Unique Research · Insight / a16z economics essay · October 11, 2023 · Full editorial rewrite of ~4,698 CJK — economic argument, numbers, and “third age” framing preserved
Martin Casado (a16z GP; Nicira co-founder/CTO, sold to VMware for $1.26B in 2012) — The Economic Case for Generative AI.
Core frame: We are entering compute’s third age — chips drove compute’s marginal cost toward zero; the internet drove distribution’s; generative AI drives creation’s.
AI economics — why no prior platform shift?
Casado recalls 2003 PhD classmates (half in AI, Bayes fashion) then “AI is dead” within three years. Seventy years of progress solved real problems — expert systems, chess, AV, vision — then the bar moved (“that wasn’t real AI”). On entity recognition AI beat humans for ~a decade. Meta, Google, Netflix used AI to expand market caps — yet unlike mobile, internet, and chips, incumbents captured value and AI-native startups did not displace them.
Part one of the thesis: capabilities existed; startup economics did not.
The AI mediocrity spiral
Standalone AI economics (ex-bigco) looked mediocre. Sexy demos were niche markets — beating Russians at chess is not a market. Worse: many classic AI uses demand correctness in a long, fat tail. Startups either build all tech or hire people — variable cost that scales with volume. Complex domains (AV) can show a perverse economy of scale: unit costs rise as the industry grows. Many AI startups ended with non-software margins — hard to scale. Robotics adds the hardware curse.
Competitor is often the human brain — ~100M years of evolution for perception (berries, lions). Investors know the AI mediocrity spiral: founder wants automation → correctness matters → hire humans first so it “looks good” → raise → board pushes growth → “AI is hard, hire more” → treadmill; automation never arrives or is only a slice. Bigcos can absorb perverse scale. Robotaxi: decades and ~$75B invested; unit economics still lose to a human Uber driver. Even the AV compute stack can draw ~1.3 kW vs ~15 W for a human brain for maybe ~10× performance — not a 10,000× economic leap that creates market transitions.
Lesson of 70 years: tech works; problems get solved; bigcos make money — startups struggle to break out; without startups, no transformation.
Generative AI economics — why this wave differs
One model takes natural language and outputs dialogue, images, anything. Productive businesses already appear in the 3Cs: Creativity, Companionship, Co-pilot.
Prior anti-startup properties fail here. Markets are large (knowledge work; games/film ~$300B). Correctness is often soft for creativity — what does “wrong” mean for a novel or game? (Keep the fingers; even sci-fi may not care.) And behavior is iterative: the person who once did the work inside the company becomes the user — cost shifts off the payroll; code can be wrong in parts because the user revises. Mostly software for now (robots not the takeoff).
Silicon vs carbon for creative-language centers (~50k years of evolution) is far more competitive than fighting a 100M-year perception brain. Example: Pixar-style self-image — model inference ~$0.001 in ~1 second vs ~$100 for a graphic artist for an hour → 4–5 orders of magnitude on cost and time. Instabase / legal brief PDF into an unstructured-doc LLM: same ~$0.001 / ~1s vs hours of lawyer time that still needs cross-check. Whole games (3D, characters, VO, music, story) — hundreds of millions of dollars and years vs dollars of inference — asymmetric like internet and chips. Not a promise of timing; a path signature of big transitions.
Hence fastest open-source projects and some of the fastest-growing companies in industry history — not only capability, but economics when marginal cost collapses.
The third age of compute
Age 1: microchips. Before computers, people were computers (log tables). ENIAC ~5,000× a human → 3–4 orders of magnitude → IBM-class reinvention and new firms. Chips → compute MC → $0. Age 2: internet → distribution MC → $0 (box games in weeks vs bits over DSL — another ~4–5 orders). Age 3: large models → creation MC → $0 — fuzzy as “creation” is, as prior ages were. Nobody predicted Amazon or Yahoo; prepare for a wave of iconic companies whose shapes we do not yet know — because the economics are that compelling.
Jevons paradox & expanding productivity
When markets scramble, jobs questions arise. Economic answer: Jevons paradox (1865: more efficient coal → more coal use when demand is elastic). Internet efficiency cut prices and raised demand and productivity. Casado’s view: creative assets and automatable work have elastic demand — more creation → more consumption → large productivity expansion, new jobs, new things — as with chips and the internet.