Ivan Poupyrev knew the timing looked terrible. March 2023. Silicon Valley still reeling from the sudden collapse of SVB. ChatGPT mania flooding every pitch deck with questionable LLM wrappers. And he was about to walk away from Google's legendary Advanced Technology and Projects group—ATAP, the skunkworks that had shipped radar gestures into millions of Pixel phones and woven touch controls into Levi's denim jackets.
But Poupyrev, along with four colleagues who'd spent years turning implausible hardware concepts into actual shipping products, had glimpsed something they couldn't ignore. The physical world generates trillions of data points every second. Sensors embedded in buildings, vehicles, factories, traffic lights, warehouses. Mostly that data just sits there, trapped in proprietary systems or simply discarded. What if there was a ChatGPT, but for reality itself?
Twenty months later, that question has evolved into Archetype AI—a Palo Alto startup with $48 million in funding, peer-reviewed research, and early deployments ranging from Japanese construction sites to suburban Seattle intersections. The company's bet: Physical AI, systems that perceive and reason about the real world through sensor data, represents the next frontier after language models conquered text. Whether that bet pays off may determine if a handful of ex-Google engineers can outrace the inevitable big-tech response to their own idea.
Leaving the Mother Ship
The founding team reads less like a typical startup roster and more like a greatest-hits compilation of ambient computing. Poupyrev, who holds a Ph.D. in human-computer interaction from Hiroshima University, led Project Soli—miniature radar for gesture control—and Project Jacquard, the interactive textiles effort, at Google ATAP. His co-founders brought equally rarified credentials: Jaime Lien ran sensing at ATAP after stints at NASA's Jet Propulsion Laboratory, with degrees from MIT and Stanford. Nick Gillian led machine learning for Soli and Jacquard; he created the widely-used Gesture Recognition Toolkit during a Fulbright fellowship at MIT Media Lab. Leonardo Giusti brought design expertise from ATAP, Samsung, and an MIT postdoc in human-computer interaction. Brandon Barbello rounded out operations and product after 13 years building consumer products at Google.
What they'd learned at ATAP—beyond how to navigate Google's bureaucracy—was how to ship impossible-sounding hardware AI into actual consumer devices. Project Soli went from lab prototype to detecting hand gestures via millimeter-wave radar in Pixel phones. But the team kept hitting the same wall. Every application required custom models trained on narrow datasets. A radar system trained for gesture recognition couldn't help with health monitoring. Camera models for safety inspection couldn't interpret accelerometer data.
The physical world, it turned out, doesn't speak a universal language the way text and images increasingly do for foundation models. Not yet, anyway.
Newton's Universal Physics
The technical challenge Archetype set out to solve sounds deceptively straightforward: create a foundation model that understands physical signals the way GPT-4 understands language. Simple to state, nightmarishly complex to execute.
In practice, that meant training on wildly heterogeneous data—accelerometers measuring vibration in hertz, thermometers reading degrees, radar tracking motion, cameras capturing scenes—and finding a way to encode all of it into a unified embedding space. The team's October 2024 research paper, published on arXiv, detailed their approach. They trained Newton (the name itself a not-so-subtle claim to foundational status) on approximately 590 million cross-modal sensor samples, fusing multimodal time-series data with natural language descriptions of physical phenomena.
The breakthrough, at least in their telling: teaching the model to perform zero-shot predictions across entirely new physical systems. No explicit physics equations required. No domain-specific training. In their experiments, Newton predicted pendulum dynamics, electrical grid demand, and transformer oil temperature from raw sensor streams it had never encountered before.
They call it a "Large Behavior Model"—a deliberate cousin to large language models, but for physical behaviors instead of words. The model doesn't just classify or detect; it reasons about context, predicts outcomes, generates insights in real-time. Perhaps more importantly for enterprise adoption, Newton runs on-premises or at the edge, processing sensitive industrial data without sending it to the cloud. A requirement, not a feature, for most infrastructure deployments.
Whether Newton actually delivers on that promise at scale remains an open question. The research demos look compelling. So did countless other AI breakthroughs that failed to survive contact with messy reality.
