Kaggle, Kaggle TPU, Hugging Face ZeroGPU and Google Colab side by side: the real limits, the cells we ran, measured speeds, and which one for which job

The companion to the reel and the long-form video. Four big companies lend you GPU time for free, with no credit card, each on its own clock. PAGE 1 — all four with links: Google Colab (we got a Tesla T4 on a free account on October 3 and ran an 8B model at about 32 tokens a second), Kaggle GPU T4 ×2 (2 × 16 GB, 30 hours a week; the P100 was retired September 15, 2026), Kaggle TPU v5e-8 (128 GB, 20 hours a week) and Hugging Face ZeroGPU (half an RTX Pro 6000 Blackwell, 48 GB, 5 GPU-minutes a day). PAGE 2 — the limits side by side and which one to use for which job. PAGE 3 — the exact Colab and Kaggle cells we ran, and the speeds we measured. PAGE 4 — turning on the TPU, using and hosting on ZeroGPU, and what fits where. PAGE 5 — seven catches and what to do about each. PAGE 6 — the proof checklist. Limits read on the pages named, October 3, 2026. Independent guide from Hyperautomation Labs — not affiliated with Google, Kaggle, Hugging Face, NVIDIA or Ollama.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — three practical AI emails a week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.