Meta’s AI push hinges on cheaper, faster infrastructure, not just a new model

Yahoo Finance notes Meta’s stock rally is powered by a quiet but transformative shift in its infrastructure economics: Reuters revealed an internal memo showing Meta plans about 14 gigawatts of compute capacity for 2026–27, with actual costs per gigawatt around $22 billion (far below many analysts’ estimates). This suggests Meta can expand AI capacity more cheaply than feared, supported by plans to manufacture in‑house Iris chips starting in fall 2026 and a cadence of new chips every six months through 2027, pointing to a highly efficient, vertically integrated AI factory and potentially strong ROI on its big capex beyond the Muse Spark model buzz.