Unique Research · Insight · August 7, 2024 · Essay by Benedict Evans (ex-a16z partner; independent analyst)
Hundreds of millions tried ChatGPT; few became habitual users. Enterprises ran pilots; few put models into core workflows. Part of that is time — but LLMs may also be a trap: they look like products when they are not. Product–market fit still has to be found the hard way.
Adoption always takes longer than the demo
Marc Andreessen’s line: every failed idea of the dot-com bubble can work today — after broadband, PCs, e-commerce infra, ads, and behavior change. U.S. retail took 20+ years to move ~20% online. Early iPhone: 5.4M units in year one; real takeoff ~2010. Cloud still only ~1/3 of enterprise workflows decades after Benioff’s browser pitch — even as Silicon Valley treats cloud as settled.
ChatGPT was faster: ~100M users in ~2 months (late 2022). OpenAI didn’t wait for new devices or carrier builds — a site/app on 25 years of internet infra. But “use” often means curiosity once or twice, or every few weeks — not daily habit.
Pilots ≠ production
Bain-style surveys separate experiments from real deployment: almost everyone is testing; few wire LLMs into core processes. Coding and marketing show clear pull; legal/HR slower. Accenture: ~$300M genAI work across ~300 projects last year (~$1M average) — mostly trials. BCG expects ~20% of 2024 revenue helping large companies “deal with” genAI. Morgan Stanley CIO survey: 30% of large-company CIOs see no product deploy before 2026. Enterprise sales cycles can outlast the time since GPT-3.5 shipped.
The “looks like a product” trap
LLMs look general, look magical, look finished — a chatbot feels like a product. Outputs can be wrong in hard-to-manage ways: you cannot ship raw prompt → raw answer. Bing’s early LLM search push (make “Google dance”) is the cautionary tale; patience may beat forced launch. Capex raced ahead of mostly experimental budgets — skipping the painful bottom of the S-curve where prior platforms learned PMF (App Store wasn’t in the original iPhone plan; Berners-Lee’s first browser was more editor than publishing platform).
Bubble — or anti-bubble?
Startup stack is Silicon Valley’s bet that LLMs are a technology needing classic customer discovery — not one chat product that ends history. If models eventually do everything, many of those startups would be unnecessary — an oddly optimistic “anti-bubble” reading. Dot-com dreams did come true eventually; a J-shaped S-curve is possible. It will not happen this year.