Eric Schmidt on the Robotics Race, Singularity Timeline, and Energy Shortage | 241
Summary
Schmidt estimates AI has delivered only 10–15% of its eventual impact, even though today’s reasoning systems are already “perfect partners for human beings—for good and bad.” San Francisco’s consensus puts recursive self-improvement two to three years away, but Schmidt stresses that it does not yet exist: limited demonstrations cannot “learn everything, discover things, and tell me what you learned.”
Software development has flipped from assisted coding to autonomous orchestration in a matter of months. The Claude Code release identified onstage as Opus 1.6 moved Bay Area workflows from “80/20” to “20/80”; one programmer now writes a specification and evaluation function at 7 p.m., then lets the system complete overnight what Schmidt says once required six months and 10 Google programmers.
AI should concentrate value at both ends of the company-size spectrum: a few enormous platforms and many tiny teams. Schmidt expects elite programmers—historically worth 10 times the next tier—to become more valuable as directors of parallel agents, while hand-coding becomes “like riding a horse.” The scarce skill shifts from producing code to defining objectives, tests and learning loops.
Electricity is the binding US constraint, with an estimated 92-gigawatt shortfall by 2030. That equals roughly 60 nuclear plants at 1.5 GW each; at about $50 billion of infrastructure per gigawatt, 100 GW implies $5 trillion over five years. Schmidt sees no demand asymptote yet because efficiency triggers Jevons paradox: better hardware and algorithms unlock more uses, computers and power consumption.
Google and Nvidia occupy unusually strong infrastructure positions because they control more of the inference stack. Schmidt says TPU version two’s decade-old design choices made it an ideal inference engine, while Nvidia accomplished what Intel could not by controlling a purchasable “complete server architecture.” Space data centers could offer effectively infinite power, but heat dissipation and radiation remain issues, and Schmidt frames ground-versus-space as a business question involving fiber, launch scale and other trade-offs.
China appears positioned to win low-cost robotic hardware through its electric-vehicle supply chain, motor expertise and “brutal competition.” Schmidt calls China a competitor, “not enemy,” but says allowing it to dominate low-end EVs was an error that America risks repeating in robotics. His boundary matters: predictable battery production can automate rapidly, whereas precision rocket assembly still depends on skilled workers exercising judgment that current robots lack.
The frontier-model market may support roughly 10 capital-intensive competitors, but their architectures and national strategies are diverging. China favors open weights and pervasive edge computing despite US chip restrictions, while America remains centered on centralized AGI and ASI. Schmidt argues that safety must be shaped without slowing the race—though a “Chernobyl-like” event might be what finally forces rival governments to coordinate.
Deep dive
1. Reasoning agents have arrived; recursive AI has not
Schmidt places the world only “10 or 15% into the impacts” of AI, with software moving faster than hardware and robotics. Even if progress stopped today—which he says it will not, and which no government, individual or corporation can stop or control—reasoning agents would already have advanced humanity, with implications for good and bad.
The “San Francisco consensus” says this is the year agents scale across work, followed within two to three years by recursive self-improvement. Its mechanism: a company with 1,000 scarce AI researchers could run perhaps one million research agents, bounded primarily by electricity and judged against explicit evaluation metrics.
Evidence for that acceleration is the Claude Code release identified onstage as Opus 1.6: Bay Area developers report that work “was 80/20, now it’s 20/80.” Schmidt attributes the jump less to context length than to an underlying model that can reason longer and produce better-quality tokens.
His RSI review remains deliberately unresolved: scientists disagree on the necessary approach, and successful lab tests are “limited cases that are kind of demos.” Real recursion would answer, “Start now, learn everything, discover things, and tell me what you learned.” That query does not work yet.
2. Coding is becoming orchestration, not craftsmanship
Schmidt’s best specimen is a UI developer who writes a specification and evaluation function, starts the agent at 7 p.m. and reviews its inventions after 4 a.m. “This stuff would have taken me six months and 10 programmers at Google,” he says; the developer sleeps through it.
Diamandis contrasted that workflow with Davos two months earlier, when AI remained autocomplete that required babysitting. Onstage, he had six concurrent Claude 4.6 jobs running in the cloud and expected them to solve six launched problems before the session ended.
Schmidt sees elite programmers gaining leverage rather than disappearing: the top tier was already worth 10 times the tier below, and now those people can direct parallel systems. The resulting market structure is “a relatively small number of very large companies” alongside many very small companies requiring fewer employees.
His proposal is to make prompt engineering a first-year course for every university student; Diamandis pushes it down to high school. Schmidt accepts the logic but flags age restrictions and vulnerable teenagers, emerging job impacts in software and customer service, and unpredictable orchestration when agents from incompatible vendors interact.
