In February, Gemini 3.1 Pro sat at the top of the Artificial Analysis Intelligence Index. It beat Claude Opus 4.6 by four points and cost less than half as much to run. For a few weeks, Google led on every chart the AI crowd likes to screenshot.
At I/O in May, Google promised Gemini 3.5 Pro for June. It didn’t arrive. In mid-July, Bloomberg reported that the model’s coding was short of where Google wanted it, and that an attempt to fix this by changing the training data had disappointed. Google’s statement talked about shipping many models quickly and keeping them “highly cost-effective for customers.”
Fair enough on the first part. Since July, Google has shipped Gemini 3.7 Flash, 3.8 Flash, speech models, a video model, a music model and a robotics model. Its API changelog, latest entry dated 22 September, still has no 3.5 Pro. In his Q2 earnings remarks, Sundar Pichai was already talking about Gemini 4 and “our most ambitious pre-training run yet.”
Yesterday I wrote on LinkedIn that I think Google’s chances of leading in AI are slim. Coming from me, that’s a change of position. I’ve bet on Google for years, and I always expected it to win this race. When the AI hype started, I was the cautious one, happy to let the novelty settle before jumping in.
I’m still cautious. I just don’t want to be blind to what’s in front of me. After 20 years in search, what I see in Google’s recent moves is a company turning into Microsoft: rich, everywhere, and slow to act against the business that pays for everything.
The replies to my post were good, and most of them disagreed. Almost all made the same case: Google owns the data and the distribution, makes its own chips, and has more cash than anyone it competes with.
All true. Microsoft had the cash and the distribution too.
I’m not predicting Google’s fall. Google will be fine. My argument is narrower: I don’t think Google has the motivation or the nerve to build the thing most likely to eat its search business, because the search business still works. The question was always whether Google is willing to break something that prints money.
Everything on the inventory is real
Let’s give the other side its full due. Google crawls the web, runs Chrome and Android, owns YouTube and sells Workspace to businesses. Its cloud backlog stood at $514 billion at the end of Q2. The Gemini app passed a billion monthly users in August, and Pichai says AI Mode has passed a billion too.
The strongest point against me comes from Comscore. On its US panel, Gemini’s share of AI assistant prompts went from 17% in January to 30% in June, while ChatGPT fell from 70% to 50%. That’s more than Gemini appearing in front of people. People are typing into it.
I accept all of it. An inventory tells you what a company can do. What it will choose to do is a separate question, and for that you watch where the scarce stuff goes.
Follow the chips
The scarce stuff in AI is compute. Google makes its own, and it rents it out. In May, The Information reported that Anthropic had committed to spend $200 billion over five years on Google’s cloud and chips. In Q2, Google Cloud revenue grew 82% to $24.8 billion, with operating income of $8.8 billion.
Now put that next to a Reuters exclusive from August. Tight computing capacity, and disputes between the people running different parts of Gemini, left areas such as coding short of resources. Cloud’s senior leaders, Reuters reported, hoped the new Gemini leadership would calm a fight with DeepMind over who gets Google’s limited chips. In April, The Information reported on a memo in which Sergey Brin told DeepMind staff that Google must “urgently bridge the gap” with Claude on agentic coding.
I don’t know how those allocation meetings went. I know how the quarter ended: Cloud up 82%, and the in-house model short of its coding targets.
As a business decision, it’s hard to argue with. A TPU rented to Anthropic is revenue this quarter. A TPU spent on another Gemini training run is a bet. That’s exactly my point. A company that must win takes the bet. A company already winning on the P&L takes the contract.
Search got an AI bodyguard
The same pattern shows on Google’s most valuable surface. When Google put generative AI into Search, it put it inside the results page: AI Overviews on top, AI Mode beside it, and ads moving into both. I covered what that means for publishers in Half Your Traffic Left: your content, consumed on Google’s surface, next to Google’s ads.
From Google’s side, it’s working. Search and other revenue grew 17% in Q2, to $63.3 billion, and Pichai credits the AI features with driving query growth. A defence that pays that well is the most comfortable place a company can be. Comfort is the problem.
