Is Google (AI) in Trouble?

My Medium friends can read this over on Medium.

When did you last Google something? I mean type it into the box, click a blue link, read the page.

I actually want the answers because it's super interesting. Many people in my circle don't do it anymore. They use ChatGPT. Or similar.

I still google, though. Often enough. But a lot of the time I take the AI summary at the top and never scroll further. And for many questions I don't open Google at all anymore. I ask Claude.

Probably many of you reading this too.

And that's not even the whole story. This summer, three of the biggest names in AI research walked out of Google within two months. Gemini 3.5 Pro was promised for June and is still not here. The stock took two big hits.

So, is Google in trouble? Let's look at it.

Exits

For a long time, if you wanted to work on AI, you went to Google. The 2017 paper that everything in this field is built on, Attention Is All You Need, came out of Google. The Transformer architecture, the T in GPT. Google's paper.

Then they missed the product moment. Bard against ChatGPT. Remember that? I worked for Google at that time. And Bard really wasn't great.

Then they caught up. Gemini got pretty good. NotebookLM got popular. AI Overviews landed on every search result page on the planet. For a while it looked like Google was back in front. I think...

And now this.

On June 18, Noam Shazeer left for OpenAI. He co-wrote that 2017 paper. He co-led Gemini. Google had paid around $2.7 billion in 2024 to get him back from his own startup, Character.AI. Less than two years later he's gone.

One day later, John Jumper left for Anthropic. Nobel Prize in Chemistry for AlphaFold, the protein-folding model. Nine years at DeepMind. That's huge.

Alphabet stock dropped about 7% that week. Roughly $250 billion. Billion with a B.

Then, in August, the big one. Jeff Dean left after 27 years. 27 years! Google's chief scientist. The guy who worked on the search index, on MapReduce, on TensorFlow, on Gemini. He took Sanjay Ghemawat, Oriol Vinyals, and Quoc Le with him, and founded Discovery Loop. Google is an investor in it, which softens the blow a bit. But still.

The same week, Demis Hassabis stepped back from running DeepMind day to day. He's chairman now, plus chief scientist for Alphabet. Koray Kavukcuoglu runs it now.

Four moves in seven weeks. At least… interesting...

Why do they leave

I can only guess, of course. But two things stand out.

Google carries a lot of legacy. And I mean huge legacy. Search, Gmail, YouTube, Android, Workspace. Billions of users, billions in ad revenue, and everything you build has to fit in somewhere. You can't just try things. Well you can. Google does all the time. But it has huge implications. There is a lot to lose. Big companies with a cash cow have always had that problem. Monopolies get comfortable. I wrote about Google's monopoly being in danger last year, and the pattern is the same.

The second interesting fact is what Discovery Loop and Jeff Dean say they want to do. Not build the next chatbot. That's boring. Instead, build AI that automates research itself. Models that improve models, experiments that run thousands at a time, the loop closing on its own.

That's a different bet than "make Gemini answer emails better." And apparently the people who want to make that bet don't feel they can make it inside Google right now.

Again, my guess.

The missing model

Gemini 3.5 Pro was announced at I/O in May. Coming in June, they said.

It missed June. It missed mid-July. It missed early August. Bloomberg reported the coding performance falls short, especially next to what OpenAI and Anthropic ship. Google retrained on new data in late June. Didn't help much, apparently.

On the earnings call, Sundar Pichai skipped the question and talked about Gemini 4 instead.

Now, a company not releasing a model can mean many things. Maybe it's not good enough. Maybe it's too expensive to run. Maybe it's fine and they don't want the launch to land in a quiet news week. We don't know.

But "coming in June" and then three months of silence is not a great look when your top researchers are leaving at the same time.

And it's a lot different from Apple. We know they don't do AI models well. But they have other leverage. What about Google?

Which Gemini?

When people say "I use ChatGPT" or "I use Claude", you know what they mean. One app, one thing.

