Google’s August 5 leadership shakeup at DeepMind, which pushed Demis Hassabis out of day to day operations and sent longtime engineering lead Jeff Dean to launch a rival lab called Discovery Loop, looked to most observers like a retreat. Tim O’Reilly, writing in The Asimov Addendum on August 8, offers a different read. He argues Google may be trading pursuit of the single most capable model for control of the infrastructure that other companies’ AI applications run on. The distinction matters because it changes what “losing” the frontier race would actually cost Google.
The retreat narrative traces to Dylan Patel’s SemiAnalysis, which framed the departures as the outcome of a chronically cautious compute posture inside DeepMind. But SemiAnalysis’s own numbers point somewhere else. The firm estimated Gemini’s annualized revenue at $12 billion in the second quarter of 2026, against a forecast of more than $73 billion in third-party cloud AI revenue and $120 billion in TPU sales for Google Cloud by 2027’s close. Alphabet’s most recent quarterly cloud revenue, $24.8 billion, grew 82 percent year over year, ahead of AWS’s 37 percent and Azure’s 43 percent. Google Cloud’s total folds in Workspace and other applications, which makes the comparison imperfect, but the growth gap is wide enough to notice.
O’Reilly’s central move is historical. He points to George Westinghouse’s rivalry with Thomas Edison in the 1880s and 1890s. Edison built the first working incandescent bulb and opened the first power station to serve paying customers, but his direct-current system only worked close to the generator. Westinghouse backed Nikola Tesla’s alternating current instead, a technology capable of running long distances at high voltage before being stepped down for household use. That system powered the lighting at an 1893 exposition in Chicago and, three years later, was sending current from a Niagara Falls generating station to homes in Buffalo. Edison held the more famous invention. Westinghouse built the system that spread electricity everywhere.
O’Reilly’s framing draws on Jeff Ding’s research on how great powers rise: not by inventing a general-purpose technology first, but by spreading it through the whole economy faster than rivals do. Ding points to Britain’s economy-wide adoption of steam-powered machinery and America’s fast diffusion of electrification and mass manufacturing, both of which outpaced narrower technical leads other countries held at the time. Japan, in Ding’s account, led semiconductors and consumer electronics through the 1980s yet still lost the broader information revolution to a US economy that was slower to build the chips and faster to put computers to work everywhere.
That framework is testable against what Google has actually shipped. AI Insiders has reported Google Cloud landing a nine-figure compute commitment from an AI lab, and Google building a transactional agent directly into Maps for a user base above two billion. Neither move chases frontier capability. Both extend Google’s infrastructure and consumer surfaces into transactions that other companies’ AI performs, which is exactly the diffusion pattern O’Reilly describes.
The weak point in the thesis is Google’s own spending. Alphabet continues to commit tens of billions of dollars a quarter to frontier training runs and TPU capacity, and Sergey Brin, per the essay, remains personally engaged in pushing Gemini forward. A company that had genuinely ceded the frontier race would not still be funding the compute to contest it. The more accurate read may be a hedge: keep the frontier effort credible while building the distribution layer that pays out regardless of who wins the model race.
For operators, the practical test is not next quarter’s benchmark leaderboard. It is whether Google Cloud’s third-party AI revenue actually approaches the $73 billion mark SemiAnalysis projects for 2027. If it does, expect Google to keep selling compute to competitors, Anthropic included, rather than trying to out-train them, a pattern worth pricing into any multi-cloud AI infrastructure contract signed over the next two quarters.
This analysis is based on Tim O’Reilly’s essay “Google’s Westinghouse Bet,” published in The Asimov Addendum on August 8, 2026.
