Teaching machines to perceive, reason and act in the physical world - not just words and pixels.
Above: the Archetype AI mark. The company's flagship model, Newton, is named for Isaac Newton and learns the laws of physics directly from raw sensor readings.
Archetype AI is a San Mateo startup with an unfashionable premise for an AI company: the most valuable problems are not in text or images, but in the physical world - the vibration of a failing motor, the heat of an overloaded transformer, the movement of a pedestrian stepping off a curb. Founded in 2024 by a group of researchers who left Google's Advanced Technology and Projects (ATAP) group, the company builds what it calls Physical AI: models that read the raw output of sensors and translate it into something a person can act on.
Its flagship is Newton, described as a first-of-its-kind foundation model for the physical world. Where a large language model is trained on the internet's text, Newton is trained on the signals streaming off cameras, radar, microphones, accelerometers, thermometers and dozens of other instruments. It fuses those streams together and pairs them with plain language, so an operator can ask a question - "is this equipment behaving normally?" - and get an answer grounded in physics rather than dashboards.
The company raised a $13M seed in April 2024 and a $35M Series A in November 2025, for roughly $48M in total, with backing from Hitachi Ventures, Amazon's Industrial Innovation Fund, Samsung, Venrock and Bezos Expeditions.
The biggest problems in the world are physical, not digital. We believe the power of AI can and should do more for us than incremental productivity gains.- Ivan Poupyrev, Co-Founder & CEO
Newton turns the noise of the "trillion sensor economy" into intelligence you can query in English. The pipeline, simplified:
Hundreds of sensor types feed raw, real-time signals.
Newton learns physical patterns across modalities - no labels.
Physical Agents deliver insight and action on-premises.
What makes the model unusual is how it learns. Rather than being taught the equations of motion, Newton discovers them - the company has shown it inferring patterns of mechanical oscillation, thermodynamics and electromagnetism from raw measurements alone. It reads a machine the way GPT reads a sentence.
Sensor modalities Newton can fuse:
The multimodal physical-world model at the core of everything - perceiving, understanding and reasoning about real environments in real time.
Where teams develop, customize and deploy Physical AI applications and agents on top of Newton - including on-premises, where the sensors live.
Tooling to build custom Physical Agents that turn multi-sensor streams into perception and action with minimal extra training.
Ready-to-deploy agents for industrial process monitoring, operator task verification and safety monitoring.
Archetype AI sells to enterprises and public agencies running sensor-rich operations. Early deployments span the map: Volkswagen for vehicle situational awareness, Kajima for construction project management, NTT DATA for industrial agents, and the City of Bellevue for pedestrian safety monitoring. Research groups like UC San Francisco's orthopaedic surgery team have also worked with the technology.
The common thread is that all of them run on the same underlying model - the whole point of a foundation model is that one thing generalizes instead of a hundred bespoke systems.
Organizations have spent a decade wiring up sensors and collecting data they never actually understand. Archetype AI's pitch is that the gap between collecting and comprehending is the business. Its agents move industrial monitoring from charts nobody reads to direct answers: is this process drifting, did the operator complete the step, is this a safety hazard.
Application areas: manufacturing, construction, energy, telecommunications, logistics, automotive and city infrastructure.
Organizations don't need more data - they need intelligence where it matters most: inside their operational environments.- Archetype AI, Series A announcement
Most AI reasons over text and images. Newton reasons over the analog, physical signals those systems ignore.
Newton discovers physical laws from unlabeled data instead of being hand-programmed with rules.
Intelligence runs where the sensors are - important for latency and data privacy in operational settings.
In a crowded field of world-model research from NVIDIA, Google DeepMind and Meta, and industrial-analytics vendors like Augury, Samsara, Palantir and C3 AI, Archetype AI's wager is on a single multimodal foundation model that generalizes across sensor types - rather than a separate model for every use case.
Five co-founders, most from Google ATAP - the group behind Project Soli's miniature radar chip.
20+ years turning R&D into products; led work at Google ATAP, Disney Imagineering and Sony.
Perceptual-sensing researcher; invented the Soli radar used in Pixel and Nest devices.
15+ years in AI and interaction technology at Google.
13+ years in hardware, software and AI product development at Google.
15+ years in design and R&D; formerly at Google ATAP and Samsung.
Capital raised by round (USD millions)
Series A led by IAG Capital Partners and Hitachi Ventures, with Bezos Expeditions, Venrock, Amazon Industrial Innovation Fund, Samsung, E12, Systemiq Capital, HLV, Gaingels and Plug and Play Ventures participating.
Former Google ATAP researchers, led by Ivan Poupyrev, set out to build physical intelligence.
Launched the Newton foundation model alongside a seed round led by Venrock.
Introduced the Archetype Platform, Agent Toolkit and Physical Agents.
A portfolio of ready-to-deploy Physical AI agents for industrial operations.
Poupyrev and Archetype AI featured in press on making everyday objects AI-readable.
It builds Physical AI - a foundation model called Newton plus a platform and agents that let AI perceive, understand and reason about the physical world by fusing real-time sensor data with natural language.
Newton is Archetype AI's multimodal foundation model that ingests signals from cameras, radar, microphones, accelerometers, thermometers and other sensors, learning physical patterns directly from raw data without human labels.
Founded in 2024 in San Mateo, California by former Google ATAP leaders Ivan Poupyrev (CEO), Brandon Barbello, Leonardo Giusti, Jaime Lien and Nicholas Gillian.
About $48M total: a $13M seed in April 2024 and a $35M Series A in November 2025, backed by IAG Capital Partners, Hitachi Ventures, Bezos Expeditions, Venrock, Amazon and Samsung.
Enterprises and public agencies in manufacturing, construction, automotive, energy, telecom and city infrastructure - early customers include Volkswagen, Kajima, NTT DATA and the City of Bellevue.