This week on Tom's Hardware Premium: October 3, 2026

OpenAI's push into custom ASICs signals a structural shift: AI compute is migrating from general-purpose rental to dedicated ownership. The technical ripple is immediate—once inference workloads are frozen into silicon, CUDA's software moat loses its grip, and the upstream bottleneck shifts to CoWoS advanced packaging and HBM3E supply, whose demand profile changes with inference chips' lower-bandwidth, higher-capacity power signature. On compliance, the design-in-US, fabricate-in-Taiwan, China-and-Korea memory chain remains the single point of failure. The A100 export ban already stress-tested this; locking an ASIC to one foundry bakes geopolitical risk premium directly into the BOM. Strategically, NVIDIA's likely counter is twofold: Rubin-generation inference optimization to defend the training side, and a CUDA-on-ASIC licensing model converting chip sales into architecture licensing. Google's TPU took eight years for internal substitution; OpenAI's window is shorter but its path more aggressive. Twelve-to-twenty-four-month outlook: by 2027, at least three frontier labs will complete ASIC production cutover, and NVIDIA's inference share will slide from roughly 85% to below 60%. The real long-tail effect isn't the chip—it's the EDA and IP-core market. Synopsys and Cadence will compete to define the new standard for AI-dedicated architecture design, reshaping a $30B toolchain landscape.