The Age of Personalized Hardware Is Coming

- **GEAstack** is a startup building custom RISC-V chips for AI workloads. Its argument: AI-specific silicon designed from scratch, not repurposed GPU silicon. - For decades, hardware was general-purpose (Intel x86 CPUs, NVIDIA GPUs) because designing custom chips was too expensive for anyone but the largest hyperscalers. But chip design costs are plunging thanks to open ISA RISC-V, chiplets, and new fab accessibility. - The article argues we are entering an era where small teams can design domain-specific hardware cheaply. General-purpose GPUs waste power and die area on features AI doesn't use (e.g., branch predictors, rasterization units). Personalized hardware means silicon tailored precisely to a model's math. - If true, this shifts computing economics: instead of one expensive general chip for all tasks, you get many cheap specialized chips. That makes inference cheaper and faster, enabling AI at the edge. For NVIDIA and AMD, this is an existential threat.