Something consequential is unfolding at the intersection of scientific ambition and corporate strategy. Google DeepMind, once the standard-bearer of frontier AI research, is losing the architects of its greatest achievements — Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals among them — as the organization that housed them quietly reorients around a different kind of power: the power to sell infrastructure rather than to discover intelligence. The departure of these figures is less a rupture than a revelation, making visible a choice that was already made: that the economics of being th
Google's AI Exodus: DeepMind Loses Top Talent as GCP Wins Compute Wars
Compute is the lifeblood of frontier AI research, and Google chose margin over moonshot.
Mark
Why would Google deliberately sell its most advanced chips to the companies it's competing against in AI?
Mimi
Because the person running Google Cloud won the internal argument. Kurian sees TPUs as infrastructure to be monetized, not as a strategic asset to be hoarded. He's right about the money—$200 billion in external sales beats $12 billion in Gemini revenue.
Mark
But doesn't that guarantee Gemini loses the race?
Mimi
It does. You can't build a frontier model without frontier compute. Google chose margin over moonshot.
Mark
So why is Jeff Dean leaving now, specifically?
Mimi
Because he can see the writing on the wall. DeepMind has lost the compute war. Starting a new lab outside Google, raising money from outside investors, and buying Nvidia GPUs on Google Cloud is now more viable than staying inside and fighting for scraps.
Mark
Is there any scenario where DeepMind recovers?
Mimi
Theoretically, yes. New leadership, clean slate, maybe they acquire another research lab and get serious about compute allocation. But Google's culture is bureaucratic and risk-averse. They had an AI chatbot a year before ChatGPT and were too scared to release it. That culture doesn't change with a few leadership swaps.
Mark
So Google Cloud wins by losing the AI race?
Mimi
Exactly. GCP becomes the infrastructure layer everyone depends on. Google makes more money as the landlord than it ever would as the tenant.
O Pulso
- Jeff Dean, Google's most celebrated engineer, is leaving to found Discovery Loop alongside three other senior researchers, marking the most significant talent exodus in DeepMind's history.
- Gemini has fallen from the world's top-ranked model in late 2025 to 8th or 9th place today, with API token growth decelerating and a flagship release quietly canceled.
- Google is selling more than 20 percent of its upcoming TPU capacity directly to Anthropic — one of Gemini's chief rivals — on long-term contracts, starving its own research lab of the compute it needs to compete.
- Google Cloud CEO Thomas Kurian has won the internal battle over resource allocation, framing TPUs as general-purpose infrastructure rather than instruments of AGI pursuit.
- With over $150 billion in TPU backlog and projected third-party AI infrastructure revenue exceeding $190 billion by end of 2027, Google's financial logic is coherent — even as its frontier AI ambitions hollow out.
Something consequential is unfolding at the intersection of scientific ambition and corporate strategy. Google DeepMind, once the standard-bearer of frontier AI research, is losing the architects of its greatest achievements — Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals among them — as the organization that housed them quietly reorients around a different kind of power: the power to sell infrastructure rather than to discover intelligence. The departure of these figures is less a rupture than a revelation, making visible a choice that was already made: that the economics of being the world's compute platform outweigh the prestige, and perhaps the peril, of racing toward artificial general intelligence.
On August 5th, Google announced a sweeping restructuring of DeepMind that amounts to something more than reorganization. Demis Hassabis stepped back from day-to-day leadership. Jeff Dean — the engineer behind the TPU program and co-lead on Gemini — announced he is leaving Google entirely to found a new research lab called Discovery Loop, taking with him Google Fellows Sanjay Ghemawat and Quoc Le, as well as Gemini co-lead Oriol Vinyals. Koray Kavukcuoglu, the remaining Gemini co-lead, will now run the entire division.
These departures are not isolated. They follow the exits of Noam Shazeer, John Jumper, and other foundational researchers — a sustained attrition of the people who built modern large language models. The timing sharpens the sting: in November 2025, Gemini 3 Pro was arguably the world's best model, alarming enough that Sam Altman reportedly issued a code red at OpenAI. Since then, the gap has widened. Gemini 3.5 Flash underperformed commercially, the forthcoming 3.5 Pro is rumored to trail Anthropic's Opus 4.5, and Gemini now ranks 8th or 9th among frontier systems. Google has quietly shelved the 3.5 Pro release and is banking on Gemini 4, though the trajectory offers little encouragement.
The business data reinforces the picture. First-party API token growth decelerated from 60 percent in Q1 2026 to 38 percent in Q2, and direct API revenue growth has followed downward. But the deeper explanation is structural, not technical. Google has been selling enormous quantities of TPU capacity — more than 20 percent of shipments scheduled through late 2027 — directly to Anthropic on long-term contracts, with hundreds of thousands more committed to Anthropic and Meta. The compute that might have powered DeepMind's research is instead flowing to its competitors.
