Google has acquired Common Sense Machines, a Cambridge-based startup building AI models that convert 2D images into game-ready 3D assets. A Google spokesperson confirmed the deal to The Information on January 24, though financial terms remain undisclosed.
The entire CSM engineering team, estimated at 12 to 17 employees, will join Google DeepMind.
Coming home
For CSM co-founder and CEO Tejas Kulkarni, this is a return trip. He worked as a senior research scientist at DeepMind from 2016 to 2019 before leaving to start CSM in 2020 with Max Kleiman-Weiner, a PhD researcher from MIT.
Kulkarni's academic background spans deep learning, probabilistic programming, and reinforcement learning. His bet in 2019 was that 3D would follow the trajectory of image generation, which was just beginning to work at scale.
CSM built Cube, a platform that transforms sketches, images, or text prompts into textured 3D meshes suitable for game engines. The company claimed first place on the 3D Arena benchmark, an independent evaluation system hosted on Hugging Face that measures how well AI models handle 2D-to-3D conversion.
The valuation question
According to PitchBook, CSM raised around $10 million total from investors including Andreessen Horowitz, Toyota Ventures, and Intel Capital. The startup's last known valuation sat at roughly $15 million.
Whether Google paid a premium or grabbed a distressed asset is unclear. The undisclosed price combined with the small team size suggests this leans toward an acqui-hire, where the talent matters more than the technology. But CSM did ship working products with paying customers, so the code has value too.
Why DeepMind wants 3D
DeepMind has been building toward what it calls "world models" for years now. The Genie project generates interactive 3D environments from text prompts, while SIMA agents learn to navigate and complete tasks in virtual worlds.
These systems need endless variations of 3D content for training, and that content is expensive and slow to produce manually. CSM's tech could help DeepMind generate training environments at scale without paying artists per asset.
The timing makes sense given DeepMind's recent publications. Genie 3, announced in late 2025, can generate multiple minutes of navigable 3D worlds in real time at 720p. The model uses techniques from neural radiance fields and Gaussian splatting, the same spatial representation methods that CSM built its pipeline around.
What happens to CSM's products
CSM was still actively operating when the deal closed. The company had partnerships with gaming studios and had recently integrated with 3D Cloud for product visualization workflows.
Google hasn't announced what happens to 3d.csm.ai or the Cube API. Existing customers will want clarity on whether their workflows keep working or need migration.
The acquisition follows a familiar pattern for small AI startups getting absorbed by tech giants. The independent product usually disappears, the team ships features for the parent company, and anyone who built dependencies on the original product scrambles to find alternatives.
