Dwarkesh Patel sees AI progress while ratepayers carry grid costs

By Imani Sutton · Reporting from Atlanta ·

Dwarkesh Patel argues self-improving AI will accelerate technical progress, but scaling data center infrastructure risks pushing energy costs onto public utility ratepayers unless regulators step in.

On his podcast Dwarkesh Patel, host Dwarkesh Patel laid out a spending metric. It should wake up every public service commissioner in America. Top researchers at OpenAI are now spending heavily on automated coding agents, up from just months earlier. On the episode "Recursive self-improvement has started," Patel argued that artificial intelligence is reaching an inflection point. Software rapidly improves itself. Patel predicted that fully automated AI researchers could equal top human experts by 2028. To maintain this pace, Patel noted that AI firms must quadruple their research and development computing power every single year.

The real cost of code loops

Patel’s argument for rapid acceleration rests on a clear observation. Machine learning research is largely a testable, empirical science. Unlike pure theoretical physics, building better AI models relies on running millions of small experimental runs in data centers. These runs show which algorithmic recipes improve learning efficiency. Patel cited internal data from Anthropic. Its Claude model now leads 26 percent of the company’s R&D, up from just 1 percent in February. Evaluations from independent benchmark organization METR suggest Anthropic’s capability gains are doubling year-over-year. In Patel’s view, deploying swarms of automated agents will compress technical progress—a "spark" scenario that accelerates chip design and materials science.

The strongest argument for Patel’s timeline is clear. We have already watched models move from failing primary-school arithmetic to winning gold medals at the International Mathematical Olympiad. Software loops can solve complex math problems like OpenAI’s Millennium Prize resolution. Advocates argue that hardware bottlenecks will be engineered away by smarter algorithms requiring far less computation to train.

The grid carries the ultimate bill

Yet this software-centric vision ignores hard physical limits. It overlooks thermodynamics and utility grids. An algorithm can improve parameter efficiency. However, running continuous, synthetic training environments demands massive, uninterruptible baseload electricity. These environments involve simulated economies where AIs compete across corporate scenarios.

Every line of self-improving code eventually hits a power line. Quadrupling R&D computing annually is not an abstract software metric. It is an industrial land rush for substations, high-voltage lines, and cooling water. Tech giants like Microsoft, Amazon, and Google build multi-gigawatt server campuses for OpenAI and Anthropic. They plug into regional utility grids already struggling under extreme weather. Regulated utilities have a clear historical record. Monopolies rush capital expansion to satisfy large industrial customers. Captive residential ratepayers end up stuck with the bill through higher monthly rates.

Public utility commissions review rate cases and land-use permits for new data centers. Voters and consumer advocates must demand hard protections. Regulators should mandate that tech corporations contract their own off-grid clean generation. They must pay full cost-of-service rates for grid infrastructure. If technology companies want to run automated software loops, their own balance sheets must pay for the grid beneath them. Family utility bills should not carry the load.

Sources

  1. Recursive self-improvement has started