Sundar Pichai and Google’s Improbable AI Comeback

The awkward part of Google’s AI story is that the underlying architecture behind the current boom was invented inside the company. The transformer paper was published by Google researchers in 2017. For several years afterward, the most visible commercial applications of that work were built elsewhere. I have known a man to invent a very good mousetrap and then stand in his own doorway watching the neighbours sell them from a cart.

That gap produced a period of unusually public criticism. Competitors shipped assistants first. Regulators asked why a company with the best research was slowest to market. Internal memos leaked describing the situation as a missed opportunity. When a company’s private doubts begin arriving in the newspapers before the board has finished reading them, you may safely assume the matter has gone beyond a difference of opinion.

The reorganisation

Pichai’s response was structural rather than rhetorical. Research teams were merged, a single model family was made the centre of the product line, and the company’s enormous distribution — search, Android, Workspace, Chrome — was pointed at the problem. Notice what he did not do. He did not make a speech.

The advantage Google had all along was never model quality. It was the fact that it already had the users. Embedding an assistant into products that a billion people already use daily is a shorter path to adoption than convincing anyone to install something new, and convincing anyone to install something new is the longest errand in the trade.

The cost of the comeback

• Capital expenditure at a scale that has unsettled investors accustomed to software margins.
• A search-advertising business model that an assistant interface potentially undermines.
• Sustained regulatory attention on both sides of the Atlantic.

The strategic bind

The deepest problem is not competitive but internal. Google’s revenue depends on people clicking links. An assistant that answers directly reduces those clicks. Pichai has spent years managing this tension, and the resolution — new ad formats inside AI-generated answers — is one of the most closely watched experiments in the industry, watched chiefly by people who would like it to fail.

It is an unenviable position: the company must disrupt its own most profitable product before someone else does. Pichai’s tenure will be judged on whether it manages that transition while the transition is still optional.

What the distribution actually buys

The company’s advantage is not that its assistants are better. It is that they are already installed. An improvement shipped to a billion devices overnight is a different kind of event from an improvement announced at a conference and adopted by volunteers, and the difference is not a matter of taste but of arithmetic.

That advantage is also a constraint. Changes must work across every device, language and accessibility requirement at once, which slows releases and makes the organisation conservative about anything that might behave unpredictably at scale.

The research record

It is worth remembering that the underlying architecture behind the current boom came out of the company’s research organisation, as did several of the techniques now standard in training and serving large models. The gap was in commercialisation, not in science. Remembering it costs nothing and explains a good deal.

That is a more uncomfortable diagnosis than a capability gap, because it points at organisational structure and incentives rather than at the quality of the researchers. Those are harder to fix and slower to show results.