What I think threatens Search is the assistant that does the job and never shows a results page at all. It writes the code, compares the contracts, books the trip, and nobody clicks an ad on the way. That’s the category where Google is furthest behind, and I don’t think that’s a coincidence. Building it properly means building the thing that makes Search optional. Every incentive inside Google points the other way.
Microsoft already ran this experiment
If distribution were enough, Microsoft would be winning. It had Windows, Edge and Office as default placement. It moved first, putting OpenAI’s technology into Bing in February 2023, while Google was still in its code red over ChatGPT.
Three and a half years later, Bing holds 4.5% of worldwide search, according to StatCounter’s August figures. Copilot, sitting right there in Windows, sends 2.8% of the AI chatbot referral traffic StatCounter tracks. StatCounter counts referral clicks rather than usage, so treat the Copilot number as rough. The direction isn’t in doubt.
Microsoft is fine, of course. It rents cloud compute to the companies building models, OpenAI among them. Sound familiar? Microsoft did very well selling shovels, and nobody calls it the leader in search or consumer AI. It failed to stay competitive on exactly the fronts that demanded it disrupt itself.
So what about Gemini’s consumer gains? They’re real, and they’re happening where Google’s distribution is strongest. Where people pick a model for paid work, the picture changes. Menlo Ventures’ 2025 enterprise report put Anthropic at 40% of enterprise LLM API spend, OpenAI at 27% and Google at 21%. In coding, it had Anthropic at 54% and OpenAI at 21%, with Google at 11%. Menlo is one of Anthropic’s biggest backers and co-runs a fund with it, so read those numbers with that in mind. Even with that discount applied, the money teams spend on getting work done has mostly gone elsewhere.
Google moves when something is on fire
To be fair, Google can move fast when it’s frightened. ChatGPT triggered a code red inside Google in December 2022. Two months later, Bard’s launch demo got a fact wrong and Alphabet shed about $100 billion in market value in a day. The recovery that followed was real. Gemini 2.5 Pro reached the top of LMArena in March 2025, and after Gemini 3, it was Sam Altman calling a code red at OpenAI.
So yes, Google can sprint. It needs a fright first. You could read the 3.5 Pro delay as Google being careful after Bard, and I’d accept that. Holding back a model that misses your own bar is discipline.
It just tells you nothing about whether Google will build the assistant that makes the results page optional. Nobody declares a code red when Search revenue is up 17%. Without a fire, Google moves at Microsoft speed.
Moving like Microsoft
Twenty years in search teach you to watch how a company moves, more than what it announces. Google’s announcements are fine. Its movements this year: a Pro model promised for June that still hasn’t shipped at the end of September, a coding gap its own co-founder flagged in April and a leadership reshuffle that only came in August, and a CEO talking up Gemini 4 before 3.5 Pro is out. To me, that’s the pace of a company protecting a lead. It’s the pace Microsoft had with Bing.
We’ve seen this before. Google had Gmail, Android, YouTube and most people’s contacts when it launched Google+ in 2011. It shut the consumer version in 2019. Google had every advantage in social except a need to win it. Go back further and you get Xerox PARC, which built the graphical interface and watched other companies ship it. Google researchers published the transformer paper in 2017. OpenAI turned it into ChatGPT.
The people seem to read the room the same way. Noam Shazeer, co-lead of Gemini, left for OpenAI in June. In August, Jeff Dean left after 27 years to co-found Discovery Loop, Demis Hassabis stepped back to chair DeepMind, and Koray Kavukcuoglu took direct control of Gemini. Fortune, using Zeki Data, reports that DeepMind’s hires per departure fell from roughly 12 in 2023 to roughly two now. Anthropic managed 22 in 2025. My read: researchers who want to build what comes next go where what comes next is the whole business.
What would change my mind
Two signals. The first is a Gemini product that does the job end to end, with no results page and no ad slot, pushed by Google as hard as it pushes AI Mode. The second is Google letting that product take queries away from Search, and saying so on an earnings call as if it were good news. If I see either, I’ll write the follow-up and eat my words.
Until then, Google will do fine. So did Microsoft. Ask Bing.