AI Chip History Not Only Rhymes but Also Repeat Itself (diblante.com)

The piece argues that the current AI chip landscape is a historical rerun: after the 1980s U.S.–Japan chip rivalry, we’re now in a geopolitically charged U.S.–China competition, but at a far larger and faster scale. Rather than a single dominant vendor, a new wave of companies — from hyperscalers expanding TPU availability to startups like NextSilicon and incumbents eyeing new ARM-based APUs on TSMC 3nm — are designing bespoke accelerators for training and inference. The result is more competitors, deeper specialization, and rapid commercial deployments that extend far beyond Nvidia’s previous dominance. For AI/ML practitioners and investors this matters technically and strategically. Custom silicon and advanced process nodes promise higher throughput and efficiency for specific model classes, but they also increase hardware-software co‑design complexity, risk fragmentation of toolchains and runtimes, and magnify supply‑chain and geopolitical fragility. The environment rewards “paranoid” attention to architecture, portability, and vendor diversity: teams must balance squeezing performance from specialized chips with the need for portable stacks and resilient sourcing. For investors, history suggests studying past cycles and focusing on durable advantages in design, software ecosystems, and manufacturing partnerships.