Free cloud compute tiers
Free cloud compute tiers
Cloud compute that never bills — for ML training without owning hardware. Quota numbers rot fast; every figure below was verified against the cited official page on 2026-07-17. Re-verify before relying on any of them. Complement of ML training on consumer hardware (the strictly-local path).
The tiers, as verified 2026-07-17
- Kaggle Notebooks — the only major tier with published, citable quotas (kaggle.com/docs/notebooks): ~30 h/week GPU (“or sometimes higher”), choice of 1×P100 or 2×T4 (4 cores, 29 GB RAM), TPU v3-8 at 20 h/week, 12 h sessions (9 h TPU), 20 GB persistent disk. GPU access requires phone verification. ⚠️ Kaggle’s own docs self-contradict on idle timeout: 20 min (notebooks doc) vs 60 min (efficient-gpu-usage doc) — plan for the stricter.
- Google Colab free — real but deliberately unpublished limits: the official FAQ commits only to “at most 12 hours” sessions and says limits vary over time. The widely-cited free T4 (16 GB), TPU v5e, ~15–30 h/week, ~90 min idle timeout are community-measured folklore, not contract.
- Lightning AI free — 15 credits/month (“up to 80 GPU h” marketing; ~22 T4-h by third-party math), single GPU, 4 h studio restarts, 50 GB storage, phone verification required.
- Modal starter — $30 free credits/month (≈50 T4-h), no card required; serverless functions rather than notebooks (steeper learning curve). Bills only if payment is explicitly added.
- AWS SageMaker Studio Lab — free T4 without an AWS account, but closed to new signups 2026-07-30; existing users unaffected, feature-frozen.
- HF Spaces ZeroGPU — 5 min/day, 60 s max per function: effectively inference-only, not a training path.
- Paperspace Gradient free — nominally alive under DigitalOcean but being folded into GPU Droplets; too churny to recommend.
The verdict is use-case-dependent
For fine-tuning, the free tier is comfortably over-provisioned (costed at measured M4 rates, 2026-07-17):
| Workload | Cost | Runs per 30-h week | Runs per 12-h session |
|---|---|---|---|
| Fine-tune Flowers-102 (full) | 4.8 min | 373 | 149 |
| Fine-tune Flowers-102 (partial) | 2.5 min | 714 | 286 |
| Fine-tune Food-101 (75,750 imgs) | 3.1 h | 9.8 | 3.9 |
| Pre-train ImageNet (600 ep) | 69 days | 0.018 | — |
A Flowers-102 fine-tune is 0.7% of one session; even Food-101 fits ~4× inside a single session with its 5 GB under the 20 GB persistent allowance. The session cap and ephemeral disk never engage — so for fine-tuning, neither checkpoint-chaining nor the ToS question below arises at all.
Where the free tier stops working
Verified 2026-07-17. For pre-training the quota is a hard ceiling, not a slow lane — the arithmetic simply does not close:
| Recipe | Epochs | Kaggle-weeks @30 GPU-h | 12-h sessions to chain |
|---|---|---|---|
torchvision mobilenet_v3_large | 600 | 55.2 | 138 |
timm mobilenetv3_large_100.ra4 | 3600 | 331.3 | 828 |
(Costed at measured M4 throughput; Kaggle’s P100/T4 is not an M4, so these are order-of-magnitude, not a Kaggle benchmark. A 2× faster GPU turns a year into six months — the conclusion survives.)
Checkpoint-and-resume across sessions: mechanically partial, practically no. Kaggle persists 20 GB via Save Version — enough for MobileNet checkpoints, not for the 144 GB ImageNet copy, which must be re-attached every session (ImageNet-scale training logistics). Colab deletes the VM entirely.
ToS, stated precisely (Colab FAQ): a single user hand-restarting a notebook and resuming from a checkpoint is not on the prohibited list. But Colab explicitly prohibits every mechanism that would make it practical at scale — “using multiple accounts to work around access or resource usage restrictions… employing techniques such as containerization to circumvent anti-abuse policies”, and for the free tier “remote control such as SSH shells… bypassing the notebook UI”. Combined with “Colab prioritizes users who are actively programming in a notebook”, the synthesis: a manual chain is permitted-but-deprioritized; an automated one is prohibited. 129 sessions is not something a human hand-restarts. Google’s own stated remedy is “purchasing a dedicated VM at GCP Marketplace” — i.e. stop using the free tier.
⚠️ Kaggle’s ToS is unverified: kaggle.com/terms is client-rendered and could not be read by any fetch method tried. No claim is made here about Kaggle’s policy on session chaining — read it in a browser before relying on one. (Kaggle’s mechanics support checkpoint-resume better than Colab’s; that is an inference from documented behavior, not ToS text.)
⚠️ Colab’s disk quota is unknowable: Google “does not publish these limits.” The widely-repeated “~100 GB” is not sourceable; community reports scatter across 35–78 GB. Do not state a number.
Rules of thumb
- For small CNNs, one Kaggle week (30 T4-hours) is thousands of CIFAR-10 training runs — quota is never the constraint at that scale.
- For pre-training at ImageNet scale the free tier is not a path at all — not slow, not awkward: ~55–330 weeks of quota and 138–828 chained sessions with no persistent room for the 144 GB dataset.
- For fine-tuning it is over-provisioned — 373 Flowers-102 runs per week’s quota. This is the use case the free tier actually fits, and the recommendation flips accordingly.
- Prefer Kaggle when a run outgrows the laptop; Colab free as fallback.
- “Free tier” here means never-bills-without-opt-in; anything requiring a credit card on file was excluded.