Sequoia Bets Machines Will Do 99.9% Of Thinking And Capture $10 Trillion

Konstantine Buhler, Partner at Sequoia Capital, published on X on last week an essay arguing that machines will soon perform 99.9% of the world's cognitive work, the exact share they took of physical work between 1800 and today. The post passed 468,000 views in 48 hours. At Sequoia's AI Ascent in April, Buhler, Pat Grady and Sonya Huang priced the target at roughly $10 trillion in services revenue that software never reached, against a $650 billion software market. PitchBook counts AI startups raising $407 billion in the first half of 2026, more than the $264 billion for all of 2025, with OpenAI and Anthropic taking $217 billion of it across three rounds. AI absorbed 86% of the $412.7 billion deployed in the U.S. in the same six months, and rounds of $100 million or more took 87.5% of every dollar. Deal count fell to 3,500 from 8,290 in 2025. As an investment thesis, the Cognitive Revolution is a bet on a few very large inputs and applications that do not exist yet. Buhler's frame is echoing what happened with steam, combustion and the electric motor that moved physical work from 99% biological to 99.9% machine over two centuries. He argues neural networks will do the same to cognition, faster, because the cost curve is steeper. He puts intelligence per watt at more than a 10x annual decline. The modern Carnegies in his essay are Nvidia, TSMC, the Transformer authors and hyperscalers spending trillions on data centers. The essay opens with his daily Waymo commute and a claim of 17x fewer serious-injury crashes than human drivers. Waymo's own safety hub reports 92% fewer serious-injury-or-worse crashes over 170.7 million rider-only miles, roughly 12x. An independent Insurance Institute for Highway Safety study published in July found 68% fewer police-reportable crashes per mile, with Austin running 4% higher than human drivers on a small sample. Monthly $12.99 Billed monthly Annual $89 Billed annually Two Year $165 Billed biennially Subscriptions renew automatically. You may cancel your subscription at any time. Where Buhler is strongest is on labor history, and where he is most candid is on speed. He cites Robert Allen's Engels' pause: British output per worker rose 46% between 1780 and 1840 while real wages rose 12%, then wages rose 123% over the next sixty years. The Industrial Revolution gave a farmhand's son forty years to become a factory hand. Buhler writes this revolution may offer five, and that education, retraining and safety nets were built for the old clock. His evidence that the transition is already producing jobs rests on software. Indeed Hiring Lab finds U.S. software development postings rose almost 15% since Claude Code launched in February 2025 while overall postings fell 7%. The same report shows 71% of the gain came from senior roles, 37% from titles mentioning AI, and postings still 27.5% below February 2020. The August snapshot puts the index at 74.4 against a 100 baseline, up from a May 2025 low of 61.1. That is a rebound for engineers who supervise machines and a closed door for those who would have learned by doing the tasks machines now do. Buhler says data on legal and financial services hiring inflections is coming next week; until it lands, software is one data point. Coding agents are the Spinning Jenny, the first skilled craft automated. The automobile of this era, Buhler writes, has not been built, and his candidates are science at machine speed, a personal agent that runs a life the way a chief of staff runs a CEO, and new forms of human coordination. The screening test is the one he posted in July 2025: cover the word AI on the deck and ask whether the business still makes sense. The trillion-dollar framing has been consistent since the 2025 keynote I covered here. What changed in 2026 is the claim that functional AGI has already arrived and that the question is now who captures the services budget. Three key moves follow for investors: 1st: Underwrite revenue against services budgets, where a customer spends $6 on outcomes for every $1 on tools, and discount any application that only makes existing software workers marginally faster. 2nd: Model an Engels' pause in your portfolio's end markets, because a decade of stagnant wages in exposed professions is a political and demand problem. 3rd: Test the 10x annual cost decline against your margin assumptions, since a company whose moat is cheap inference has no moat by next year. For founders the instruction is the exact opposite; build for a world where cognition is free and price the outcome, because Buhler's long tail of un-thought-about problems, the radiologist who gets hours instead of minutes, the small business that gets an analyst, is where demand expands to meet supply. Then plan for the human that remains. The essay's most defensible line is that what stays economically human longest was never cognition: wanting things, choosing between them, being accountable and being trusted.