That is pretty much what happened to the mainframe. It didn’t disappear. A lot of ordinary computing simply moved someplace else, leaving the mainframe with the jobs that still made sense there.
Before we build power plants and transmission systems intended to last half a century, we ought to ask whether we are wiring the country for the mature form of artificial intelligence or for the architecture it happens to have right now.
The mainframes of 2030 may be busier than ever. They just may not need to answer every email.
Bibliography
1. OpenAI. “Building AI Infrastructure with the Effingham County Community.” July 22, 2026. Describes Project Camellia’s planned 3.2-gigawatt electrical supply and OpenAI’s commitment to pay the infrastructure and service costs associated with it.
2. U.S. Energy Information Administration. “Electricity Use in Homes.” Provides average annual U.S. household electricity consumption used for comparison with Project Camellia.
3. Computer History Museum. Materials on Digital Equipment Corporation, Data General and the minicomputer revolution. The museum describes DEC’s emphasis on interactive, affordable computing and documents Data General’s 1968 introduction of the Nova minicomputer.
4. NVIDIA and Google Cloud. Documentation on NVIDIA’s H100-class AI accelerators and Google Tensor Processing Units. These sources describe purpose-built AI processors and the high-speed memory and interconnects used to combine large numbers of them into training systems.
5. Stanford Institute for Human-Centered Artificial Intelligence. The 2025 AI Index Report. Reports a greater-than-280-fold fall in GPT-3.5-level inference cost, approximately 43 percent annual growth in machine-learning hardware performance, 30 percent annual improvement in price-performance and 40 percent annual improvement in energy efficiency. Epoch AI separately finds that leading AI hardware energy efficiency has roughly doubled every two years.
6. Apple Machine Learning Research. “Introducing the Third Generation of Apple’s Foundation Models.” June 8, 2026. Describes Apple’s sparse on-device architecture and larger cloud models.
7. Microsoft. Fiscal Year 2026 earnings materials. Reports more than 1.6 billion active Windows devices and describes Windows as a platform for increasingly “unmetered intelligence at the edge.”
8. Meta. “Open Source AI Is the Path Forward.” July 23, 2024. Describes Llama 3.1’s 405-billion-, 70-billion- and eight-billion-parameter versions and the use of large models for distillation and synthetic-data generation.
9. International Energy Agency. “Key Questions on Energy and AI.” April 2026. Projects global data-center electricity consumption rising from approximately 485 TWh in 2025 to 950 TWh in 2030, while consumption at AI-focused data centers triples.
10. Microsoft. Fiscal Year 2026 Third Quarter Earnings Conference Call. Discusses the concentration of capital spending in CPUs, GPUs and other relatively short-lived equipment.
11. Microsoft. Fiscal Year 2026 Fourth Quarter Earnings Conference Call. Reports the extension of estimated useful lives for data-center and office buildings from 15 years to 25 years.
12. Federal Energy Regulatory Commission. Large Load Show Cause Orders and Commissioner Rosner’s remarks, June 18, 2026. FERC is examining cost allocation for major new loads and protections intended to keep ordinary customers from absorbing infrastructure costs when planned data-center demand fails to materialize.