Archetype AI

AI goes into the real world 🌎 Today, we're starting a new chapter of Archetype as we celebrate our $35M Series A. The physical world generates vast amounts of sensor and video data every day but most of it goes unused. After years of AI living behind screens — in dashboards, apps, and digital abstractions — intelligence is finally stepping into the physical world. We're bringing the Archetype platform to more customers and evolving Newton from a powerful foundation model into an intelligence layer for the physical world—one that perceives, understands, and reasons across real-world systems, enabling hundreds of intelligent applications. And this is just the beginning. This funding round, led by IAG Capital Partners and Hitachi Ventures, with participation from Bezos Expeditions, Venrock, Amazon Industrial Innovation Fund, Samsung, E12, Systemiq Capital, HLV, and others, enables us to scale physical intelligence globally and continue advancing Newton as the premier foundation model for the real world. Thank you to our team, partners, investors, and all of you for following our journey to making the world truly intelligent. Onwards! 🚀

Welcome to Archetype AI, Ethan Chang! Joining the team as our first Design Fellow, Ethan is a part of MIT's Design Intelligence Lab and Ideation Lab and has an impressive portfolio spanning AI cohabitants, generative games, and intelligent systems.

One foundation model, adapted across very different operations. You can now fine-tune Newton, to the specific behavior of: 🔹 A manufacturing line — learning the failure signatures of your specific equipment and the tolerances of your specific processes. 🔹 An energy asset — internalizing the wear patterns of your turbines, pumps, or compressors across your operating conditions. 🔹 An industrial facility — capturing the environmental and operational fingerprint that makes one site different from the next. The mechanism is the same across all three. Newton's general understanding of physical signals, adapted with your data, running on your infrastructure. Years of sensor history that previously sat unused becomes a working model of how your operations actually run, and one that keeps adapting as those operations evolve.

Welcome to Archetype AI, Michał Łowicki! Michał joins us as a Member of Technical Staff after nearly six years at Datadog and before that, leading rapid release engineering for Facebook. Welcome aboard! 🚀

@_jaimelien_ RSVP here: luma.com/gtvcxn0s

Join us on May 20 for the ‍Newton World Model discussion featuring Archetype's Co-Founder and Chief Scientist @_jaimelien_. She will share new research highlighting the model capabilities and introduce our latest Newton Fine-Tuning features — an ability to turn operational data into a model adapted to your specific machines and workflows.

Meet the Task Verification Agent, powered by Newton, Archetype's foundation model for the physical world. In this demo with @XbloomCoffee the agent watches a real workflow, checks each step against the intended task, and flags deviations the moment they happen. No labeled training data or explicit training; Newton can generalize across use cases and facilities and create a dynamic representation of what's happening in the real world. The same approach generalizes to manufacturing and many other environments where completing steps in a correct order matters.

A gearbox slowing down could mean it's failing. Or it could mean a technician just opened the outer bracket to start a repair. Without fusing multiple sensor streams, a model has no way to tell the difference. This is the real challenge in Physical AI and that’s why it goes far beyond robots. The physical world doesn't produce one type of data. A single factory generates simultaneous, context-dependent signals from cameras, pressure gauges, vibration sensors, rotations, and more. Making sense of that environment requires a model that understands what's actually happening across the whole system instead of just one data stream. Archetype AI CTO and Co-Founder @NickGillian on why that distinction matters 🎧 #MLOpsCommunity

We benchmarked Newton, our Physical AI foundation model, across a number of smaller devices including CPU-only. Read on for what the results show and why deployment flexibility is now essential for industrial-grade Physical AI. The logarithmic chart below brings components with very different throughput regimes — including the MAC-class components — onto a single plot. The main finding: Newton is deployable across the entire stack. Mac M4 configurations deliver the highest throughput, as expected. The chart below also shows that CPU-plus-GPU edge servers deliver more than enough capacity for sustained, multimodal monitoring at the site level. On a constrained CPU like a Raspberry Pi 4, Newton still produces meaningful inference for scoped tasks; that mean it can be leveraged for use cases that previously had no realistic AI path at all.