When people say "I use Gemini"… which one? The Gemini app. Gemini in Search, now called AI Mode. AI Overviews. Gemini in Gmail. Gemini in Docs. NotebookLM, now Gemini Notebook or whatever it is....

And the list goes on: Antigravity, the coding tool. Spark, the agent. What? WHAT?

I exaggerate a bit. But not much.

That's the legacy problem again. Google can't start on a blank slate like OpenAI did. Every AI feature has to live inside a product that already exists, and there are dozens of them.

And in the products themselves, the picture is mixed. I have Claude Code as my default for building things. Antigravity was back on my Mac since the new version, and it works better, but it still is an IDE with agents inside. You don't run your workday from it. NotebookLM is popular for a reason: it takes something complicated, chatting with your own documents, and makes it simple.

Spark is… I don't know. A background agent that uses your Gmail, Calendar, Drive, YouTube history, and does things for you while your laptop is closed. Since July 30 it's available on the $20 Pro plan in more than 160 countries.

Not in the EU. Not in the UK. Not in Switzerland.

So I can't tell you if it's any good. I wrote about Spark as Google's answer to Claude Cowork, and my take hasn't changed: it exists, most of us over here can't use it.

Google has the data to make an agent like that great. Whether the model is good enough at using tools, and whether the context is clean enough, is another question. Everyone building agents right now knows that part is the hard part.

Money

It's always money.

Google's business is not in trouble. Not even close. Q2 2026: $119.8 billion in revenue, up 24%. Search alone $63.3 billion, up 17%. Cloud $24.8 billion, up 82%. The Gemini app has 950 million monthly users. AI Mode in Search passed a billion.

A billion people a month use Google's AI answers. And the ad business grew anyway. Crazy...

The stock still lost around $300 billion in value after those numbers. Because Alphabet raised its 2026 spending on data centers and chips to $195 to $205 billion. Per year. With more coming in 2027.

So here's a bet on data and chips. Is it a strategic mistake or does management know something we don't? My gut says my own AI use went up a lot in the last year, and since agents started to work properly it went up again. And I'm probably part of a small group of heavy, heavy users right now. If the rest of the world catches up, that hardware won't sit idle.

Then the chatbot numbers. market share this year, at 46.4% in May according to Sensor Tower. Gemini is at 27.7%. Claude at 10.3%. Other counts, web traffic instead of app usage, put ChatGPT higher and Claude higher. Depends on who's counting.

Either way, Google went from "Bard, LOL" to almost a third of the chatbot market in two years. Losing is something else.

For me, as a writer, it has other implications as well. Google Search is my biggest traffic source. SEO beats every social feed I have ever tried. And every question that gets answered by an AI summary, or inside a chatbot, is a click that doesn't happen. Google's own AI Overviews eat Google's own clicks. And mine.

Downsides, of course

Talent churn is not a Google-only problem. Meta, xAI, OpenAI, Anthropic. Researchers move like soccer players in transfer season. Somebody should introduce a transfer window.

The $200 billion could be wrong. If AI demand grows slower than the data centers do, that's a lot of expensive, empty racks.

Gemini 3.5 Pro might land next week and be great. Could prove me wrong. We'll see.

And the EU... We get every Google agent feature months late, or never. That's partly regulation, partly Google's choice. Either way, from Germany, the exciting Google stuff is something I read about, not something I use. Unfortunately.

What Google has, and nobody else does: the whole stack. Search, Gmail, YouTube at the top. Workspace for companies. Google Cloud underneath. Their own TPU chips at the bottom. OpenAI and Anthropic rent their compute. Google builds it. They're not in trouble.

The Bottom Line

Yeah, Google is not in trouble (yet). The money says so. The stack says so.

But the people leaving say something too. And the model that isn't there.

I use Google every day. Search, mail, AI. I'd like the products to move as fast as the numbers do.

Or is that already the trouble?

If you're not already, follow me on Medium.

Get Substack SEO