DeepMind's Bad Year Is Now Your Vendor Risk

Demis Hassabis, Google DeepMind's cofounder and longtime CEO, stepped back into a chair role last week and handed day-to-day control to chief technology officer Koray Kavukcuoglu, who reports straight to Sundar Pichai rather than holding a standalone DeepMind chief title ^[1][2]. Minutes later, according to Fortune, news broke that chief scientist Jeff Dean was leaving too ^[3]. The org chart is not the story; what sits under it is.

Gemini 3.5 Pro, unveiled with fanfare at I/O in May, has now blown through three release dates ^[4]. That is roughly three months of slippage on a flagship that was announced before it was ready ^[5]. Meanwhile the independent benchmarking firm Artificial Analysis told Fortune that Google's best shipped model, Gemini 3.6 Flash, currently trails Anthropic, OpenAI, xAI, Meta and at least one Chinese lab on raw intelligence ^[6]. That is at least five labs ahead of the shipped product ^[7], and a reversal from the brief stretch last year when Google topped the leaderboard ^[8].

The talent ledger is worse than the two headline exits. In a single week in June, Google lost Gemini co-lead Noam Shazeer to OpenAI and AlphaFold co-inventor John Jumper to Anthropic ^[9]; two more AlphaFold veterans, Jonas Adler and Alexander Pritzel, followed Jumper out ^[10]. Counting Hassabis's move and Dean's departure, that is six named senior figures who have left or changed seats ^[11]. Engineers told Fortune the causes are aggressive poaching by cash-flush rivals, frustration at falling behind on coding benchmarks, and people chasing pre-IPO equity before those rivals list ^[12]. The same Fortune newsletter that reported the exits also listed River AI raising $1.1 billion for personal AI tools and Thrive Holdings raising $2 billion ^[13][14]. The equity those engineers are chasing is not hypothetical.

Several engineers also said the reshuffle compounds a slow pull of power from London, where DeepMind was founded, to Mountain View, and that losing Hassabis means the firewall between DeepMind and Google is breaking down ^[15]. A Google spokesperson disputed that framing, saying London remains central and DeepMind keeps its research autonomy under the new structure ^[16].

Two caveats before anyone writes Google off. It still has the chips, the cloud and the distribution that no AI-native rival can match, and those structural advantages can offset a weaker showing at the model layer ^[17]. And this is the second time in four years that Google has had to prove it can move fast, the difference being that it now does so without the two people most associated with its scientific identity ^[18][1][3].

For operators, the practical read is narrow. A single-vendor frontier model dependency now carries a roadmap risk you cannot price: a slipped flagship is three missed commitments that were presumably in someone's planning deck, and a shipped model ranked behind five competitors is a quality gap you are passing to customers ^[4][6][7]. Anything you have promised on the assumption that Gemini 3.5 Pro ships should be tested against the assumption that it does not.

What to watch: whether the fourth date for Gemini 3.5 Pro holds, whether Kavukcuoglu's line into Pichai produces faster shipping or more London attrition, and whether Artificial Analysis moves Google back up its ranking on the next shipped model rather than the next announced one ^[2][4][6].