3. Google’s enduring advantage came from architecture and evaluation
Looking back at Google’s transformer, TPU and DeepMind work, Schmidt says, “When you’re making history, you typically don’t know it.” He credits Larry and Sergey’s technical standards—including once calling his proposal to hire Java programmers “the stupidest idea we have ever heard.”
TPU version one was essentially a matrix multiplier; version two changed the algorithm and became particularly effective for inference. Schmidt pairs that with Nvidia’s Rubin architecture, arguing Nvidia achieved what Intel never did: control of the complete server architecture and delivery of a functioning supercomputer.
The host recalls DeepMind’s $600 million acquisition being mocked as a zero-revenue Go bet, then reportedly paying for itself through data-center cooling savings. At the 2016 Go match, Schmidt watched the win estimate move from 50% to 51% to 52%; the team’s answer was, “We just planned for it to get to infinity.”
The Go example’s larger lesson was the need for a defined validation function, not merely a command to write a winning program. The same team then worked on protein folding; Schmidt says they ultimately produced AlphaZero, which essentially self-learns. The host says protein folding reduced a four-year PhD task to roughly an hour, while Schmidt calls Gemini 3 the broadest non-Chinese system across multilingual and multimodal depth.
4. The AI buildout is a power-and-capital supercycle
Schmidt’s congressional estimate is a 92-GW US power shortage by 2030—about 60 nuclear plants at 1.5 GW each, while America is building “essentially zero or one.” Universities, talent and finance are available; electricity is the scarce resource.
One gigawatt corresponds to roughly $50 billion of hardware, software and data centers, putting 100 GW near $5 trillion over five years. Schmidt believes America might finance that “on a wing and a prayer”; data-center construction already represents about 1% of US GDP growth.
A standard new facility is roughly 400 MW, half a mile long and 500 feet wide. It is effectively an airflow machine with internal liquid cooling: chips draw about 2 kilowatts, while HBM3E and HBM4 memory produce enough heat to require water cooling.
Efficiency does not rescue the grid because Jevons paradox converts cheaper computation into new demand. “When does the asymptote arrive?” Schmidt keeps asking; no slowdown is visible. Space data centers could offer infinite power, but radiation remains an issue, and Schmidt says heat dissipation is technically understood. He frames space versus ground—with its fiber and other advantages—as a business question involving launch scale and trade-offs.
5. China has the low-end robotics lead America surrendered in EVs
Schmidt’s geopolitical distinction is emphatic: China is America’s “competitor, not enemy.” It has capital, technical talent, an equal or stronger work ethic and dominance in key industries; shielding Americans from Chinese cars does not erase their competitiveness or the error of losing low-end EV manufacturing.
Robotics inherits the EV supply chain because robots are fundamentally brains plus actuators—the same stepper motors and related systems produced by the electric-vehicle industry. Schmidt therefore sees China winning low-cost robotic hardware, with Unitree’s robot dancing with humans as a visible example.
Diamandis argues that highly automated gigafactories make “the robot building the robot” feel imminent. Schmidt narrows the claim: batteries are predictable and scalable, but rocket assembly still requires exceptional workers who understand tubes, tolerances and defects. “Low-skilled labor of any kind gets swept up”; skilled mechanical judgment may be among the last capabilities replaced.
6. Frontier competition and safety will be shaped together
Using numbers he explicitly calls “made up,” Schmidt thinks the world can support at least 10 frontier companies: a few in China, a US majority, perhaps one or two in Europe despite electricity costs, and maybe one in India. Russia is excluded from his estimate because of the war.
China’s strategy combines open-source, open-weight models—including DCV4, Quinn, Kimi 2 and Kimi version 3—with more edge computing around users. The US pattern is more centralized and AGI/ASI-focused; Schmidt does not expect the two approaches simply to track together.
Microsoft and Google can fund the race from enterprise cash flow; Anthropic has raised heavily from other companies, uses Google TPUs and leads in the Claude API within the enterprise-agent system. OpenAI is adjusting its strategy, but Schmidt’s honest forecast is uncertainty: “In a year, we’ll know better.”
His “Chernobyl-like” warning is descriptive, “not an endorsement”: a hopefully small biological or nuclear attack spawned by these systems might finally compel China and the US to coordinate. Diamandis calls for experts in politics, history, psychology, governance and ethics to help shape alignment; Schmidt responds with faster energy permitting, high-skill immigration and shaping rather than slowing the race, while treating harm to minors as a line that cannot be crossed.