The San Francisco Consensus is a term coined by Eric Schmidt in "The San Francisco Consensus" (Digitalist Papers Vol. 2, December 2025) for a three-part set of beliefs Schmidt describes as shared among most leading AI developers in Silicon Valley. Schmidt presents it as a consensus position — broadly held but not necessarily true — drawing an analogy to prior economic consensus periods such as the postwar Keynesian consensus and the neoliberal Washington Consensus.
The three beliefs
As stated by Schmidt, the consensus comprises:
- Scaling laws will drive continued rapid progress — larger models trained on more data with more compute will reliably yield better performance (see Scaling Laws).
- Timelines to superintelligence are short — Schmidt writes that many experts now see 2–5 years to superintelligence (see AGI Timelines).
- Transformative AI will deliver unprecedented benefits — exponential scientific, financial, and human progress (see Compressed 21st Century, Machines of Loving Grace).
Schmidt characterizes a consensus as a function of both observable truths and underlying ideological commitments, and argues that its existence does not imply its truth. He notes that Silicon Valley has been wrong before, citing the technology industry's optimism about the social impacts of the internet through the 2000s, a vision he says was undercut by disinformation, weaponization, and mental-health crises.
Three revolutions framework
Schmidt decomposes AI progress into three axes, each on a different timeline:
- Language revolution — already occurred. Computers can make sense of, replicate, and interact with human language.
- Agentic revolution — currently underway. AI becomes an actor rather than only a tool, with interconnected AI systems managing entire workflows. Schmidt frames this consistently with Mustafa Suleyman's account.
- Reasoning revolution — described by Schmidt as the most consequential and most speculative. Scaling up existing architectures yields qualitatively new reasoning capabilities, which Schmidt frames as unleashing AGI as a property of scale.
Constraints on the consensus
Schmidt outlines several ways the consensus could fail, grouped by the resource or assumption involved.
On hardware and energy, he writes that chip progress continues (citing Blackwell and Rubin) but that energy is harder to scale, estimating that the US may need the equivalent of 92 additional nuclear power plants. He argues that political and logistical constraints may either force rapid innovation, such as small modular reactors (SMRs) and fusion, or throttle the trajectory.
On data, he notes that pre-training has absorbed most of the public internet, so future progress depends on synthetic data, multi-agent systems, and potentially embodied learning — AI acquiring tacit knowledge by moving through the world.
On algorithms, he allows that LLMs may not be the final architecture, citing Yann LeCun's view that LLMs are fundamentally incapable of creative invention and Fei-Fei Li's experimentation with non-LLM approaches.
Dissenters
Schmidt explicitly names where the consensus is not held:
- Europe, which he describes as more skeptical about social impacts and the pace of progress.
- China, which he describes as much less preoccupied with AGI and more focused on deploying AI for practical applications across sectors.
- The Peter Thiel tradition of broader tech-progress pessimism, which Schmidt summarizes with Thiel's framing: "will AI be flying cars or a 140-character Emily-Dickinson recipe?"
- AI skeptics, including LeCun and Narayanan and Kapoor in their book.
Relation to race dynamics
Schmidt argues that if the consensus holds, the next 5 years are "the most important of the next 1,000," because recursive self-improvement could lock in an enduring advantage for the first group to reach AGI. This framing intensifies race dynamics and motivates calls for US–China cooperation, a theme developed in several Digitalist Papers essays (Graylin's Beyond Rivalry essay, the Abraham/Kavner/Moon/Matheny game-theory essay, and the Beyond Rivalry framework).
Empirical tests
Subsequent results have been read both as reinforcing and as contesting the consensus.
Read as reinforcing it: Opus 4.7, whose capabilities continued to scale and ranked first on GDPval-AA; GPT-5.4 and GPT-5.3 Codex, on a similar trajectory; and Claude Mythos, which Anthropic described as a substantial step in cyber capability (Source: nytimes.com).
Read as contesting it: a 2025 empirical check grading AI 2027's predictions returned mixed results (Source: blog.aifutures.org); MIT NANDA — The GenAI Divide (State of AI in Business 2025) reported that 95% of enterprise AI pilots fail, suggesting value creation is harder than the consensus implies; The Least Understood Driver of AI Progress argues algorithmic progress is narrower than a scaling-law-only framing; and AI as Normal Technology argues that diffusion will be slow rather than rapid.
Schmidt's own position
Schmidt treats the consensus sympathetically but not dogmatically, writing as a meta-observer whose investment decisions depend on the consensus being broadly right even if specific predictions are wrong. His Vol. 1 Democracy 2.0 essay implicitly assumes the consensus, arguing how democracy should evolve if transformative AI arrives on a short timeline.
Relationships
- supports: The Digitalist Papers (Stanford, Volumes 1–2) — Schmidt's source essay.
- supports: Scaling Laws / AGI Timelines / Law of Accelerating Returns (LOAR) — components.
- supports: Situational Awareness: The Decade Ahead / The Intelligence Age (Altman) / Abundant Intelligence (Altman) / Machines of Loving Grace — representative consensus holders.
- related: AI Race Dynamics / Fast-Follow Problem / AI 2027 — downstream implications.
- contradicts: AI as Normal Technology / AI Snake Oil — Narayanan and Kapoor (2024) / Why I Think AI Take-Off Is Relatively Slow (Cowen) / (Source: blog.aifutures.org) — explicit dissenters.
- related: Eric Schmidt — entity page for Schmidt.