The cost of compute has fallen by roughly an order of magnitude per decade for over half a century, and the trend continues. AI models that required a data center to train in 2020 can now be run on a laptop. Models that required a laptop in 2024 can now be served from a phone. The same compute that drives drug discovery, scientific simulation, financial analysis, and language processing is on the same falling-cost curve that brought computing into the home over the past forty years.
What is new is what compute now does. AI models trained on vast bodies of human knowledge are making expertise accessible at near-zero marginal cost. A medical question that would have required hours of clinical reading can be answered in seconds. A legal precedent that would have required a paralegal day to find can be retrieved in minutes. A coding problem that would have stopped a non-programmer can be solved in conversation. Expert knowledge, in the form that machines can hold, is becoming something close to free for those with access.
The ownership question here is acute and structural. The largest AI models are trained on data scraped from the public commons — books, articles, code, conversations, the accumulated written work of human civilization — and then sold as proprietary services by a small number of well-capitalized firms. The compute required to train frontier models is concentrated in a handful of corporate data centers. Whether AI-mediated knowledge becomes a commons or remains a private rent is being decided now — in court cases, in licensing agreements, and in the slow accumulation of dependencies on platforms whose terms of service can change at will.