Meta invests $10 billion in a new Indiana AI data center campus, featuring custom MTIA silicon and liquid cooling at massive scale.
What a Campus-Scale AI Build Looks Like
Meta’s $10 billion commitment to an AI data center campus in Indiana is less about a single hall full of racks and more about treating compute as a multi-building system. A campus design groups power delivery, cooling plants, networking, and operations so capacity can grow in stages without redesigning the whole site each time. For AI training and large inference workloads, that layout matters: clusters need dense interconnect, stable power, and physical room to expand as model size and job parallelism increase.
Choosing a large Midwestern site also reflects practical constraints—land, transmission access, and the ability to stage construction—rather than putting every facility next to existing coastal hubs. The useful takeaway for teams planning their own capacity is to separate “how many GPUs we want next quarter” from “what power, cooling, and network topology we can still support three expansions later.” Campus thinking forces those questions early.
Custom MTIA Silicon and the Case for Homegrown Accelerators
The campus is expected to run Meta’s custom MTIA silicon alongside the usual mix of general-purpose accelerators that hyperscalers still rely on. Purpose-built chips aim at the shapes of work Meta actually runs: training and serving models at high utilization, with less emphasis on being a universal GPU replacement. That can improve performance per watt and simplify software stacks when the hardware and compilers are co-designed with the models and frameworks in use.
The tradeoff is real. Custom silicon needs a long design cycle, specialized tooling, and careful capacity planning so the fleet is not stranded if model architectures shift. Organizations without Meta’s volume should not copy the chip strategy wholesale; they should copy the discipline: measure which kernels dominate cost and latency, then decide whether off-the-shelf hardware, cloud instances, or a narrower custom path fits the risk. For most teams, the practical step is tighter profiling of training and inference jobs before any hardware lock-in.
Liquid Cooling at Massive Scale
Dense AI racks push air cooling toward its limits. Liquid cooling—whether cold plates, rear-door heat exchangers, or facility-level loops—moves heat more efficiently so racks can sit closer and power density can rise without runaway fan energy or hot spots. At campus scale, that means designing the mechanical plant, water or coolant loops, and maintenance procedures as first-class systems, not afterthoughts bolted onto an air-cooled design.
- Plan rack and row layouts for coolant distribution, leak detection, and service access from day one.
- Treat cooling redundancy like power redundancy: failures cascade into thermal throttling and job loss.
- Budget operational skill: liquid systems need different procedures, parts inventory, and vendor support than traditional CRACs alone.
Even teams leasing cloud capacity benefit from understanding this shift. Higher density and liquid-cooled regions change instance availability, thermal limits, and sometimes pricing. Knowing why a provider favors certain form factors helps you place multi-node jobs where interconnect and cooling headroom actually exist.
How Builders and Platform Teams Should Respond
You do not need a $10 billion campus to apply the same design rules. Map your power and thermal budget per rack before you order more accelerators. Prefer software that can schedule across heterogeneous devices so you are not stuck if one accelerator line is constrained. Keep model training and serving pipelines portable enough that moving between cloud regions, on-prem pods, or mixed silicon stays a configuration change rather than a rewrite.
Meta’s Indiana project is a clear signal that frontier AI capacity is being built as long-lived industrial infrastructure: custom chips, liquid cooling, and multi-building sites sized in the billions. The durable response is operational: design for density, plan cooling and power as co-equal with compute, and keep your stack flexible enough to use whatever silicon and facilities are available next.