At Google, AI Has Never Cost So Much (and Morale Has Never Been So Low)

Google beats its estimates and its cloud explodes at +82%. Its AI models are falling behind, its best researchers are leaving to compete, and its cash flow is turning red for the first time in twenty years.

For a long time, the story that Google loved to tell was a few simple steps: Demis Hassabis takes the helm of the DeepMind-Google Brain merger in 2023, Gemini arrives to dethrone ChatGPT, and Mountain View regains command of an AI race in which it had never really won. For a few quarters, the narrative held: Google Cloud was accelerating, Gemini was making progress on benchmarks, and competitors seemed to be falling behind. Then Anthropic came along with Mythos. Then Google signed a contract with the Pentagon in April authorizing the American army to use its technologies. On July 23, 2026, the most paradoxical quarter in Alphabet’s recent history opened to the general public: $119.8 billion in revenue+24%, a cloud whose revenues jumped 82%, and a stock which still lost 5% after the close. The reason? Alphabet raised its annual AI spending envelope to $205 billion, its free cash flow went negative for the first time in nearly two decades, and six anonymous and cautious inside sources chose this moment to describe what’s really going on in the halls of DeepMind.

When the military contract ignited the powder

What these sources describe is two threads that eventually come together. The first is competitive: Gemini 3.5 Pro, the most powerful model in preparation, is several months behind schedule. Meanwhile, Google this week delivered a series of lighter, cheaper Flash models to mixed reactions. Alexandr Wang, chief AI officer of Meta, summed up the mood with one sentence on X: « Gemini who ? » In the ranking of the most used models on the OpenRouter platform, no Gemini model appears in the top 10. The reason given internally: Google has devoted considerable energy to defending its Search market share against ChatGPT, to the detriment of agentic code in which OpenAI and Anthropic have taken a head start.

On the other side of the picture is the part that Google would prefer to see less exposed: the ethical crisis surrounding the contract with the Pentagon. Signed in April 2026, it authorizes the American army to deploy Google technologies, and it has provoked several resignations, the reason for which comes up in exit interviews. Alex Turner, a former researcher at DeepMind, publicly confirmed that he left the company for this reason. He also questioned the lack of internal presence of Demis Hassabis with his teams, considering that Sam Altman and Dario Amodei maintain a closeness with their collaborators that the boss of DeepMind does not have (in a race where the market in recent years has shown that talent retention is the most critical variable, the observation deserves to be taken seriously). For at least one of the sources interviewed, opposed to the military contract, the situation resembles a “permanent battle” which gave rise to a “emotional exhaustion”. The most painful names to pronounce on the Google side are known: Noam Shazeer, co-lead of the Gemini team, who left for OpenAI after Google had picked him up at a high price less than two years earlier ($2.7 billion in August 2024, for the record); and John Jumper, Nobel Prize winner in chemistry for his work on AlphaFold, now at Anthropic. Every start in this category is a difficult equation to compensate for in a race where brains are the only variable that really matters.

Google provides a formal denial of these accusations. Attrition among AI teams in the first half of the year would be lower than the same period last year, and more than 90% of candidates made an AI offer would accept it. “We are pleased with our latest Flash launches, our roadmap and the incredible demand for our models”said a spokesperson. This is not inaccurate: the Gemini app claims 950 million monthly active users and the APIs process 22 billion tokens per minute. The problem is not that no one uses Gemini. That’s because everyone is waiting for 3.5 Pro.

205 billion in spending, and Google in the red for the first time in twenty years

The figures for the quarter give the measure of what is currently at stake. Google Cloud jumped 82% to $24.8 billion, with operating profit tripled and a margin of 36%. Search increased by 17% to 63.3 billion, YouTube increased by 13% to 11.1 billion. Google Cloud’s contract backlog hits $514 billion, a signal that capacity under construction already has buyers waiting. This record would be the dream of any tech company, but investors had their eyes on something else.

What they read was that Alphabet raised its full-year capital spending guidance to $205 billion (compared to 180 to 190 billion announced previously), and that the past quarter alone weighed 44.9 billion, approximately double that of a year earlier. The direct consequence: the free cash flow fell to -5.9 billiona negative figure that has not been seen in Alphabet’s accounts for almost twenty years. The stock fell around 5% in after-hours trading. The posted net profit of 112.1 billion seems impressive until we notice that 98 billion corresponds to an uncashed latent gain on the stake in SpaceX’s capital, which can go up or down depending on valuations: remove it, and the underlying results take on a much more ordinary air.

What Alphabet is doing resembles an industry gamble unlike any other in its history: spending more on investments over a year than it generates in actual net profit, in the hope that the cloud becomes self-sufficient before the capex wave crests. For Gemini’s European users, who number in the hundreds of millions, the question is reformulated differently: an AI which costs its manufacturer 205 billion per year, whose most anticipated model remains absent from the rankings, and whose teams anonymously describe low morale, says something about the trade-offs that have been made. And on those that Google will have to decide soon.