Meta's AI Ambitions Drive Unprecedented 14 GW-Scale Infrastructure Push, Fueling Power Demand and Grid Pressures

A leaked internal Meta memo, as reported by Reuters via ZeroHedge, outlines plans to double AI computing capacity from 7 GW to 14 GW by 2027, with 2026 capital expenditures reaching as high as $145 billion. This aligns with Meta's publicly raised guidance for 2026 capex in the $125-145 billion range, explicitly tied to AI infrastructure expansion including data centers and accelerators.

Meta's in-house MTIA (Meta Training and Inference Accelerator) program, including the Iris generation, is advancing with production at TSMC and a rapid cadence of new chips every six months through 2027, supplementing external GPUs like those from AMD. Official Meta announcements detail multiple MTIA generations deployed or slated for 2026-2027, emphasizing cost reductions and workload coverage beyond ranking/recommendation systems.

The energy implications are substantial. Meta has secured deals for multi-GW power, including nuclear agreements potentially unlocking up to 6.6 GW by 2035 via TerraPower, Vistra, and Oklo partnerships, alongside gas-fired and solar projects for campuses like Hyperion (targeting 2-5 GW in Louisiana) and Prometheus (over 1 GW in Ohio). These efforts underscore AI data centers' voracious power needs—equivalent to hundreds of thousands of homes per GW—straining grids and prompting corporate nuclear investments previously rare at this scale.

Additionally, Meta's Meta Compute initiative, reported by Bloomberg, aims to monetize surplus capacity through cloud services, reflecting both aggressive buildout and potential overcapacity concerns. Long-term supplier contracts (e.g., memory and components) amid shortages further signal sustained commitment.

While the exact 14 GW figure remains internal, corroborated elements reveal AI scaling's hidden infrastructure costs: power procurement, supply chain lock-ins, and capex discipline debates amid market reactions to spending forecasts.