Standard Matter

Where AI meets human ingenuity

What excites us about AI is where it meets human ingenuity. It supports the people doing important work for the whole economy. It understands systems too complex to program by hand. It expands what people can discover, design and build.

AlphaFold has achieved this in cell biology, and GenCast in atmospheric prediction.

The frontier for complex physical systems

We are now at the frontier of this breakthrough for complex physical systems, and especially ones with people at their center. This opens an era of enormous potential for manufacturing, industrials and energy.

Physical AI is adaptive. It is useful for handling variance: the unpredictability inside a factory and outside it. Manufacturing took adaptation out of the process to make itself static and linear.

The linear process was a trade-off. It expanded capacity without needing proportionally more labor, and it tied manufacturers to products they can make in large quantities and processes they can repeat at high volume. Manufacturing and physical AI have to change in tandem, and the change can be gradual.

The detail our models hold

The physical world and expert intuition are more richly detailed than we have been able to capture. Today our models can represent those details: their structure, their state, and how they react to change.

A model that holds the state of a plant, or of the supply network around it, can say what happens next when something changes.

Much of this detail is what people pick up on their own when they work in a factory: when they operate equipment, when they design a process and get an outcome from it, when they meet a demand. These small data points give a person intuition about their own role. That data can now be replicated and simulated, at a wider scope than any one person can manage.

A person is limited in scope, and by their most recent intuition. The relationships within the data have to be obvious before a person can make sense of them. For machine learning they don’t have to be obvious, and the data doesn’t have to be filtered or focused first.

Everything goes in as one input. The input is turned into vectors, numerical representations of information, and the vectors are related to each other. We can look at wide trends and at small ones, and find small trends inside the wide ones. Connecting everything reveals relationships we didn’t know existed.

Efficiency is just efficiency

A gain in one part of a plant doesn’t reach the output if it moves the constraint to another part. The same holds for a plant inside its supply network.

A factory that optimizes alone is still bottlenecked by every system around it. It runs out of supplies, and then of buyers, before it sees the real gains. So we model the whole plant and the whole network around it.

Built with the people on the floor

Turning this data into the right mathematical representations is not only a technical problem. It takes working hand in hand with plant operators, engineers and the people on the factory floor who best understand the processes behind the data.

A lot of this looks like removing humans from the loop. What we’re removing is the decision load.

Today every decision waits on a person. When decisions are prepared in advance, people stop being the bottleneck in gathering all of that information, and they have the leverage to make the best decision possible.

People now manage production, largely through their labor. We see them managing capacity instead, as they do in the Shenzhen ecosystem.

We focus on technology that lets a manufacturer empower its own people. Those people become the engineers who run the factories they are building with our technology. Our growth does not depend on the size of our own engineering team working inside those factories. Manufacturers develop their own workforces and run their plants themselves, and the people already on the floor gain the skills to take on the engineering roles.

This is also the opportunity to redefine what it means to work in manufacturing, and to win back the next generation of skilled people. We’ve had conversations across the Midwest where parents have told their children not to get into manufacturing. Ten years ago, parents told their kids to learn to program. Ten years from now, we want them telling their kids to learn process engineering, to learn how to control robots, and to learn how to create demand and think creatively about products.

Our research platforms

We have internal research platforms that take the rich and complex data of the physical world and calibrate it against a vast public and proprietary data set, so the complexity of the physical world is modeled accurately.

They run on two foundation models, made to be used together. The first holds the state of the processes inside the plant: every part, machine, shift and breakdown, and how each process depends on the others. The second holds the state of components and materials across the supply network: what each part is made from, what depends on it, what can substitute for it, and where the risk sits.

The two are aligned, so a decision in one is weighed in the other. A change on the plant floor shows up in the network as more or less capacity at that plant, and the network says what that capacity is worth in deliveries. A material that’s going to be late shows up on the floor as the hours it will cost and the work that can be done in its place.

The two also share one time scale. The network model says when a lot arrives, the plant model says when it is needed, and a gap between the two dates shows up while there is still time to act.

Whole factory data and sensing feed both models. Robots act on what the models show.

