Google's cloud backlog hit $514 billion at the close of Q2, up 4.75X year on year — and Google is rewriting its Gemini leadership to chase it. The Mountain View giant disclosed this week that Jeff Dean, employee number 30 and co-leader of the Gemini effort, is leaving after 27 years to run a new AI startup called Discovery Loop.

Dean is taking three senior researchers with him: Sanjay Ghemawat, a Google Senior Fellow with two decades on hardware infrastructure; Quoc Le, co-founder of Google Brain; and Oriol Vinyals, DeepMind's deep learning lead who co-led Gemini with Dean and who originated knowledge distillation and sequence-to-sequence learning.

The reshuffle lands as Google's flagship Gemini 3.5 Pro, announced in May and originally slated to ship in June, remains undelivered with no public date. Anthropic's Claude and OpenAI's GPT models have pulled ahead in the public conversation, pressuring a team that invented the transformer architecture underpinning all generative AI.

"We have to accelerate all this work and stay focused on the AI frontier," CEO Sundar Pichai wrote in a blog post announcing the changes. "At the same time, there's never been a more important moment to shape the future of AGI and science."

What does the new structure look like?

Demis Hassabis, co-founder of the DeepMind unit, steps up from DeepMind CEO to chairman and takes the chief scientist role at Google that Dean held. "Demis has described us as standing in the foothills of the singularity, and has been spending a lot of his time engaging externally," Pichai wrote. "He and I have been long discussing a role that allows him to put his full attention on actively shaping the future of AGI."

Koray Kavukcuoglu, formerly DeepMind's CTO, now carries the title chief AI architect and runs Gemini directly. Amin Vahdat stays as SVP overseeing infrastructure, including the TPU accelerators and TensorFlow stack Dean originally built.

The split between Google Brain and DeepMind, kept distinct since the acquisition, now appears to be resolving in DeepMind's favor. Brain's founders are leaving; DeepMind's veterans run the models.

What does Wall Street see?

Google is reportedly preparing a $25 billion bond offering on Wall Street this week to bankroll the AI infrastructure push. CFO Anat Ashkenazi told analysts last week that Google is pulling planned capacity forward into 2026, lifting the full-year capex range from $180–$190 billion to $195–$205 billion. The increase reflects timing shifts, not unit price hikes, Ashkenazi said.

In Q2 alone, Google spent $44.92 billion on capex — every dollar routed through Google Cloud. Servers took $26.95 billion (60 percent); datacenter shells and networking ate the remaining $17.97 billion (40 percent). Both line items ran almost exactly 2X year-ago levels.

How is the cloud unit actually doing?

Google Cloud posted $24.77 billion in Q2 revenue, up 81.8 percent year on year and 23.7 percent sequentially. Operating income tripled to $8.81 billion, a 35.6 percent margin — a record for a unit that long dragged profits. Cloud now delivers 20.7 percent of Alphabet's revenue and 21.6 percent of operating income.

Alphabet reported $119.8 billion in total Q2 revenue (up 24.2 percent) and $40.77 billion in operating income (up 30.4 percent). Reported net income hit $112.11 billion, but roughly $77.1 billion of that came from markups on equity holdings — a $94 billion SpaceX stake chief among them. Google rents capacity on SpaceX's "Colossus" supercomputer at $920 million a month.

Does the reshuffle fix Gemini's lag?

Not by itself. Kavukcuoglu inherits a model line that has slipped against Claude and GPT even as Google pours capital into the silicon and shells behind it. Pichai's message reads as a verdict: the old structure slowed them down, and the people who built it are gone.

The real test lands the moment Gemini 3.5 Pro finally ships — and whether a DeepMind-led team can close the gap before OpenAI and Anthropic push another frontier generation out the door.