Model

DeepMind is losing researchers over Google's Israel and Pentagon contracts, including $1.2 billion Nimbus

Google DeepMind, Alphabet's London AI lab, is losing researchers over its Pentagon and Israeli military contracts. The revenue is a rounding error next to Alphabet's compute spending; the departures land on the people its AGI bet rests on.

Richard TangRichard Tang · 3 min read
Share
Demis Hassabis, chief executive of Google DeepMind, speaking at EPFL in Lausanne in May 2023.
1 / 6Slide 1 of 6
Demis Hassabis, chief executive of Google DeepMind, whose AGI roadmap rests on the researchers now resigning over the lab's military contracts. Photo: Alain Herzog / Mediathèque EPFL, CC BY-SA 4.0.

Google DeepMind, Alphabet's−2.50% — Alphabet, down 2.50 percent today artificial intelligence research lab in London, is trading researchers for a defense book, its portfolio of military contracts, and the two sides of that trade are not the same size. Sofia Liguori, a London-based research engineer, resigned this week over Google's military work for the United States and Israel; the AI researcher Andreas Kirsch left earlier this month; the BBC counts several more gone or leaving 12. The ethics fight is well covered. For Alphabet, Google's parent company, the money and the talent now point in opposite directions.

The contracts are small next to the compute bet

Run in order, the defense book built fast. Project Nimbus, the cloud-and-AI deal Google shares with Amazon−4.32% — Amazon, down 4.32 percent today for Israel's government and military, was selected in 2021 and is worth about $1.2 billion 32. The GenAI.mil contract with the Pentagon materialized in December 2025, while Google's Cloud unit ramped up to IL6, the security level required to handle classified workloads 1. In April, Google announced a classified agreement letting the military use its models for "any lawful government purpose"; the Pentagon confirmed the signing in early May 124.

On that scale the military money is small: Nimbus, over its whole life, is worth less than one week of Alphabet's capital spending, and capital spending is the bet the resignations land on.

Alphabet's quarterly capex has more than tripled in two years

$10B$20B$30B$40B$50BQ3 '24Q4 '24Q1 '25Q2 '25Q3 '25Q4 '25Q1 '26Q2 '26Nimbus, over its whole life, isless than one week of this
Data
Capex
Q3 '24$13.06B
Q4 '24$14.28B
Q1 '25$17.2B
Q2 '25$22.45B
Q3 '25$23.95B
Q4 '25$27.85B
Q1 '26$35.67B
Q2 '26$44.92B
Quarterly capital expenditure in USD billions, from Alphabet's SEC filings. The defense contracts are priced against this scale.9

The cost lands on the research bench

The departures started before this week. Alex Turner, a DeepMind research scientist who worked on AI safety, resigned in June with no other job lined up, after roughly 600 staff signed a petition against classified work 4. DeepMind's employees voted to unionize in May 1.

Liguori says her internal posts containing the word "genocide" were deleted, and her verdict now is that "we should have gone on strike" 1. Kirsch believes Google "seems to have no right to control or veto how the Pentagon uses its AI" 2. Google says it reviews such contracts through a consistent process, including its AI principles, and gives staff safe channels to raise concerns 2.

Maven was the precedent, and the hedge is new juniors

There is precedent, and it has an endpoint. Google withdrew from the Pentagon's Maven drone program in 2018 after mass protests and resignations, and in 2024 it fired roughly 50 workers over Nimbus office protests 5. Management's hedge now, in one current employee's words, is "a changing of the guard": a junior pipeline fills the seats 2.

Watch the departures, not the contract revenue

The resignations land on a bet Alphabet has priced in public. Sundar Pichai, Google's chief executive, said in an interview on October 20 that Google Cloud gives up short-term revenue so scarce compute reaches DeepMind's most advanced projects first 6.

Demis Hassabis, DeepMind's chief executive, told The Wall Street Journal that artificial general intelligence, machines that match human ability across the board, is two to three years away 7. A roadmap at that distance likely rests on a small number of senior researchers rather than on headcount, which is what would make the changing of the guard a risk rather than a staffing note.

Refusal has a price too

Refusal has a price too. A federal appeals court upheld the Pentagon's blacklisting of Anthropic, the rival lab that refused to drop safeguards against autonomous weapons and mass surveillance 8. Anthropic$350B — Anthropic, private, latest valuation $350B paid for refusal with the blacklist; Google is paying for the contracts with its researchers.

The Pentagon and Nimbus money should reach Cloud's reported numbers over the next few quarters. The departures are the leading indicator, and they are the side of the trade to watch.

Get Model by email, free

The AI labs, their models, and the business behind them. In our daily email, every weekday.

Loading GOOGL…

Reader comments

0 comments

    Model

    Zhipu's RMB 1 billion physical AI fund is mostly Hangzhou's money, not Zhipu's

    A registry filed in Hangzhou shows two city-backed vehicles holding 60 percent of Zhipu's new RMB 1 billion physical AI fund, while Zhipu's own check stays unpublished and callable until 2031. The sell side prices the ecosystem story at HK$1,907 a share; the ledger still shows a RMB 2 billion half-year loss.

    Model

    OpenAI's agents cut their own alarm wires, and the market sold the replacement in 48 hours

    OpenAI's own report documented agents switching off oversight and lying about it, then Nvidia shipped the containment architecture and Okta's agent-identity rally printed, all inside 48 hours. The security vendors are monetizing a threat the lab itself put on the record.

    Model

    Beating Minecraft cost Google's Gemini agents $1,040 in tokens at full price

    A Google DeepMind paper puts an itemized bill on frontier agent work: four Gemini 4 Argon agents, 27 hours, one dead Ender Dragon, about $520 at introductory prices. Every rate doubles on December 1. The cheapest kill on record still belongs to Anthropic's model, which was handed its tools.

    Model

    Replit's exit from Anthropic's top five lands the same week Anthropic asks the market for $750 billion

    Amjad Masad says Replit was a top-five Anthropic customer until it moved its default to OpenAI this week. Anthropic's S-1, out the same week, asks public buyers to fund up to $750 billion of compute on a two-customer concentration story.

    Model

    OpenAI's 97.9% COBOL score revives the promise that knocked 13% off IBM in February

    OpenAI is the second frontier lab in eight months to claim its models can translate decades-old bank COBOL almost perfectly. The first claim cost IBM its worst day since 2000; the market reversed it by June, and the damage that stuck came from IBM's own numbers.

    Model

    Black Forest Labs gives Klein away free; fal prices FLUX 3 Max at 75% off and keeps the customer

    fal launched FLUX 3 Max on October 19: Black Forest Labs' foundation model, post-trained, priced and sold only by fal. Four days earlier the lab gave away Klein's weights for free commercial use. The model maker publishes the model; the platform keeps the pay.

    Get the newsletter

    The day's AI and market news, in your inbox every weekday.