Manifold: whole factory optimization

Manifold lets us run simulations on our foundation models to explore the world of possibilities.

Inside a factory, many people make decisions, and those decisions are almost never coordinated. The coordination that exists is reactive: a person can only see the demand, or the problem, right in front of them. It’s hard for any one person to see the downstream effects and the upstream implications.

Whole factory optimization solves that at the level of the factory floor. It lets us understand the factory’s capacity, and properly allocate it at all times.

A running plant can’t be rearranged to see what happens. Manifold simulates it. One optimum hides the trade between automation and workforce, and the people deciding need to see that trade. So Manifold keeps the best plant it finds for every combination of the two.

Tributary: supply and demand optimization

Tributary combines the forecasting abilities of frontier models with the possibilities Manifold models, so a manufacturer is as resilient as possible.

A plant is bottlenecked by who is mining, who is refining, and who is willing to sell to it at what price.

Everything that’s made requires 200 other things, and those 200 things require another 200 things.

Tributary runs on our model of that network: every component, material and site, and what each one needs from the others.

One of our favorite examples is a late truck. A plant is waiting on a supply shipment, and the truck says it will be four hours late. Supply and demand optimization is aware of that truck and knows, from history and extrapolation, that it will probably be more like six. Whole factory optimization says what capacity is available and what the factory can do.

The decision then becomes clear: how to reallocate capacity around the part that’s missing, and what can be prepared upstream or downstream in the meantime. A plant that planned for four hours and reallocates by hand is down for two, even if it is tightly run.

CONTRAST: physical AI for robotic integrations

Our third research platform, CONTRAST, uses physical AI to improve robotic integrations.

Implementing robotics in manufacturing is bottlenecked by engineering capacity. CONTRAST lets engineers perform more integrations of robotic automation, and maintain those integrations at scale.

Through contrastive state, CONTRAST makes intelligent visual decisions in real time, which gives engineers a more intuitive experience.

This architecture improves the metrics engineers care about, and lets them focus on more important parts of the process.

Right now, the only way to get precision and repeatability is programmatic robotics: robots programmed from fixed states, unable to make decisions for themselves. A robot operated by physical AI keeps that precision and repeatability in adaptive situations. It perceives its surroundings and handles the variance in them.

The cost of engineering is mainly engineering time. We keep the motion that already works in a robot cell and learn only the decisions between its steps, so a new part doesn’t mean teaching the whole task again. The time to program comes down, and so does the time to evaluate the benefits and downsides of an integration.

Because physical AI allows adjustment in real time, robots can guide and adjust themselves on cheaper motors and cheaper arms. That lowers the hardware investment and opens the technology to small and mid-sized manufacturers.

The next era of economic growth

Developing and aligning these frontier models together, as researchers and operators, gives industrial experts the tools to predict system behavior, optimize production and deploy autonomy where it creates capacity.

Our aspiration is to help them reduce delivery backlogs, meet demand, expand industrial output and lay the foundation for the next era of economic growth.

Each piece produces benefits immediately, while the things the later pieces depend on are being solved in parallel. Whole factory optimization is a step change on its own. So is supply and demand optimization.

It will feel piece by piece, and then all at once.

The data, the aggregation, the intelligence and the decision making reach critical scale together, and the benefits cascade. We see a scaling law at play. There’s something exponential in collecting enough data, in using that data properly, and in figuring out what it can teach us that we may not have foreseen.

To us, the endgame is a manufacturing sector that is reactive, proactive, and prepared. Manufacturers meet the demands they need to meet, create the demands they want to create, and de-risk as they do it. People do the same, at the same level as manufacturers.

We believe adaptive and accessible manufacturing is a uniquely American edge. It comes from our fundamental rights and our ability to participate as sovereign citizens: the right to make things, the right to modify things, the right to choose with our money which things are bought in the marketplace, the right to answer any demand and figure it out.

America can be the best industrial state, because of those rights and Americans' drive to create things. It is the place where dreams are built, and the place people go when they want to make things. Manufacturing faces barriers on the capex side, in policy and regulation, and in the number of people to hire. We are building the solutions that lower them.