K-Culture GlossaryGWhere everyone starts
GPU
The chip used to build and run AI. The reason NVIDIA became the world's most valuable company.
In plain words
GPUs were originally built to render game graphics. The core of graphics work is doing millions of simple calculations at the same time — and it turns out AI training is exactly the same kind of task. So a chip made for gaming became the key resource of the AI era.
If a CPU is "one professor solving hard problems one at a time," a GPU is "thousands of students solving easy problems all at once." AI training overwhelmingly favors the latter approach.
NVIDIA has effectively cornered this market. That's why it became the world's most valuable company after the AI boom took off, and why news about "the race to secure GPUs," "tens of thousands of H100s," and "export controls" never stops. In AI news, you can read "GPU" as shorthand for "the capacity to build AI."
How it shows up in the news
"Company X breaks ground on a data center with 100,000 GPUs" — this reads as a declaration that the company is building the scale needed to train large models on its own.
GPU vs. NPU vs. TPU: What's the Difference?
Is it a general-purpose processor?, Is it a low-power dedicated chip inside a device?, Is it a chip Google built for its own use?
AI chip articles often mention all three in the same paragraph, making it hard to tell who's actually competing with whom. They're not fighting over the same turf — one dominates the data center, one sits inside your phone or laptop, and one was built by Google to run its own services.
| Aspect | GPU | Neural Processing Unit (NPU) | Tensor Processing Unit (TPU) |
|---|---|---|---|
| Where it lives | Data centers. Cloud servers | Inside your phone or laptop | Google data centers |
| Who makes it | NVIDIA, AMD | Device makers like Apple, Qualcomm, Intel | Google (manufactured with Broadcom's help) |
| Who uses it | Nearly everyone building and selling models | Anyone using a device — usually without realizing it | Google and Google Cloud customers |
| What it's good at | The heaviest training and large-scale serving | Running small models while saving power | Large-scale training and serving, cheaply |
| In the news | "GPU shortage race," "NVIDIA earnings" | The reality behind "AI phone" and "AI PC" | "Challenging NVIDIA's dominance" |
| Can you buy one? | Yes, though supply falls short of demand | It's already built into any device you buy | No — you rent it through Google Cloud |
Rule of thumbIf a model is built in a data center, it's a GPU. If it runs inside your own device, it's an NPU. If it's behind a Google service, it's a TPU.