In AI, defaults win

It may not matter who builds the best AI model.

For the past few years, the AI conversation has revolved around benchmarks, parameters, tokens, datasets and agent autonomy. Every few weeks a new model claims the top spot. A new leaderboard. A fancier agentic workflow. A new launch event.

While the world debated model performance, Google quietly built something far more powerful: an AI ecosystem moat.

The full stack

Think about the Google stack, from silicon to models to products to distribution to data to feedback loops.

• Custom TPUs, designed for AI workloads.
• Frontier models like Gemini.
• Real-world signals from billions of users.
• A global distribution engine built into everyday products.

And not just any products. Infrastructure products. The ones we rarely think about, yet reach for the moment we need something. YouTube. Google Maps. Chrome. Android. Chromebooks. Search.

Each one is a data engine, a distribution engine and a training loop at the same time. Together they make something incredibly difficult to replicate: a self-reinforcing AI flywheel.

Defaults win

The real strategic shift may still be ahead. If Apple integrates Gemini into Siri, Google's models could reach billions of devices almost overnight.

Not necessarily for being the best model. For being the default layer.

In technology, defaults win.

History is full of examples. Google did not dominate search simply because there were no alternatives. It won because search became synonymous with Google. We don't search for things. We "google" them.

Defaults are powerful precisely because nobody thinks about them. They decide which browser opens, which map loads, which assistant answers. Most users never change them. A small edge in quality rarely beats the habit of a few billion people.

Why benchmarks mislead

Benchmarks are useful for researchers. As a guide to strategy, they mislead in three ways.

First, they measure the model in isolation. Customers never experience a model in isolation. They experience a product: the interface, the speed, the integration with their calendar or their codebase, the trust they place in the brand behind it.

Second, leads are short-lived. A model that tops the charts this quarter is often matched within months, sometimes by an open-weight alternative that costs a fraction to run. A position built on being slightly better is a position that has to be re-won every few months.

Third, "good enough" arrives quickly for most tasks. Once a model can draft the email, summarise the meeting or answer the question well enough, the next ten percent of quality matters far less to most users than convenience and cost.

When the core technology commoditises, value migrates. It moves towards whoever controls the customer relationship, the data and the workflow. This has happened in personal computers, in mobile phones and in cloud computing. There is little reason to expect AI to be the exception.

Where the moats are forming

If models alone are not the moat, what is? Looking across the industry, four sources of durable advantage stand out.

• Distribution. Owning the operating system, the browser, the productivity suite or the messaging app where people already spend their day.
• Proprietary data and feedback. Not just training data, but the stream of real usage that shows where the product fails and how to fix it.
• Workflow ownership. Being embedded so deeply in how work gets done that switching means retraining a whole organisation.
• Cost to serve. Owning the chips, the data centres or the energy contracts that make each answer cheaper than a rival's.

Very few companies have all four. The ones that do are playing a different game from those that have one.

What this means for everyone else

This is not an argument that the race is over. Model quality still matters, and great products can still break a default. But for any company competing in AI, a few questions matter more than the next benchmark.

• Where does your distribution come from, and who controls it?
• What proprietary data do you see that nobody else does?
• How fast does real usage feed back into a better product?
• Which layer do you want to own, and which ones are you happy to rent?

For most companies, the honest answer is that they will not own the default layer. That is fine. The opportunity then is to own a narrower default: the go-to tool for a specific profession, industry or workflow, where depth of understanding beats breadth of reach. Vertical champions can be very valuable businesses, as long as they are clear about which giant's platform they sit on, and what happens if that giant moves into their space.

These are classic strategy questions. Where to play, and how to win. AI has not changed them. It has made them urgent.

A different game

The AI race might follow the same pattern as search, mobile and the browser before it. Not purely a battle of models. A battle of ecosystems, distribution and feedback loops.

On that front, Google may already be playing a different game. The rest of the industry has to decide whether to challenge the default, or to build something valuable alongside it.