Google DeepMind invested approximately $75 million in film studio A24 for a research collaboration on AI filmmaking tools, but without access to A24's library. The partnership focuses on the creative process, aiming to teach AI which frictions to keep and which to remove. This signals Silicon Valley's growing valuation of judgment.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
When Silicon Valley is used to eliminating all friction with software, Google DeepMind has invested $75 million in A24, a company known for preserving 'creative friction,' without gaining access to its film library. This deal reveals a core question: How can AI video tools learn to distinguish which friction should disappear and which must remain?
In June 2026, Google DeepMind announced a long-term research collaboration with film studio A24, investing approximately $75 million. A24 is known for arthouse films like Moonlight and Everything Everywhere All at Once. Surprisingly, Google did not gain access to A24's film library and cannot use its content to train models. A24's filmmakers are also not obligated to use the related tools.
The collaboration focuses on the creative process: DeepMind researchers will test, iterate, and develop tools together with filmmakers. This means what Google truly wants from A24 may be a discernment that models have not yet mastered—which frictions in filmmaking should disappear and which must remain.
AI video has moved past the stage of 'whether anyone uses it.' In the US, Higgsfield's ARR grew from approximately $50 million in September last year to over $500 million in June this year, a tenfold increase in ten months. In China, Kling's global users exceeded 100 million, with nearly 50,000 enterprise clients, and Q1 revenue this year exceeded 650 million RMB, up over 300% year-on-year.
Product boundaries are also advancing: Google's Flow has moved from single-shot generation to character control, scene extension, and story orchestration; Adobe in June 2026 rolled out Creative Agent into Firefly, Premiere, and other tools. Silicon Valley capital is beginning to price 'judgment' separately—for example, San Francisco-based AI film production platform Flick completed a $6 million seed round with participation from True Ventures, GV, YC, and Lightspeed.
There is plenty of friction in filmmaking that should be eliminated, such as keying, transcoding, and asset retrieval. But another type of friction occurs before judgment is formed: why a character is silent, which beautiful shot should be cut, whether an uncomfortable pause is the most authentic few seconds of the entire scene. These have no clear process and are hard to write into efficiency metrics.
A 2024 experiment in Science Advances showed that participants who received AI help created stories that were on average more readable but more similar to each other. Visual design experiments also found that after exposure to AI-generated images, participants generated fewer ideas and showed reduced originality. AI assistance improves single-output quality while pulling different people toward similar local optima.
A24's value lies in its willingness, at critical moments, to let a temporarily unfavorable choice persist a little longer. Most studios fear that works are not mature enough; A24 sometimes fears that they mature too quickly.
The collaboration also has another side: drafts submitted by filmmakers, discarded storyboards, revision records, and reasons for rejection may contain higher-density judgment. Who owns these process data, whether they will be used to train general-purpose products, and whether filmmakers can opt out on a per-project basis—current public information has not provided complete answers. If co-development ultimately becomes artists providing free feedback that the platform then packages into general features, this collaboration will only accelerate the industrialization of aesthetics.
A better outcome: Google uses A24's real workflows to calibrate the boundaries of tool intervention, letting the model bear execution costs while leaving the exploration path and final decisions to creators. The next capability that creative software needs to develop is recognizing when a problem should not yet have an answer.
This article is primarily based on an original analysis from GeekPark, which cites some public data and experiments. However, some commercial data (e.g., Higgsfield ARR) are source claims and have not been independently verified. Experimental conclusions come from a Science Advances study, but the sample is limited and cannot be directly extrapolated to the film industry.
The core of Google's investment in A24 is not to acquire content, but to learn how to let AI preserve necessary 'friction' in creation—the hesitation and trade-offs that allow works to develop judgment. This suggests that the next competitive frontier for AI video tools is not generation quality, but a deep understanding of the creative process.
Primary report
Primary source