(Read time: 6 minutes)

Crypto and AI are a power duo—but to really see the magic, you need to understand the engine underneath.

That’s where GPUs, training, and model types come in. Today, we’re breaking it all down. Let’s go!

Key Concept #1: What are GPUs?

Call of Duty (COD) and FIFA. 

Two of the best games I’ve ever played (pls don’t judge me!)

If you’ve ever been blown away by the hyper-realistic explosions in COD or hit that iconic “Siuuu” celebration in FIFA, you can thank one thing: your GPU.

Short for Graphics Processing Unit, a GPU is the powerhouse behind all those insane visuals. Originally built for gaming, GPUs crunch thousands of complex calculations simultaneously

And GPUs aren’t just for gaming anymore. They’re now the backbone of AI, processing massive datasets at lightning speed.

Unlike CPUs, which work through tasks one at a time, GPUs can handle multiple operations at once. That parallel processing makes them perfect for training AI models, which need to sift through mountains of information to “learn.”

In short, GPUs are like supercharged calculators.

Key Concept #2: Training vs Inference

Training: The Learning Phase

Training is when an AI model “learns” how to recognize images, generate text, or do whatever it’s designed for. 

It’s kinda like AI’s awkward high school years. Absorbing knowledge, making mistakes, and gradually figuring things out.

This process is insanely expensive. AI models consume mountains of data and require a ridiculous amount of computing power, powered by fleets of GPUs.

Training frontier models like GPT-4 costs upwards of $100M in hardware—enough to buy a private island and staff it with AI researchers.

And the GPU arms race is real. Meta announced plans to acquire 350,000 H100 GPUs, leading to a potential total expenditure of around $12-$14 billion for these GPUs alone.

Meanwhile, Elon Musk is assembling a 200,000-strong H100 GPU cluster, called “Colossus”—set to be the world’s largest supercomputer.

Inference: The Doing Phase

Once trained, an AI model enters its "adult" phase – inference.

This is when it takes everything it has learned and starts applying it in real-time, making predictions or generating responses.

It’s like how you just know your native language without consciously thinking about grammar rules.

Every time you type a prompt into ChatGPT, the model instantly processes it and delivers an answer—that’s inference.

Unlike training, inference doesn’t require nearly as much computing power, which is why we can use AI models in real-time.

Think of it like learning to play the piano:

• Training = Practice: Repeating scales, learning songs, making mistakes—slow, painful, but necessary.

• Inference = Performance: Once skilled, you can read new music and play songs smoothly, impressing everyone (or at least your mom).

Key Concept #3: Open-source vs Closed-source

In AI, models come in two main flavours: closed-source and open-source.

Closed-Source Models

Closed-source models, like OpenAI’s GPT-4 or Anthropic’s Claude, are like secret recipes owned by big companies.

You can drink Coca-Cola, but you can't make the exact same one at home because its recipe is not publicly available.

With closed-source AI, you get access through ChatGPT, Claude, or APIs, but you can’t tinker with the model itself. These are powerful, well-funded models, but they’re black boxes—we have to trust the companies’ claims about how they work and what their limitations are.

Plus, because they run only on the company’s servers, you can’t host them yourself, customize them, or tweak them for specific needs. 

Right now, closed models are generally more advanced, which is why most AI-powered apps rely on them.

Open-Source Models

Open-source models are like a legendary family recipe that’s shared with the world.

The “weights” (parameters that make the model tick) are released publicly, meaning anyone can use, modify, or improve them—no permission needed.

But here’s the catch: Once a model goes open-source, the creator can’t directly monetize it anymore. Instead, the real value comes from community-driven improvements and adoption.

A popular example of an open-source language model is Llama 3, developed by Meta (thanks Zuck!). Llama has become a favourite for developers, who adapt and fine-tune it for specific tasks.

But it’s not just Llama getting love—DeepSeek, a new contender, recently dropped and immediately grabbed attention across the world. Worth a spin (you can try it here).

The AI game is playing out a lot like iPhone vs. Android—one’s locked-down but polished, the other’s open but endlessly customizable.

Most businesses don’t need an AI that writes Shakespearean sonnets while solving quantum physics problems—they just need practical AI for tasks like emails, customer service, and data analysis.

Open-source is already crushing those use cases.

On the other hand, closed-source models are backed by billion-dollar war chests; they’ll keep chasing the next AI breakthrough. Meanwhile, open-source models will continue getting better, cheaper, and more adaptable.

Bottom line? We’re headed for a hybrid AI future—where both approaches push each other forward, giving users the best of both worlds.

Further Reading

If you’re interested in exploring open-source models, check out Hugging Face—it’s the world’s largest library of open-source models, with over a million options to experiment with

What’s Next?

You just learned three of the most important building blocks in AI—congrats! 👏

In the next lesson, we’ll demystify how AI actually learns—no PhD required. Think of it as peeking under the hood of the machine.

Cheers,

Teng Yan