Meta and Amrize say an AI-designed concrete mix went into a Minnesota data center slab

March 30, 2026 — Meta says a concrete mix its AI designed with Amrize, poured at scale in a slab at its Rosemount, Minnesota, data center in 2025, reached full structural strength 43% faster than the original formula and cut cracking risk by nearly 10%. Here's what happened, what it cost, and what went wrong.

Cost and outcome (as of March 30, 2026)

Outcome: The AI-optimized mix was used at scale in a site support section of one data center building slab, and data showing it meets requirements cleared it for other areas of the site, the companies say. (Engineering at Meta, July 16, 2025; Amrize press release, July 16, 2025)
Cost: Not disclosed. Meta's 2026 post says AI design lets suppliers optimize for performance, cost and carbon, but gives no price.
Claims: 35% lower carbon footprint for the concrete, an Amrize estimate. (Amrize press release, July 16, 2025)

What the companies did

Meta built an AI tool using Bayesian optimization, on its open-source BoTorch and Ax frameworks, to design concrete mixes, and developed it with Amrize and the University of Illinois Urbana-Champaign, it said. Professor Nishant Garg's lab generated the training data; Meta says the pipeline has produced data on more than a hundred unique mixes. (Engineering at Meta, July 16, 2025)

Amrize built a pipeline at its batch plant near St. Paul, Minnesota. Data-center slabs have to be flat, level, smooth and durable, so the formulas had extra finish constraints. Meta says that within two iterations, with minor human adjustments, the AI found formulas that beat standard low-carbon industry formulas on strength, speed and sustainability. (Engineering at Meta, July 16, 2025)

The pour

Amrize worked with Mortenson, the general contractor, to test workability and finishability. Successful slab tests led to the at-scale pour, and Meta says formal tests showed the team exceeded all technical requirements. Amrize says the ECOPact mix is "estimated to reduce the total carbon footprint of the concrete by 35%." (Engineering at Meta, July 16, 2025; Amrize press release, July 16, 2025)

"AI-driven mix design lets us optimize concrete for performance, cost and carbon in one step," said Garg, quoted by Amrize. Jaime Hill, president of Amrize Building Materials, said the work with Meta "is just the beginning." Julius Kusuma, a Meta research scientist, said the company would keep "testing different ways we can improve our material design in future projects with Amrize." (Amrize press release, July 16, 2025)

The March 2026 update

In March 2026 Meta released a new model, BOxCrete, and the foundational data behind the Rosemount mix. It said the mix, using domestically sourced materials, reached full structural strength 43% faster than the original formula and cut cracking risk by nearly 10%. Meta framed the update around domestic supply, saying roughly 20-25% of U.S. cement consumption is met by imports. It also said Pennsylvania-based Quadrel has built its framework into ready-mix software. (Engineering at Meta, March 30, 2026)

Meta's own strength-explorer tool warns that its predictions reflect controlled lab conditions and "may differ from field performance." (Engineering at Meta, March 30, 2026)

What went wrong

Slower curing is the usual catch. Meta says low-carbon formulas typically face slower curing, surface-quality issues and supply-chain complications, and its own example curve shows the greener mix starting weaker. (Engineering at Meta, July 16, 2025)
Humans still in the loop. Experts reviewed each stage, refined constraints and adjusted binder amounts and water-to-binder ratios, Meta says. (Engineering at Meta, July 16, 2025)
Lab versus field. Meta warns that its model's predictions may differ from field performance because of batching variability, curing environment and aggregate properties. (Engineering at Meta, March 30, 2026)
What we don't know yet: the measured, third-party-verified carbon and strength results for the poured slab, and the cost compared with Amrize's standard mix.

Our follow-up question for Amrize

What were the measured strength and embodied-carbon results for the Rosemount slab, and what did the mix cost per cubic yard against the standard mix?

Sources

Amrize, “Amrize and Meta Partner on First-of-its-kind AI-Optimized Advanced Concrete Mix for Data Center in Minnesota,” July 16, 2025. amrize.com
Engineering at Meta, “Using AI to make lower-carbon, faster-curing concrete,” July 16, 2025. engineering.fb.com
Engineering at Meta, “AI for American-Produced Cement and Concrete,” March 30, 2026. engineering.fb.com

Disclosure: [not in copy] Editor placeholder