The scarce resource is the people who design the architecture

On June 18, Noam Shazeer announced he's leaving Google for OpenAI to work on next-generation model architectures. If the name doesn't land, the work will: he's a co-author of the 2017 paper "Attention Is All You Need," which introduced the Transformer — the design under ChatGPT, Gemini, Claude, and nearly every model you've used. Google had paid roughly $2.7 billion to bring him and his team back less than two years ago. He left anyway. And in the same three days, Google also lost a Nobel laureate to Anthropic.

When companies pay billions for individual people and still can't keep them, that tells you what's actually scarce.

Compute is buyable. The design isn't.

The AI race looks like it's about money and chips — and a lot of it is. But chips you can order. Data you can gather. The thing you can't simply buy is the small number of people who can look at the current approach and invent the next one. Transformers didn't come from more GPUs. They came from a handful of people rethinking how a model pays attention.

That's why a single architect can move markets: the people who design the shape of the system are worth more than almost any other input, because the shape decides what everything else can do.

This is true at every scale, not just frontier labs

You're probably not designing a new attention mechanism. But the same law governs your work: the highest-value decisions are architectural, and they're made by people, not bought off a shelf.

I've felt this directly. Building a real-time system for ten thousand players in one world, the win wasn't a faster server — it was naming the real bottleneck (traffic, not CPU) and designing around it. Get the architecture right and the rest gets easy. Get it wrong and no amount of compute or clever code saves you.

What it means for hiring and for you

• The leverage is in the design, so protect time for it. Teams that treat architecture as a quick step before "real work" optimize the cheap part and skip the expensive one.
• A great architect is worth a crowd of typists — especially now, when agents make the typing cheap. The scarce skill isn't producing code; it's deciding what should be built and how it should be shaped.
• Invest in judgment, not just output. The person who picks the right structure saves you the rewrite a year later. That's the role the labs are paying billions to fill.

The bottom line

A billion-dollar tug-of-war over one researcher isn't a celebrity story. It's the market pricing the rarest input in AI.

Compute and data are buyable; the people who design the architecture are not — which is exactly why the design is where the value lives, at every scale. The labs just put a very large number on it. Treat your own architecture decisions like they matter that much, because they do.