Why Big AI Labs Are Quietly Cooking Cheaper Models

The Kitchen Gets Quieter

Last week, the mood in Google's Mountain View headquarters was thick with farewells. Employees lined up for one-on-ones with Jeff Dean and Quoc Le, two names about to leave for a new startup. Many of those meetings came from DeepMind, and the vibe was tense. People worried about their jobs, their projects, and what comes next.

Dean's new company, Discovery Loop, is basically doing what DeepMind does. Three other co-founders are also senior Google folks. So those meetings weren't just goodbyes—they were chances to hand over a resume, maybe get a referral, or schedule an interview before the door closes.

From Grand Ambitions to Flash

Here's what we've learned from sources: DeepMind is no longer chasing the frontier of AI models. Instead, it's focusing on Flash—lighter, cheaper models that don't need a supercomputer to run. And with this pivot, DeepMind might cut its workforce by a third or more.

The goal is to trim people who were hired for algorithm work but aren't actually doing it. DeepMind has around 7,000 to 8,000 employees. Nothing is final yet, but the writing is on the wall.

Why the Pivot?

Training a massive model like Gemini 3.5 Pro is expensive and slow. It's also not winning. Internal benchmarks show it trailing competitors like Meta's Muse Spark. Gemini has dropped out of the top three in North America. Grok is doing better. Honestly, Google seems done with the frontier race.

That's not a disaster. Google's core products—Search, Gmail, YouTube, Maps—don't need a trillion-parameter model. They need something smart, fast, and cheap enough to run billions of times a day. That's what Flash is for. It's the difference between a Michelin-star tasting menu and a reliable daily special. One is flashy; the other feeds the crowd.

The Cost of Cooking

Compute is the secret ingredient, and Google's TPU clusters are the ovens. Search uses custom models to understand queries; YouTube relies on TPUs for recommendations and ads; Google Photos needs them for tagging and search. These services serve billions of people, and their AI needs are massive.

DeepMind, on the other hand, has been struggling to get the resources it wants. The team's OKR score last cycle was 0.5 out of 1. That's not a team you hand more money. So the message is clear: stop chasing the giant models, focus on what works.

Not a Retreat, a Rethink

Google isn't giving up on AI research. It's still a pioneer in architectures, machine learning, and hardware. But the era of unlimited spending on frontier models is over. The company is reorienting around using AI to improve existing products, not just to win benchmarks.

Leadership changes reflect this. Demis Hassabis steps back from day-to-day DeepMind management to become Chief Scientist. Koray Kavukcuoglu takes over, but with less power. Jen Fitzpatrick, who runs Search, now has more say over AI. That's a big deal—she's a product person, not a research person.

What This Means for You

So what does this mean for the rest of us? For one, AI features might get cheaper and faster. Google can roll out Flash-powered tools that don't drain your battery or your wallet. For another, the pace of big model releases might slow down. Instead of a new flagship every month, we'll see incremental improvements in smaller models.

This is a shift from the AGI hype of the past few years. Companies are realizing that scale isn't everything. Sometimes, the best dish is the one you can make every day, not the one that wins awards once a year.

As for DeepMind employees? Some will move to other teams, some will leave. The uncertainty is real, but so is the opportunity. In the kitchen of AI, the menu is changing. And Google is betting that a good, cheap meal beats an expensive one that only a few can taste.