Goldman Sachs sees AI capex topping $500B by 2026

Goldman’s projection that hyperscalers will pour over half a trillion dollars into AI and data-center capex in 2026 is a blunt measurement of how seriously industry takes the race to build ever larger models and agents. Spending at that level effectively makes AI infrastructure a macroeconomic force on par with national energy systems or telecoms buildouts. It also implies that the next two years will be defined by an arms race for GPUs, custom accelerators, and power—precisely the ingredients that determine who can train frontier‑scale models.

For AGI-watchers, this forecast reinforces that we’re not in a pause; we’re in a scale-up. Even if per-dollar model quality gains diminish, a 30–50% jump in capex year‑over‑year buys a lot of experiments in agents, world models, and alignment techniques. At the same time, Goldman notes that investors are becoming more discriminating, rewarding companies that can clearly translate AI capex into revenue. That pressure may push big platforms to monetize aggressively—through agents embedded in productivity, search, and vertical SaaS—which will accelerate deployment even if headline model releases slow.