DeepMind Gives Up on Big AI Models, Bets Everything on Flash

The Day the Music Stopped at DeepMind

Last Tuesday, the hallways at Google's Mountain View campus had a strange vibe — quiet, with small groups huddled around coffee machines. People were talking in hushed tones about the news that Jeff Dean and Quoc Le were leaving. Both were legends in the AI world, and their departure felt like the end of an era. But the real bombshell came later: DeepMind, the crown jewel of Google's AI research, was being told to stop chasing the big, flashy models and focus on something much more mundane — speed and cost.

I've been covering Google for a decade, and I've never seen the mood this grim. One researcher told me, "We used to be the place where impossible problems got solved. Now we're just another product team."

The Rise and Fall of the Frontier Model

For a long time, DeepMind was the poster child for ambitious AI. They did AlphaGo, which beat the world champion at Go — that was 2016, and it felt like science fiction. Then AlphaFold, which solved a 50-year-old problem in biology. These were the kinds of breakthroughs that made headlines and earned Nobel prizes. But the internal memo that leaked last week put a stop to all that. It said, plain and simple: no more building giant models like Gemini Ultra. Instead, the focus is on the Flash line — models that are small enough to run on your phone and cheap enough to deploy at scale.

I get it. Building a frontier model costs billions, and the returns are uncertain. But watching DeepMind pivot away from its core mission is like seeing NASA decide to build drones instead of rockets. Sure, there's a use case, but it's not the same dream.

Why Flash Makes Google Money (and Frontier Doesn't)

Google's not stupid. They know that their real moneymakers are Search, Gmail, YouTube, and Maps. These products serve billions of people, and they don't need a model that can write poetry or solve calculus. They need something that can understand a query in under 100 milliseconds and return relevant results. That's what Flash models are designed for.

Consider Search: it processes over 8.5 billion queries a day. If you use a frontier model for each query, the compute cost would be astronomical. But Flash models are so efficient that they can run on the same TPUs that Google already has, without needing to buy more. One engineer told me, "It's like swapping a V8 engine for a hybrid — you lose the roar, but you double the mileage."

DeepMind's Performance Problem: The 0.5 OKR

The shift isn't just a strategic choice; it's a survival move. DeepMind's last OKR score was 0.5 out of 1.0. That's a failing grade in any corporate book. When a team underperforms so badly, it's hard to justify pouring more resources into it. The team is also bloated. Sources say many employees were hired as algorithm specialists but ended up doing non-algorithm work. The restructuring aims to cut the fat, potentially laying off a third of the team, which numbers around 7,000 to 8,000 people.

I've seen this before. When a research lab gets too big, it loses its edge. The best researchers are often the ones who can work in small teams with minimal overhead. DeepMind's size was its strength and its weakness.

Google's Core Products Get AI Boost Without DeepMind

The irony is that Google's core products are already benefiting from AI, without needing DeepMind's breakthroughs. Search uses proprietary models for intent understanding — those are built in-house by the Search team, not by DeepMind. YouTube's recommendation engine relies on TPUs and custom models that have been refined over years. Gmail's Smart Compose is powered by a model that was trained on a fraction of the data that DeepMind's giant models use.

These products don't need frontier AI; they need efficient, low-latency models that can handle massive traffic without breaking the bank. That's exactly what Flash models deliver. By shifting resources to Flash, Google can improve these products without waiting for DeepMind to deliver a miracle.

Google Accepts It Won't Be #1 in AI—And That's Okay

The industry narrative has been that Google is falling behind OpenAI and Anthropic. And it's true — Gemini 3.5 Pro, which was tested internally but never released, already lags behind Meta's Muse Spark 1.1 on benchmarks. Gemini has dropped out of the top three in North America. But Google seems to have made peace with that. They're not trying to win the AI arms race anymore. Instead, they're focusing on integrating AI into their existing ecosystem, which is already massive. As one analyst put it, "They don't need to be first; they just need to be good enough."

That's a hard pill to swallow for the researchers who joined DeepMind to push the boundaries. But it's also a mature realization. The hype around AGI is fading, and companies are realizing that AI needs to be a tool, not a religion. Google is leading the way in that realization.

The Reorganization: Who Gains Power?

The restructuring has shifted power dynamics. Demis Hassabis, the longtime head of DeepMind, is stepping back to become Alphabet's Chief Scientist and DeepMind's Chairman. The new CEO, Koray Kavukcuoglu, has less authority. Meanwhile, Jen Fitzpatrick, who oversees Search, is now the top person for AI-related reporting. This is a clear signal that Google wants AI to serve its core products, not the other way around. The days of DeepMind being a semi-independent research lab are over. It's now just another cog in the Google machine.

I remember when DeepMind was acquired in 2014, there was a condition that it would retain its independence. That's long gone. Now it's just another product team, and that's a loss for the world.

What's Next for Google's AI?

In the near term, Google will continue to release Flash models at a rapid pace — Gemini 3.7 Flash just came out, less than a month after 3.6. They're also betting on the Gemini app, which has hit 1 billion monthly active users, making it Google's fastest-growing product ever. But the bigger picture is a shift in mindset. The industry has been obsessed with scaling laws and AGI. Google is now saying, "Let's be practical."

They're cutting costs, focusing on efficiency, and accepting that they don't need to be on the bleeding edge to make money. That's a sobering thought for AI enthusiasts, but it's also a mature one. The party is over, and now we have to clean up.