The global race for artificial intelligence supremacy is increasingly being measured not just in algorithms, but in the sheer scale of capital expenditure. As hyperscalers and sovereign nations scramble to secure the hardware necessary for the generative AI era, a critical question looms: when will the industry’s annual capital expenditure (Capex) breach the trillion-dollar mark? To answer this, analysts are looking at the only entity that sits at the center of the entire ecosystem—Taiwan Semiconductor Manufacturing Company (TSMC). As the primary foundry for Nvidia, AMD, and Google, TSMC’s revenue stream serves as the ultimate leading indicator for the global hardware build-out.
Currently, the market is absorbing Nvidia’s Blackwell architecture, specifically the GB200 NVL72 racks. For TSMC, a single GB200 rack represents roughly $170,000 in revenue, with the foundry capturing value through both front-end wafer fabrication and advanced CoWoS (Chip on Wafer on Substrate) packaging. As the industry transitions to the Rubin architecture later this year and into 2025, the revenue per rack for TSMC is expected to surge to approximately $260,000. This increase is driven largely by the expanding physical area of these chips and the necessity for more complex packaging solutions to manage heat and interconnect density.
To derive a global Capex forecast from TSMC’s earnings, economists use a multiplier effect. Every dollar of revenue earned by TSMC on an AI GPU cascades through the supply chain: first through Nvidia’s high-margin sales to system integrators, and finally into the total cost of ownership for a data center, which includes switches, cooling, and power infrastructure. Research suggests a multiplier of roughly 18 to 24 for Nvidia-based hardware. When these variables are aggregated, the trajectory becomes clear: global AI Capex is projected to reach $544 billion in 2025 and $794 billion in 2026, before finally crossing the $1.1 trillion threshold in 2027.
This growth is underpinned by a significant lead-time advantage in foundry data. Because the fabrication and assembly cycle for high-end GPUs takes several months, TSMC’s current production schedules provide a window into the state of the market three quarters ahead. Based on TSMC’s own guidance of a 55% to 60% compound annual growth rate for its AI segment through 2029, the industry is looking at a world where AI-related capital spending could approach $2 trillion by the end of the decade. This represents a fundamental shift in the global economy, where compute is increasingly treated as a primary utility, requiring infrastructure investment that rivals the historical build-out of the power grid or the interstate highway system.