Become an AI Expertv2.1
0 of 0 steps Continue
Reference

Compute and costs

You do not need a GPU laptop. fast.ai's official line: "You don't need any special hardware or software." Classical ML runs on any CPU. Prioritise RAM over GPU if buying a laptop. Rent a GPU by the hour only when a specific project needs one.

Prices below were checked in the first week of November 2026 and move monthly. Treat them as a planning range and re-check at signup.

Free tier, enough for the first 6 to 12 months

Service What you get Notes
Google Colab (free) T4-class GPU, up to 12h sessions Limits are unpublished and fluctuate. Fine for Zero to Hero, Raschka's book, smol-course.
Kaggle Notebooks About 30 GPU-hours a week (dynamic), T4x2 or P100, 12h sessions No credit card. Check your account page for the current quota.
Modal (Starter) $30 a month in free credits, no base fee Per-second billing. Enough for several fine-tuning runs a month.
Lightning AI (Free) One-time credits worth about 50 interruptible GPU-hours, plus a free CPU workspace Changed in September 2026 from monthly to one-time. Older guides are wrong.

Rule: Colab free plus Kaggle plus Modal's $30 covers everything in Path A modules 1 to 4 except the nanochat run. Pay only when a project needs more than a T4 for more than about 12 hours.

Provider T4 L4 RTX 4090 A100 H100 8xH100 node
Modal $0.59 $0.80 $2.50 (80GB) $3.95 ~$31.60
RunPod Community $0.44 $0.34 $1.19 (PCIe) $1.99 PCIe / $2.69 SXM ~$21.50
Vast.ai marketplace ~$0.36 ~$1.73 to $2.14 varies; interruptible 50%+ cheaper
Lambda (from 16 Nov 2026) $1.99 (40GB) $4.29 $31.92
Lightning AI $0.55 $0.79 $2.19 (40GB) $5.45
Colab Pro $9.99/month for 100 compute units (a T4 burns about 1.76 units an hour)

What the course's projects cost

Project Hardware Cost
Zero to Hero, Raschka book 1, smol-course, ARENA chapters Free Colab / Kaggle $0
Tabular project, RAG project, agent project Laptop CPU plus API calls $5 to $30 in API usage
nanochat GPT-2-grade speedrun 8xH100, about 2 to 3 hours $48 on demand per the README; about $15 on spot; Karpathy quoted about $73 for a 3h run in Jan 2026
nanochat larger tiers (optional) 8xH100, 12 to 33 hours $300 to $800
SFT and GRPO on your nanochat model 1 to 8 GPUs, a few hours $10 to $50
LoRA fine-tune of a 7B to 8B model One A100 or RTX 4090, 2 to 4 hours $3 to $10
GRPO on a 1.5B to 4B model with LoRA One 48GB GPU, several hours $10 to $40
KV-cache and kernel benchmarks One GPU, hours $5 to $20

A realistic whole-course budget is $100 to $300 of GPU time plus $50 to $150 of API calls over the course and the first year after it. The Redditor who trained three LLMs from scratch spent about $750; that is the upper end.

Spot and interruptible instances

Spot pricing cuts cost by half or more but the machine can be taken away. Checkpoint every 10 to 15 minutes and write a resume script before you start. nanochat supports resuming.

Use other people's money

  • Ask your employer for a learning budget tied to a pilot project (median learning spend per employee fell to about $299 in 2026, so ask early and be specific).
  • Hugging Face, Modal and most clouds run periodic credit programs for open-source and student work.
  • MATS pays about $2,000 a week in compute to fellows.

Sources

Free to use: code MIT, writing CC BY 4.0, third-party media under its owners’ terms. Licence and terms.