The AI that we'll have after AI (pluralistic.net)

Cory Doctorow argues that the current AI boom is a finance-driven bubble: companies have spent hundreds of billions (approaching a trillion) on GPU-heavy data centers while the industry only earns an estimated ~$45B/year. Those purchases are often showmanship for investors, not product economics — firms claim multi‑year GPU lifespans while real duty cycles look closer to 2–3 years (and under extreme load some cards “burn out” in as little as 54 days). To recoup investments at that scale would require roughly $2 trillion in revenue, a gap that makes the current model unsustainable and risks a broad market crash that will ripple into the research and engineering workforce. What survives will matter for practitioners: cheap GPUs and hardware sold off in bankruptcies, experienced applied statisticians and engineers looking for work, and a thriving open‑source model ecosystem that’s already proving efficient and capable. Doctorow highlights examples where optimization wins — Chinese models like Deepseek use far less compute and still impressed markets, and hardware demos (e.g., Pete Warden’s chatbot on a low‑single‑digit‑dollar Synaptics SoC) show voice assistants and inference can run locally with tiny power and privacy benefits. The technical implication for the AI/ML community is clear: efficiency, model compression, and on‑device deployment will drive the next phase, not ever‑bigger foundation models tied to uneconomical GPU fleets.