Stealth, Then Speed
Archetype emerged from stealth in April 2024 with a $13 million seed round led by Venrock, with participation from Amazon's Industrial Innovation Fund, Hitachi Ventures, and others. By then the company already had early customers—Infineon became the first to offer Newton's developer platform alongside sensor hardware. Fortune Global 500 companies were running quiet pilots in automotive, construction, logistics, retail.
The Wired profile that accompanied the launch captured some early demos that straddled the line between genuinely useful and tech-demo parlor tricks. A package dropped at a front door triggers an alert without cameras—just motion sensors detecting the distinctive thud pattern. A car's embedded sensors infer a driver approaching and automatically pop the liftgate. Volkswagen was running experiments, though not for autonomous driving (a clarification that suggested someone at VW's legal department had thoughts). Amazon's investment arm highlighted potential for robot perception and warehouse optimization.
By March 2025, Archetype launched "Lenses"—configurable applications built atop Newton to transform sensor streams into domain-specific insights. Japanese construction firm Kajima deployed a Lens to analyze 5 terabytes of data across 27 cameras and four years of construction footage, measuring project progress and identifying blockers. The goal: enable half of project management to happen remotely. The City of Bellevue, working with partners including Khasm Labs, AT&T, and Dell, deployed a pedestrian safety Lens processing 43 gigabytes per intersection per day on edge servers. The aim: proactive traffic signal adjustments and instant incident response.
Then came the November 2025 Series A, which told the story of rapid scaling. IAG Capital Partners and Hitachi Ventures co-led a $35 million round, joined by Bezos Expeditions, Samsung Ventures, and returning investors. Total raised: $48 million across two rounds in nineteen months. The company announced Physical Agents—tools to build and deploy AI agents that sense, understand, and act through Newton's platform. Early enterprise customers included NTT DATA deploying agents across warehouses, construction sites, and streets. The roadmap teased deeper fusion of physical signals and language, plus generating continuous time-series data from natural language prompts. Steps toward AI that doesn't just perceive but acts on the physical world.
Against the Wave
Archetype positions itself carefully against the broader "physical AI" wave cresting through venture capital. While companies like Physical Intelligence—which raised $400 million in late 2024 and reportedly hit a $5.6 billion valuation by late 2025—focus on robotic manipulation and embodied AI, Archetype targets something wider and less defined. Newton isn't a robot brain. It's infrastructure for any system with sensors.
Poupyrev framed the distinction in a June 2025 essay for the World Economic Forum: Physical AI should augment human intelligence for complex environments, not just automate repetitive tasks. A convenient rhetorical position for a company that doesn't actually build robots.
The competitive landscape validates the category's momentum, if not Archetype's specific approach. Nvidia launched its Jetson Thor platform for physical AI at the edge. Google DeepMind stood up a world-modeling team for games and robot training. The trillion-sensor economy that investors referenced in Archetype's seed round is no longer speculative hand-waving—it's coming, though who captures the value remains unclear.
Challenges remain, and they're not trivial. Archetype hasn't disclosed revenue, pricing models, or how many pilots have converted to paid deployments beyond the named partners. The company's team hovers between 11 and 50 people according to LinkedIn—tiny compared to competitors with deeper pockets and established customer relationships. And the core technical promise—zero-shot transfer across radically different physical domains—needs validation at scale beyond research demos and friendly pilot customers.
But the ex-Google team has shipped ambient AI hardware to millions of devices before. They spent two years since leaving ATAP building something more universal than any single sensor application. Whether Newton becomes the foundational model for the physical world, or simply another specialized tool in an increasingly crowded market, depends on questions the next twelve months should start answering.
Can it truly generalize across industries and use cases? Will enterprises pay meaningful money for it? And can a startup outrace inevitable efforts from Google, Microsoft, Amazon, and Nvidia in the same space?
The early deployments suggest the technology works, at least in controlled environments with motivated partners. Now comes the harder part—proving the market does too. Poupyrev and his co-founders have left Google before. They know what it takes to ship the impossible. This time, they just don't have Google's resources to do it.
