Studio Lab’s compute limits reset each 24-hour period (4 GPU hours/day and 8 CPU hours/day), and storage persists while you have access to the service.

Yes. If you have AWS credits you won’t use before they expire, you can list them on AI Credit Mart and sell them at up to 70% of face value. Companies regularly list surplus credits from startup programs and enterprise agreements.

AI Credit Mart has discounted AWS credits available from companies with surplus allocations. Prices are typically 30-70% below retail.

When Studio Lab session time runs out, your running computations stop, but your files and installed packages remain in your persistent storage.

No. Only one runtime session can be active at a time, which means you have to choose CPU or GPU per session.

Studio Lab requires approval, and AWS says review can take up to about 5 business days (though many requests clear faster). If you’re in a hurry, a referral code from a workshop or hackathon can bypass the wait and grant instant access. Also check your inbox carefully: after approval, the registration link expires in 7 days, and if it expires you have to submit a new request. One more gotcha is SMS verification on first launch; AWS supports 240+ countries but has reported delivery issues in some regions, and VoIP numbers typically don’t work.