Meta targets 6.5 gigawatts of AI compute capacity by 2026 following infrastructure efficiency breakthroughs.
Scaling to 6.5 GW of compute capacity by 2026 requires massive data center infrastructure, but achieving this with higher-than-expected efficiency is the critical signal. If Meta has optimized power-usage effectiveness (PUE) or cooling bottlenecks at this scale, it fundamentally lowers the CapEx ceiling for training next-generation foundation models. Competitors will now be forced to either match this infrastructure efficiency or burn excess capital on raw power provisioning.
An internal Meta memo reviewed by Reuters and highlighted by Bank of America reveals that the company expects to deploy approximately 6.5 gigawatts (GW) of AI compute capacity in 2026, with 5.5 GW coming online in the second half of the year alone. More importantly, the memo indicates Meta has found ways to expand this capacity more efficiently than previously estimated, prompting a surge in stock price and an $835 price target from BofA.
Technical Implications To put 6.5 GW into perspective, a standard hyperscale data center consumes between 50 to 100 megawatts (MW). Meta is effectively planning to power the equivalent of 65 to 130 massive data centers dedicated strictly to AI compute within a single year. The mention of "efficient expansion" strongly implies breakthroughs in physical infrastructure and power utilization. This likely involves a combination of higher rack densities, advanced direct-to-chip liquid cooling, optimized power delivery networks, and potentially better integration of their custom MTIA silicon to maximize performance-per-watt.
Why It Matters From an engineering standpoint, power generation and grid interconnects have replaced GPU availability as the primary bottleneck for scaling AI. If Meta has genuinely optimized its Power Usage Effectiveness (PUE) or found novel ways to cluster compute with lower energy overhead, they fundamentally alter the CapEx math for training next-generation foundation models (like Llama 4 and beyond). This structural advantage means Meta can train larger models faster and cheaper than competitors who are brute-forcing their scaling and bleeding capital on inefficient power provisioning.
What to Watch Next Monitor Meta's upcoming quarterly CapEx guidance to see how this efficiency translates to their balance sheet. On the infrastructure side, look for downstream supply chain signals from liquid cooling vendors, power supply unit (PSU) manufacturers, and optical interconnect providers. Finally, deploying 5.5 GW in H2 2026 will put immense strain on local utility grids; watch for Meta to announce strategic energy partnerships, potentially involving nuclear or large-scale renewables, to guarantee this massive power baseline.