We're excited to welcome Nathan (Dae Hyun) Nam to the Archetype's Product team! Nathan brings deep expertise in product management for developer platforms and API infrastructure, with 10+ years scaling connector ecosystems and data tooling, previously at Observe and Confluent. Welcome to Archetype AI Nathan!

Thanks to @Politico for including @PhysicalAI in this week's piece on the world models and the key use cases for Newton. Read on: politico.com/newsletters/di…

In offshore drilling, Invisible Lost Time can cost tens to hundreds of thousands of dollars per instance. It hides in the gap between what sensors record and what operators see in real time. Drilling Monitor, our new Newton-powered demo, runs on open data from Equinor's Volve field — 14 wells in the North Sea, 2007–2009. Nine sensor channels (ROP, RPM, pressure, hookload, flow rate, bit depth, block position, hole depth, weight on bit) stream into Newton in 25-sample windows. The Process Monitoring Agent runs few-shot classification against 500 drilling and 500 not-drilling examples, returning rig state in real time: DRILLING or NOT_DRILLING. No years of labeled history required. One foundation model and 1,000 reference examples drive live rig state across every well. Traditional ML means one model per asset, condition, and site, constantly retrained. Newton fuses multichannel sensor data into a representation that transfers across wells, so the same Process Monitoring Agent extends to stuck pipe, pump anomalies, or any other physical state next.

Learn more: archetypeai.io/blog/newton-th…

Introducing Newton Fine-tuning — a new way to adapt Physical AI foundation model to your operations. Out of the box: strong priors over physical signals — vibration, current, thermal, acoustics — across a wide range of industrial assets. After fine-tuning on your proprietary operational data: a local world model that has internalized the specific behavior of your systems, the failure signatures that matter at your facility, and the process variables unique to your production environment. Runs entirely within your infrastructure. No data movement required.

.@ipoupyrev at @semafor World Economy on why physical AI has to be sovereign by design: "Recently there was a conversation that we are running out of data for language models. In the physical world, we don't have that problem. But when you talk with companies in the physical world, like semiconductor manufacturing, the data they collect from operations is core IP, top secret. They're unlikely to give that data to large hyperscalers for training AI models."

At #HumanX last week, our Chief Scientist @_jaimelien_ spoke about one of the core ideas behind how we build Physical AI: every machine collecting sensor data can develop its own local world model, learned entirely from its own data.

How can AI understand the real world if it only sees images? There is a distinction between the world as perceived and the world as objectively construed. As Dr. Fei Fei Li noted in her essay on Spatial Intelligence, training AI solely on language can be limiting because language is not the world itself but only its representation. While not typically classified as “language”, it could be argued that images offer us a representation of the world as well. As Wittgenstein wonders “Can I see something as it, which is possible to be represented in an image?” VLMs so far have yet to give a concrete answer. That's why I keep coming back to the idea that real understanding — for humans or AI — can’t come from images alone. Our world is full of signals we can’t always see, or don’t see until it’s too late. So what could get us closer to the world as it actually is? The answer could be sensor data: the continuous, embodied stream through which reality reveals itself spatially and temporally. This is the bridge we're exploring between the physical world and how AI can understand it.

🎉 We are excited to welcome Sylvain Lebresne to our Platform Dev Team! Sylvain brings a wealth of experience in distributed systems and database infrastructure from IBM, DataStax, and Apollo GraphQL. Sylvain, welcome to the team!

In an orchestra, every musician can be skilled, but without a conductor all you get is noise. That’s the state of most industrial infrastructure today. Factories, plants, and energy grids are already full of smart devices. Sensors capture vibration, temperature, current, and motion in real time, each doing its job. The problem is they don’t talk to each other and there is no way for operators to understand what a combination of signals actually means for the system as a whole. Physical AI is the intelligence layer that changes this. It isn't about robots or humanoid forms. It's about giving industrial infrastructure the same kind of general reasoning capability that foundation models gave the digital world. That's what Newton is built for; a Physical AI foundation model trained on real-world sensor data that reasons across modalities and surfaces patterns no single signal source could reveal on its own. Learn more about our approach to #PhysicalAI: archetypeai.io/blog/language-… #FoundationModel