What AI does Google use?
Google uses Gemini, its own family of multimodal models, built in-house by Google DeepMind. Gemini ships in size tiers — including Nano for on-device work and larger models for complex tasks — and powers Search AI Overviews and AI Mode, the Gemini app, Workspace, Android, and Chrome. Google trains it on its own TPU chips.
Why — the first-principles explanation
The important thing to understand is that Google is vertically integrated in AI in a way almost no one else is. Most companies rent someone else's model or someone else's chips. Google owns the whole stack: it designs the chips (TPUs), runs the data centers, does the research (DeepMind), builds the models (Gemini), and owns the distribution surfaces (Search, Android, Chrome, YouTube, Workspace). That is why questions like "which AI does Google use?" have a simple answer — it uses its own.
Gemini was built multimodal from the start, meaning text, images, audio, and video were part of training rather than bolted on afterward. Google released it in size tiers, originally Ultra, Pro, and Nano, and has kept that shape through later generations. The tiers exist for a physical reason: where the model runs determines how big it can be. Nano is small enough to run on a phone with no network round-trip, which is why it powers on-device Pixel features like Recorder summarization and Gboard Smart Reply. Larger models live in data centers where memory and power are effectively unbounded, at the cost of latency and per-query expense. Every deployment decision is a trade between those.
The second thing worth knowing: Google does not use one model everywhere. Search alone routes different requests to different models by cost. A billion-query-a-day product cannot afford to run its largest model on every search — that would be economically absurd. So cheap, fast models handle the common cases and the expensive ones get reserved for hard queries. When Google announced Gemini in Search, it highlighted a 40% latency reduction, which tells you what they actually optimize for at that scale: speed and cost, not maximum capability.
Worth separating from all this: Google's older AI is still everywhere and is not Gemini. Ranking, spam filtering, translation, speech recognition, and YouTube recommendations run on specialized systems refined over two decades. When people ask what AI Google uses, they usually mean the chatbot. But most of the AI Google actually runs is invisible plumbing that predates Gemini and has nothing to do with chat.
An example that makes it click
Think of a restaurant chain that grows its own vegetables, raises its own cattle, trains its own chefs, and owns every location. When you ask "where do they get their beef?" the answer is boring: from themselves. That's Google in AI — it makes the chips, the models, and the places you use them.
And the size tiers work like knives in that kitchen. You don't bring the enormous cleaver to slice a lemon. A phone gets a small, fast model because it has to fit in a pocket and answer with no internet. The data center gets the giant one because it has unlimited counter space. Same kitchen, different knife for the job.
Key facts
- Gemini is Google's own model family, developed by Google DeepMind, and described by Google as its most capable and general model, built to be multimodal.
- Gemini was released in size tiers — originally Ultra (most capable, complex tasks), Pro (scaling across a wide range of tasks), and Nano (most efficient, on-device) — a structure retained in later generations.
- Gemini Nano runs on-device on Pixel phones, powering features such as Summarize in Recorder and Smart Reply in Gboard.
- Google reported a 40% latency reduction when testing Gemini in its Search generative experience.
- Google trains and serves Gemini on its own TPU (Tensor Processing Unit) hardware rather than relying solely on third-party chips.
- Google reported more than $150 billion in annual capital expenditure in 2025, reflecting the infrastructure cost of running AI at this scale.
▶ The 60-second explainer (script)
What AI does Google use? Its own. It's called Gemini, built in-house by Google DeepMind. And the reason that answer is so short is that Google is vertically integrated in a way almost nobody else is. Most companies rent someone else's model, running on someone else's chips. Google owns the entire stack — it designs the TPU chips, runs the data centers, does the research, builds the models, and owns where you use them: Search, Android, Chrome, Workspace, YouTube. Gemini was built multimodal from the start. Text, images, audio, video — all in training, not bolted on later. And it ships in size tiers: Nano, Pro, Ultra. Those tiers exist for a physical reason. Where the model runs decides how big it can be. Nano is small enough to run on your phone with no internet round-trip — that's what powers Pixel features like Recorder summaries and Gboard Smart Reply. The big models live in data centers with effectively unlimited memory and power, but you pay in latency and cost. Here's the part people miss. Google doesn't use one model everywhere. Search routes different questions to different models based on cost. At a billion queries a day, running your biggest model on every search is economically insane. When Google put Gemini in Search, the number it bragged about was a forty percent latency cut. That tells you what they actually optimize for at that scale: speed and cost. And one more thing — most of the AI Google runs isn't Gemini at all. Ranking, spam filtering, translation, YouTube recommendations. Twenty years of invisible plumbing that has nothing to do with chat.
What authoritative sources say
People also ask
Does Google use ChatGPT or OpenAI models?
No. Google builds and uses its own Gemini models. OpenAI is a direct competitor.
Is Bard the same as Gemini?
Bard was the earlier product name. It was rebranded to Gemini, and the assistant runs on Gemini models. The name Bard is retired.
What chips does Google train on?
Primarily its own TPUs — Tensor Processing Units it designs in-house. This is a major structural difference from competitors dependent on purchasing third-party GPUs.
Does Google use the same model for every search?
No. Different queries are routed to different models based on difficulty and cost. Running the largest model on every one of a billion daily queries would be economically impractical.
Is all of Google's AI Gemini?
Not remotely. Search ranking, spam filtering, translation, speech recognition, and recommendations run on specialized systems developed over two decades that have nothing to do with Gemini.