Google vs. Nvidia: The Real War for AI

Google is building TPUs and not GPUs to run inside their own data centres. Let me explain the difference.

GPU:
A GPU is like a Swiss-army knife.
It can do many different things (graphics, games, AI, video, etc.).
Because it does so much, it’s complicated and expensive to build.

TPU:
A TPU is like a single-purpose power tool.
It does one job really well: multiply big sets of numbers for AI.
Because it only does that one job, it’s much simpler and cheaper.

Google’s business model:

Google owns important assets across the full stack of AI:
1. TPU chips.
2. Data from 3.5 billion humans using its services
3. Gemini models
4. Distribution across multiple surfaces like Search, YouTube and Cloud.

Nvidia’s business model:

Nvidia primary sells GPUs
But many of its biggest customers are unprofitable.
And its chips can be 80% of their cost base.
That’s hard to sustain when investor confidence wobbles.
Hence wants the AI narrative to continue so that its easier for these customers to raise money.

Where they collide?

Reading an MRI: $150 by a doctor.
$0.15 by an AI model.
That's a 99.9% cost collapse.
But who captures that value?
The chip maker? The model? The app? The end user?

Google wants lower chip costs. TPUs while primitive run perfectly in their datacentres. They don’t need the fancy add-ons normal GPUs have for most of it. Hence Google wants the value in AI to be captured not by the chips but other things like Gemini models, Google search etc.

Nvidia only has chips. Its wants its GPUs to capture most of the value in the AI value chain. It wants its customers to raise money so that they can afford the expensive GPU chips that Nvidia builds and that Nvidia can sustain its margins.

Google’s view: value should flow up the stack — to models, products, and distribution.

Nvidia’s view: value should stay in the chip, where it earns its margins.

That’s the real battle shaping the future of AI.