Train your own AI for $0, $48 or $5.6 million: what each one gets you

People say training AI costs millions. At the very top, it does. But there are two cheaper ways in, and almost anyone can do them.

• Free. Take an AI that someone already made. Teach it with your own examples. This is called fine-tuning. Google lends you the chip that does the maths (a GPU) for free.
• About $48. Build a small AI from zero, so every part of it is yours. You rent eight big chips for about two hours.
• About $5.6 million. This is what DeepSeek says the last big training run of its top AI cost. That number leaves out all the tests and failed tries that came before it.

The two cheap ones are not ChatGPT. But the steps are the same at every price: get data, train, test. This guide shows what you get at each price, the exact steps for the two you can do yourself, how long they take and what they cannot do.

None of the notebooks, code or papers below is made by webridgeAI, and we have no deal with any of them. Every number comes from the page it links to, read on 18 October 2026. Free plans and rental prices change, so check the page before you plan around a number.

The three tiers at a glance

The first two columns are things you can do this week. The third is there so the word "millions" has a real number and a real meaning behind it.

Tier 1, free: fine-tune a small open model on a free GPU

What you get. A copy of an existing model, here Meta's Llama 3.2 3B Instruct, adjusted on examples you choose so that it answers in your format, tone or domain. You are not teaching it to think. It already can. You are teaching it how you want it to answer.

Why it is free. Google Colab's FAQ says Colab "is free of charge to use", and the free tier includes GPU time on a T4. Unsloth's notebooks load the model in 4-bit and train only a small set of extra weights (LoRA adapters) rather than the whole model, which is what lets a 3 billion parameter model fit on that GPU. The notebook's saved run peaked at 6.5 GB of the T4's 14.7 GB.

Free fine-tune: the steps

1. Open the notebook. Unsloth's notebook list has dozens. Start with the Llama 3.2 (1B and 3B) conversational notebook. Its first line says to press "Run all" on "a free Tesla T4 Google Colab instance".
2. Pick the GPU. In Colab: Runtime, Change runtime type, T4 GPU, Save.
3. Run it once as it is. Runtime, Run all. Out of the box it installs Unsloth, loads unsloth/Llama-3.2-3B-Instruct, and trains for 60 steps on the mlabonne/FineTome-100k conversation dataset. Watch the loss number fall in the training cell.
4. Talk to it. A later cell asks the fine-tuned model a question and prints the answer, so you can see the result in the same notebook.
5. Swap in your own data. This is the step that makes it yours. Replace the load_dataset(...) line with your own conversations: a few hundred examples of a question and the answer you wish you got, in the same "conversations" format the notebook uses.
6. Train for longer. The notebook says it runs 60 steps "to speed things up", and that you can set num_train_epochs = 1 for a full run instead. A full pass takes much longer than the demo; on free Colab that may mean one session, or more than one.
7. Save it. The notebook ends with cells to save the adapters, or to export a GGUF file (it recommends q4_k_m) that you can run on your own laptop in llama.cpp, Ollama or LM Studio. Change if False to if True on the cell you want.

Other models on the same free tier. The same page has free notebooks for Qwen3 4B and Gemma 3 4B, both of which say they run on a free T4. The Gemma 3 notebook's saved run took 17.81 minutes for 30 steps. Unsloth's page notes that its largest models "exceed Colab's free 15 GB VRAM tier" and need a paid plan.

Free fine-tune: what it cannot do

• It is still the base model. A 3B model fine-tuned on your examples is a 3B model. It will not reason like a frontier chatbot, and it can still state wrong things fluently.
• The free GPU is not promised. Colab's FAQ says resources "are not guaranteed and not unlimited", that limits change and are not published, and that a free session runs for 12 hours at most. Save your work to Google Drive as you go.
• The 7 minutes is a demo. It is 60 steps, a small slice of the dataset. A fine-tune that changes behaviour in a way you can measure usually needs your own data and more steps.
• The weights are not all yours. You add a small layer to Meta's model, under Meta's licence. Read the licence before you ship it in a product.

