AI Is Burning Through Graphics Cards (www.nelsx.com)

AI workloads are rapidly eating through datacenter GPUs: generational investments have pushed millions of NVIDIA cards into 24/7 inference duty, but these chips were built for bursty rendering, not continuous model execution. Lindahl estimates 3.5–4.5 million NVIDIA data‑center GPUs are in production today, with hyperscalers like Meta, Microsoft and Google each running hundreds of thousands. Continuous inference magnifies wear — failing fans, dried thermal paste, silicon aging and persistent thermal cycling — so many cards show significant degradation within one to three years, a window that often mirrors warranty periods. Even surviving hardware becomes economically obsolete as new architectures quickly double efficiency and throughput, suggesting much of today’s fleet will be retired, resold, or downgraded by 2027–2028. That replacement cycle has massive cost, supply-chain and sustainability implications: refreshing millions of GPUs every few years stresses foundries like TSMC, increases e‑waste, and boosts power demand. The industry is already pivoting toward purpose‑built accelerators (ASICs), FPGAs and other designs optimized for continuous inference with lower thermal and aging profiles; longer-term contenders include photonics, neuromorphic and quantum approaches. The next inflection in AI infrastructure will be driven less by raw peak speed and more by durability, energy efficiency and lifecycle impact as teams seek hardware that can endure relentless inference at scale.