AI Is Changing How American Concrete Gets Made

The United States pours about 400 million cubic yards of concrete every year. It's in the bridges, the data centers, the highway overpasses you drive under without thinking about them. And for the most part, the concrete itself is made here. The problem is the cement that goes into it.

Roughly 20 to 25 percent of U.S. cement consumption comes from imports. That's the binding agent, the ingredient that actually holds concrete together, and a significant share of it is manufactured somewhere else. Domestic producers have known this for years. The harder question has always been what to do about it, because switching cement sources isn't as simple as swapping a supplier. Different cements have different chemistries. A mix that works with one can crack, set too slow, or lose strength with another. Reformulating means months of lab work, and most producers don't have the time or budget to run that experiment speculatively.

That's the actual problem Meta's research team set out to solve, not the cement import gap directly, but the design bottleneck that makes it so hard for producers to move away from what they already know works.

What Bayesian Optimization Actually Does Here

The tool they've built is called BOxCrete, short for Bayesian Optimization for Concrete. The technique behind it, Bayesian optimization, is well-established in machine learning. The idea is that when you're searching a large space of possibilities, you don't have to test randomly. You can build a probabilistic model of which regions of that space are most likely to contain good answers, test there first, update the model with what you learn, and iterate. It's the difference between searching a room by opening every drawer in order versus looking in the places where people usually stash things.

Applied to concrete, it means the AI learns from historical mix data, proposes formulations most likely to hit target specs, and gets better with every lab result. Engineers still run the tests. They still sign off. Code compliance doesn't change. But the number of iterations to reach a viable mix drops considerably, and that's what actually unlocks the ability to work with unfamiliar domestic materials without betting months of lab time on it.

The latest version of BOxCrete also predicts concrete slump, which is basically a measure of how workable a mix is before it sets. That matters because a mix that looks great on paper can still be a nightmare to pour. Adding that to the model makes it more useful to the people actually on job sites.

Rosemount

The clearest proof point so far is a data center Meta built in Rosemount, Minnesota. They used BOxCrete to design the concrete for the building's foundation slab, which has to support the weight of thousands of servers and cooling systems. Structural requirements don't get much more demanding than that.

The AI-optimized mix, made entirely from domestic materials, hit full structural strength 43 percent faster than the original formula. Cracking risk dropped by close to 10 percent. It's now qualified for use in additional sections of the same data center.

That's a meaningful result, and it matters more than a lab study because it happened under real construction conditions with real stakes.

Beyond One Company's Buildings

Meta open-sourced its concrete AI framework back in 2023 under an MIT license, and that decision is paying off in ways that are arguably more significant than anything Meta does with it internally. A Pennsylvania company called Quadrel, which builds enterprise software for the ready-mix industry, has embedded the framework into daily mix design and quality control workflows for its customers. The models improve over time as field results feed back into training. Producers using Quadrel are running AI-assisted mix design without having to become AI researchers.

That's the path to actual industry impact. Not one company optimizing concrete for its own data centers, but the tools becoming part of the standard software stack that producers already use.

Where This Goes

There's a reasonable skepticism worth naming here. AI getting applied to concrete sounds like a press release more than a revolution. The construction industry has seen plenty of technology promises that mostly stayed in pilot programs.

The difference in this case is that the work is already out in the field, the code is open, and the adoption is happening through existing commercial software rather than requiring producers to change their workflows. The barriers are low enough that the question isn't really whether this spreads. It's how fast.

Domestic cement producers are already responding to the moment. Amrize, the largest cement and concrete manufacturer in North America, launched a Made in America cement label and announced close to $1 billion in capital investment for 2026 to expand domestic production. If supply is growing and the design tools to work with domestic materials are getting better and cheaper to use, the import share starts to look less fixed than it has been.

Concrete doesn't get much cultural attention. But it's one of those materials where small improvements at scale add up fast. Forty-three percent faster cure time on a foundation slab isn't a headline. Multiply it across a significant fraction of 400 million cubic yards a year, and it starts to matter.