Ex-Google CEO: What Artificial Superintelligence Will Actually Look Like w/ Eric Schmidt & Dave B
Summary
Eric Schmidt’s core infrastructure call is that AI is underhyped because a learning machine in a network-effect business accelerates until it hits electricity, “not chips.” He estimates the US needs another 92 GW—roughly 92 large nuclear plants—while almost none are starting and a proposed 300 MW small modular reactor would not arrive before 2030. Even better chips will be consumed by costlier reasoning, planning, and test-time compute: “Grove giveth and Gates take it away.”
The capability timetable is aggressive even after Schmidt stretches the “San Francisco consensus” by 1.5 to 2 times. He expects world-class AI mathematicians within one year, world-class programmers within one or two, specialized savants across every field within five—“pretty much in the bag”—and digital superintelligence within 10. If generally available and safe, it means “the sum of Einstein and Leonardo da Vinci in the equivalent of your pocket.”
Enterprise AI threatens the connective tissue of software before it eliminates every programmer. Schmidt says Model Context Protocol can connect an enterprise’s databases to a model that writes much of the necessary code, putting roughly 100,000 middleware and enterprise-software companies under pressure; junior programming work goes first, while senior engineers remain necessary for oversight—for now. The near-term opportunity is complete workflow refactoring, from call centers to dynamically generated interfaces.
China is much closer than Schmidt previously believed, making energy, algorithms, and open weights as important as chip controls. After Gemini 2.5 Pro topped intelligence leaderboards, Schmidt says DeepSeek moved slightly ahead a week later using hardware available in China, distillation, and Huawei Ascend chips among other resources. “A year ago, I said they were two years behind. I was clearly wrong”; with enough money and power, China is “in the game.”
The central security fork is whether frontier intelligence remains concentrated in guardable, multigigawatt facilities or proliferates onto small servers. A ten-model world could be monitored and partly nationalized, but trained weights may run on four or eight GPUs, while quantization, distillation, or a 100-fold inference improvement can magnify capability after export. Schmidt’s proposed response is “mutual AI malfunction”: reciprocal cyber capacity, tracked chips and training runs, and deterrence before either side crosses a sovereignty-threatening line.
For startups, Schmidt distinguishes hardware from software: patents, inventions, power systems, and robotics can form slower deep-tech moats, while software’s durable moat is learning velocity rather than brand. In markets with rapid feedback, a product that learns from every click, trade, or sensor reading can become “essentially unstoppable”; because the slopes are exponential, a competitor only a few months behind may still lose. He expects perhaps another 10 Google- or Meta-scale consumer companies built on such loops, while slow-feedback government and education markets remain structurally harder.
Schmidt rejects near-term forecasts of net job collapse, although he expects painful displacement and immediate pressure on routine white-collar work. His five-to-10-year case rests on adoption lag, automation raising output and wages, AI assistants enabling retraining, and shrinking workforces—South Korea at 0.7 children per two parents, China at one, and India around 2.0. Dave Blundin’s sharper warning is temporal: workers whose skills may be automated within two or three years need to retrain now, because “wait and see” transfers the cost to the employee.
Peter and Schmidt frame the deepest risk as less a Terminator event than the erosion of agency, attention, and judgment by systems that know how to persuade each person. AI can lower creative costs and expand prosperity—Peter cites estimates of 20% to 30% year-over-year economic growth—but it can also produce misinformation engines, emotionally compelling companions, and “a form of virtual prison.” Their counter to the fear of purposeless abundance is that challenges will migrate: “This notion that we’re all going to be sitting around doing poetry is not happening.”
Deep dive
Not yet available upstream; scheduled sync will retry.