I didn’t expect to be writing about concrete this week, but here we are. Meta just released BOxCrete, an artificial intelligence model for designing concrete mixes, and it’s actually kind of fascinating. Not because concrete is sexy (it’s not), but because this is one of those rare cases where AI is solving a genuinely unsexy, high-impact problem that affects jobs, supply chains, and infrastructure at scale.
The U.S. pours about 400 million cubic yards of concrete every year. That’s enough to circle the Earth with a two-lane highway multiple times. But here’s the kicker: we import nearly 25% of the cement that goes into that concrete. This matters because cement is the binding agent in concrete, and different cements have wildly different chemistries. A mix that works perfectly with domestic cement might crack or fail with imported stuff.
The traditional way to design a new concrete mix is painfully slow. Engineers rely on trial-and-error lab work, intuition, and decades of accumulated tribal knowledge. If you want to switch from imported cement to a domestic supplier, you’re looking at months of testing to validate a new formulation. That’s expensive, time-consuming, and a massive barrier to reshoring.
Bayesian Optimization Meets Concrete
Meta’s approach uses Bayesian optimization through their Adaptive Experimentation (Ax) platform. Instead of randomly testing concrete mixes or relying purely on human intuition, the AI intelligently navigates the massive space of possible formulations. It suggests which experiments to run next based on previous results, learning as it goes.
BOxCrete improves on Meta’s earlier concrete AI models with better robustness to noisy data and a new ability to predict concrete slump, which is basically how workable the mix is when you’re pouring it. This isn’t just academic research either. Meta partnered with Amrize (the largest cement and concrete manufacturer in North America), the University of Illinois, and construction firm Mortenson to test this at scale.
They used BOxCrete to design a concrete mix for Meta’s data center in Rosemount, Minnesota. The AI-optimized mix, made entirely from domestic materials, reached full structural strength 43% faster than the original formula while reducing cracking risk by nearly 10%. That’s not a marginal improvement. That’s the kind of performance boost that changes project timelines and construction economics.
The mix was used in one of the most demanding parts of the build: the massive foundation supporting thousands of servers and cooling systems. With field data now confirming it meets all structural requirements, it’s qualified for broader use across the data center.
The Software is Already in Production
What’s more interesting to me as a developer is how quickly this moved from research to production tooling. Meta open-sourced their concrete optimization framework in 2023 under the MIT license. Now, Pennsylvania-based Quadrel, a SaaS platform serving the ready-mix industry, has integrated Meta’s AI framework into their software.
Quadrel is using it for real-world tasks: data preprocessing, batch normalization, feature engineering, and customer-specific model training. The models continuously improve as field test results flow back in, and they’re embedded directly into daily mix design and quality control workflows. This is machine learning infrastructure becoming part of the standard operating procedures for concrete production.
That’s the kind of adoption curve you want to see. Not AI as a science project, but AI as a commodity tool that makes existing workflows faster and cheaper.
Why This Actually Matters
Reshoring and foreign direct investment have brought over 1.1 million jobs back to the U.S. since 2020. Manufacturing has one of the highest economic multipliers: every dollar spent in manufacturing adds $2.69 to the economy. The cement and concrete sector alone contributes more than $130 billion annually and supports roughly 600,000 jobs.
If AI can help domestic concrete producers rapidly reformulate their mixes to use U.S.-made cement instead of imports, that’s not just a technical win. It’s an economic one. Amrize recently announced nearly $1 billion in capital investments for 2026, partly to increase domestic cement production. They’re even launching a “Made in America” cement label guaranteeing U.S. standards, domestic manufacturing, and American materials.
I’m generally skeptical of corporate press releases about AI changing the world, but this one has the receipts. The model is open source on GitHub, there’s a pre-print paper describing the methodology, and Meta released the foundational dataset used to develop the Rosemount mix. They’re presenting at the American Concrete Institute convention with Amrize and UIUC researchers.
Meta’s AI work in concrete won a 2025 Building Innovation Award for Best Partnership and a Slag Cement Award for Sustainable Concrete Project of the Year. The industry is paying attention.
The Unsexy Problems are the Best Problems
This is the kind of AI application I wish we saw more of. Not chatbots or image generators or another venture-funded attempt to disrupt email, but tools that help real industries solve real bottlenecks. Concrete mix design is an incredibly boring problem with massive economic leverage. If you can cut the time and cost of reformulating mixes, you unlock domestic material substitution, faster construction timelines, and more resilient supply chains.
The fact that this is happening through open source tooling makes it even better. Quadrel didn’t need to rebuild everything from scratch. They took Meta’s framework, adapted it to their customers’ needs, and shipped it. That’s how infrastructure should work.
I have no idea if BOxCrete will become the industry standard for concrete mix design, but the early signs are promising. When a Fortune 500 construction materials company is investing billions in domestic production and using your AI model at scale, you’re doing something right.