Despite the enormous progress that has been made in generative AI frontier models in the past four years, attaining artificial general intelligence or superintelligence is not an easy goal. Very few of us are behind the scenes at AI model builders and fewer still understand the intricacies of trying to push models further and further toward human and superhuman capabilities.
I strongly suspect that the egos of the people involved are understandably strong, that people are driven, and that like people, models do not always behave as they are expected to. For each of the model builders, the push towards AGI has caused delays in models being released because the models do not behave appropriately, or have bad emergent behaviors that are difficult to guardrail.
Given all of this, it is not really a surprise that many of the top AI executives at Google have just jumped ship amid what seems to be a constant game of musical chairs across top model researchers for the past two years. Moreover, Google’s top-end Gemini model is falling behind Anthropic’s Claude and OpenAI’s GPT GenAI models – even though Google invented the transformer model that underpins all GenAI models – with its flagship Gemini 3.5 Pro model, which was announced in May and expected to ship in June still not available and no delivery date been given.
“We have to accelerate all this work and stay focused on the AI frontier,” Google chief executive officer Sundar Pichai explained in a blog post announcing executive changes across its various AI units. “At the same time, there’s never been a more important moment to shape the future of AGI and science.”
Significantly, Jeff Dean, who was employee number 30 at Google and who steered the development of many of the core search engine, data analytics, and hyperscale infrastructure – including MapReduce, Bigtable, Spanner – and who founded the Google Brain research group at the company in 2011 and brought TPU accelerators and the TensorFlow AI framework into being. In recent years, Dean has been co-leader of the Gemini effort. After 27 years at Google, Dean is leaving to be head of an AI startup called Discovery Loop that will try to automate the workflow of scientific research using AI, and he is taking three key AI researchers with him: Sanjay Ghemawat, who is a Google Senior Fellow and who worked on Google hardware infrastructure for more than two decades; Quoc Le: a co-founder of Google Brain; and Oriol Vinyals, who has led the deep learning research at DeepMind, another deep learning effort that Google acquired in the early days of deep learning and who created knowledge distillation (having a small fast model trained by a big slow model) and sequence to sequence learning (one of the foundations of the transformer). Vinyals was the co-lead alongside Dean on the Gemini models.
Google is making an investment in Discovery Loop and everybody is talking nice, but clearly Google is building a new Gemini team because it is not happy with what it is getting out of the current team.
But, it is also possible that in addition to this, after all of these years, these Google supergeniuses might just want to do something not at hyperscale and that is unequivocally and demonstrably good for the world. Speeding up research to help scientists discover stuff faster and engineers to build stuff faster is perhaps more benign than creating AGI, which will augment some people and but likely replace a much larger number of people in the long run.
Now, Demis Hassabis, who was one of the co-founders of the DeepMind acquisition and who was the chief executive officer of that DeepMind division, is moving up to chairman of DeepMind and taking over the chief scientist job at Google from Dean.
“Demis has described us as standing in the foothills of the singularity, and has been spending a lot of his time engaging externally,” Pichai explained in his post. “He and I have been long discussing a role that allows him to put his full attention on actively shaping the future of AGI.”
It is important to note that the longtime hardware and software infrastructure guy (Dean) is being replaced by one of top AI minds in the world (Hassabis). Amin Vahdat, who is chief technologist for AI infrastructure and who masterminded many networking innovations at the company, is a senior vice president firmly in charge of Google infrastructure and is well acquainted with what Google has done and what it needs to do on that front.
And finally, Koray Kavukcuoglu is now chief AI architect at Google and is in charge of the Gemini models. He was previously a researcher at DeepMind and rose through the ranks to be its chief technology officer and obviously has worked very closely with Hassabis over their careers.
Here’s the thing. DeepMind has been kept distinct from the Google Brain research arm for reasons Google has not really explained over the years. Having two different units doing what seems like the same thing may not be a good thing. GenAI and the race towards AGI means putting research into production as soon as possible. Google needs to have one AI group and get focused. Competing teams is sometimes useful in driving innovation, but it leads to fiefdoms and infighting. From where I sit, it looks like DeepMind won and Google Brain left the building.
Meanwhile, the AI buildout continues, and this week Google is reportedly hoping to do a $25 billion bond offering on Wall Street to support its ambitious AI infrastructure investments this year and into the future. In its financial report from last week, Google’s chief financial officer Anat Ashkenazi said that it was boosting its capital expenditures for 2026 from somewhere between $180 billion to $190 billion to somewhere between $195 billion to $205 billion, and due mostly to pulled planned capacity from the future into 2026 – not from price increases in infrastructure. (The price increases for 2026 were baked into the capex spending plan increased that were updated in the 2026 spending projections from the first quarter.)
None of this means, by the way, that Google is not making money on AI. It certainly ism and AI is certainly driving the breadth and depth at its Google Cloud group, which is the part of Google we care about most.
In the second quarter, Google Cloud revenue was up 81.8 percent to $24.77 billion, and that was a 23.7 percent sequential growth from Q1 2026 by the way. Operating income for the cloud group rose by a factor of 3.1X to $8.81 billion, which was 35.6 percent of revenues – a new high for Google Cloud, which has been a drag on profits for years but is now bolstering the company. By the way, Google Cloud represented 20.7 percent of Google’s overall revenue, a new high, and 21.6 percent of its operating income, also a new high.
The health of the overall Google businesses matters to Google Cloud, since the profits from other groups provides the company with the cash to invest in the cloud in general and AI in particular as well as the ability to borrow from the capital markets to not burn down its cash hoard. Overall, Google’s revenues – really Alphabet’s revenues, but like many of you, we still think of Google as Google and not as Alphabet with Google underneath it – were up 24.2 percent to $119.8 billion, operating income was up 30.4 percent to $40.77 billion, and thanks to the increasing value of various investments that have to be put on the books, net income was up by just under 4X to $112.11 billion.
With the big cloud builders and IT suppliers who have invested in AI model builders, net income is starting to become largely paper gains, which makes it had to reckon how the underlying company really did. The underlying net income for all of Google was probably on the order of $35 billion, with $77.1 of the net income coming from such investments. A lot of this funny money came from a $94 billion holding in SpaceX as well as some contribution from its stake in Anthropic. Interestingly, SpaceX stock was initially up as it went public last month because Google was renting capacity on the SpaceX “Colossus” supercomputer for $920 million a month.
In the quarter, Google spent $44.92 billion on capital expenses and all of it was for Google Cloud. Once again, about 60 percent of that figure ($26.95 billion) was spent on servers, and 40 percent ($17.97 billion) was spent on datacenters and networking. (It is funny how the network is part of the datacenter enclosure and not part of the server racks the way Google thinks of it. A datacenter is just a big server to Google, and has been for decades.) That spending on servers and on datacenters and networking was almost precisely 2X what it was a year ago.
That capex level is driven entirely by the Google Cloud backlog, which was at $514 billion as the second quarter came to a close, up 4.75X year on year and up 11.3 percent sequentially.