Google Cloud CEO Thomas Kurian appears to have prevailed in the internal contest over these resources. In his framing, TPUs are general-purpose infrastructure, and selling them to Citadel, the Department of Energy, or Anthropic is simply what a platform company does. The financial case is difficult to argue with: Google Cloud grew 82 percent last quarter, driven heavily by TPU system sales, and with over $150 billion in backlog, the division is projected to generate more than $190 billion in external AI infrastructure revenue by end of 2027 — dwarfing Gemini's estimated $12 billion in annual recurring revenue. Google's leadership has made its choice. The question that remains is not whether DeepMind can close the gap, but whether it will ever be given the means to try.
On Wednesday, August 5th, Google announced sweeping changes to DeepMind's leadership structure that signal something far larger than a routine reorganization. Demis Hassabis, the lab's co-founder and former chief executive, stepped back from day-to-day operations. Jeff Dean—widely regarded as Google's finest engineer, the architect of the TPU program, and co-lead on Gemini—is leaving the company entirely to start a new research lab called Discovery Loop. He's taking three other senior figures with him: Sanjay Ghemawat and Quoc Le, both Google Fellows, and Oriol Vinyals, another Gemini co-lead. Koray Kavukcuoglu, DeepMind's former chief technology officer and the remaining Gemini co-lead, is being elevated to lead the entire division.
These are not the moves of an organization confident in its direction. The departures represent the latest chapter in a longer story of attrition that has drained DeepMind of its most accomplished researchers. Noam Shazeer and John Jumper, among the lab's top reinforcement learning specialists, have already left. The pattern is unmistakable: the people who built the foundations of modern large language models are choosing to leave Google rather than stay.
The timing matters. In November 2025, Gemini 3 Pro was arguably the world's best model—good enough that Sam Altman issued a code red at OpenAI. Since then, the gap has widened dramatically. Gemini 3.5 Flash was a commercial disappointment. The forthcoming Gemini 3.5 Pro is rumored to perform at roughly the level of Anthropic's Opus 4.5, which is already behind competitors like OpenAI's latest offerings and the top-tier Chinese open-source models. Gemini 3.6 Flash, released as a bridge model, trails Muse Spark 1.2, Grok 4.5, and several other systems. By most counts, Gemini now ranks 8th or 9th among frontier models. Google has quietly canceled the Gemini 3.5 Pro release and is now pinning hopes on Gemini 4, but there is little reason to expect that trajectory to reverse.
The business metrics tell the same story. Gemini's first-party API token growth decelerated sharply in the second quarter of 2026. In the first quarter, token consumption grew 60 percent, from 10 billion to 16 billion tokens per minute. In the second quarter, growth slowed to 38 percent, reaching 22 billion tokens per minute. This deceleration has cascaded into declining revenue growth for Gemini's direct API business.
The core problem, according to those tracking the situation closely, is not technical incompetence but organizational conviction. Compute is the lifeblood of frontier AI research. Every lab pursuing artificial general intelligence is desperately acquiring as much processing power as possible. Google, however, made a different choice. Rather than hoard its TPU capacity for DeepMind's research, the company has been selling enormous quantities of chips to Gemini's fiercest competitors on long-term contracts. More than 20 percent of total TPU shipments scheduled from the third quarter of 2026 through the fourth quarter of 2027 are being sold directly to Anthropic. This excludes the hundreds of thousands of TPUs that Google Cloud already rents to Anthropic today, plus hundreds of thousands more committed over the next six quarters to Anthropic and Meta.
Thomas Kurian, Google Cloud's chief executive, appears to have won the internal political battle over compute allocation. Kurian is not, by his own account, focused on artificial general intelligence. In public remarks, he has argued that it is beneficial for TPUs to become general-purpose infrastructure serving customers like Citadel, the Department of Energy, and high-performance computing firms. When asked why Google Cloud was selling compute to Anthropic despite the direct competition with Gemini, he characterized it as the natural consequence of Google being a platform company. That philosophy has now prevailed.
The financial implications are substantial. Google's estimates suggest Gemini generated roughly $12 billion in annual recurring revenue in the second quarter of 2026. By the end of 2027, Google Cloud is projected to generate over $73 billion in third-party AI infrastructure revenue, plus another $120 billion from TPU system sales—$200 billion in external sales at margins in the high 30s, compared to Gemini's $12 billion today. The company's leadership has made a clear choice: the short-term financial benefits of monetizing compute capacity exceed the long-term value of frontier AI competitiveness. Google Cloud grew 82 percent last quarter, driven substantially by TPU system sales to external entities running datacenters for customers like Anthropic. With over $150 billion in TPU system backlog, this revenue stream is expected to drive Google Cloud's growth rate into the mid-100s in 2027, far exceeding Wall Street consensus. The question now is not whether DeepMind can catch up, but whether it will ever be allowed to try.
Citações Notáveis
Google is a platform company, and it is natural that we sell compute to customers like Anthropic despite competing with them in AI.— Thomas Kurian, Google Cloud CEO (paraphrased)
DeepMind's core issue has always been a fundamental lack of conviction about compute allocation.— SemiAnalysis analysis
Quer a matéria completa?
Leia o original em SemiAnalysis ↗
Technology