Tier 2, 48 dollars: train a small LLM from scratch

What you get. A language model where every weight came from your own training run, through every stage: tokenizer, pretraining, fine-tuning, evaluation and a chat interface. Andrej Karpathy's nanochat repository says a model of GPT-2's ability "cost approximately $43,000" to train in 2019 and can now be trained "for only $48 (~2 hours of 8XH100 GPU node)".

Where the $48 comes from. The README does the sum: about $3 per GPU per hour makes an eight-GPU machine about $24 an hour, so two hours is about $48. It adds that on a cheaper spot instance the total "can be closer to ~$15".

From-scratch run: the steps

1. Rent the machine. Start a machine with eight H100 GPUs from a cloud provider. The README names Lambda as the one Karpathy uses; any provider with 8x H100 works.

2. Get the code and install it.

   git clone https://github.com/karpathy/nanochat.git
   cd nanochat
   uv sync --extra gpu
   source .venv/bin/activate

3. Start the run inside screen, so it keeps going if your connection drops. The whole pipeline is one script:

   screen
   bash runs/speedrun.sh

4. Wait. The README says the run takes about 1.5 hours now.

5. Talk to your model.

   python -m scripts.chat_cli

6. Shut the machine down. The meter keeps running while it is on. This is the step that turns $48 into $480.

The free version. The repository also has runs/runcpu.sh, which "dramatically shrinks" the model so that it trains on a CPU or an Apple Silicon laptop in tens of minutes. The README is plain about it: "You will not get strong results in this way." It is still the full pipeline, and it costs nothing. Our longer nanochat learning map puts one lecture and one book's worth of code before this run, so that you know what each line does.

From-scratch run: what it cannot do

• It is not ChatGPT. It matches GPT-2, a model from 2019. Karpathy describes talking to it as "a bit like talking to a kindergartener".
• It makes things up. It will answer confidently and wrongly.
• The price is a sum, not a quote. It rests on one day's rental rate. A run you forget to stop costs more.
• Smaller GPUs need changes. The README says GPUs with less than 80 GB of memory need the batch size turned down, and a single GPU works but takes about eight times longer.

Tier 3, 5.6 million dollars: a frontier model's final run

What it is. The DeepSeek-V3 Technical Report puts a price on its own training. Table 1 lists 2,788,000 H800 GPU hours for the full training (2,664K for pretraining, 119K for extending the context, 5K for post-training) and, "assuming the rental price of H800 is $2 per GPU hour", a total of $5.576 million.

What that run built. A mixture-of-experts model with 671 billion parameters in total, 37 billion of them active for each token, pretrained on 14.8 trillion tokens on a cluster of 2,048 NVIDIA H800 GPUs.

What the number does not cover. The paper says so itself, right under the table: the costs "include only the official training of DeepSeek-V3, excluding the costs associated with prior research and ablation experiments on architectures, algorithms, or data." So the $5.6 million is:

• One run, not the project. Every smaller test run and failed experiment before it is left out.
• A rental price, not a bill. It multiplies GPU hours by $2 an hour. It is not a claim about what DeepSeek spent on buying its GPUs, on staff, or on data.
• Still millions. At the very top, training does cost millions. The point of the other two tiers is that the top is not the only place training happens.

You will not run this tier. It is here because the same three things happen at every price: you prepare data, you run a training loop, and you test what came out. The $0 notebook and the $48 run teach all three.

Which one to start with

• You want a model that answers your way, this week: tier 1. Swap in a few hundred of your own examples.
• You want to understand how a model is built: tier 2, ideally the free CPU version first, then the full map.
• You want to know what "millions" pays for: read Table 1 and the paragraph under it in the DeepSeek-V3 report. It is two minutes of reading and settles most arguments about training costs.

Sources

• Unsloth, notebook list, and the notebooks Llama 3.2 (1B and 3B) Conversational, Qwen3 (4B) Instruct and Gemma 3 (4B), including their saved outputs, read on 18 October 2026.
• Google, Colab FAQ, read on 18 October 2026.
• Andrej Karpathy, nanochat, the repository and its README, read on 18 October 2026.
• DeepSeek-AI, DeepSeek-V3 Technical Report, arXiv 2412.19437, the abstract and Table 1.