人工智能吞噬世界

May 2026www.ben-evans.com AI eats the world Capital Tech moves in platforms shifts Every 10-15 years, a platform shift reshapes technology What happens in a platform shift? Who is affected, and how much? Platform shifts reset the tech industry Microsoft dominated the PC era, but then smartphones made it irrelevant “The very worst case wouldbe that we have just pre-built for a couple of years” “The risk of under-investingis significantly greater thanthe risk of over-investing” Mark Zuckerberg A capex explosion $700bn planned in 2026 for the big four (for comparison, global Telecoms is ~$300bn and Oil & Gas is ~$1tr) (And can’t get TSMC to increase capacity fast enough) A new semiconductor investment cycle? Unprecedented surge in chip demand (but semis have always been a cyclical industry) Driving yet more investment Semiconductor capex is surging to meet (though not as much as the customers might like) Overtaking the office cycle US data centre construction spending (not including the compute!) is now overtaking office construction Challenging financial gravity These were asset-light businesses that funded capex from free cashflow - not any more Here comes the ‘structure’ As spending surges, companies scramble to manage the balance sheet Benedict Evans–– May 202614 OpenAI tries to join the club Deals to build 30GW+ of capacity at $1.4tr? Or$600bn by 2030? at $20bn/GW = ~$1tr annually? and a lot of plate-spinning Deployment hits bottlenecks everywhere Broader supply chains cannot keep up with sudden demand for capital deployment “It’s been almost impossible to buildcapacity fast enough since ChatGPTlaunched” Kevin Scott, Microsoft CTO Meanwhile - unprecedented growth Explosive growth, with demand far ahead of supply (though business models are nowhere near equilibrium) “Wait, how much did wespend?!” Price, use and capex are out of equilibrium How long does usage growth outpace efficiency gains? How much do we spend to chase the frontier? History doesn’t repeat, but it rhymes Press reports: “Usage surge from a new application overwhelms infrastructure and drives new pricing models” Mobile networks in 2010 = LLMs in 2026? How do you manage explosive demand with marginal cost and limited capacity? And then what? "We see a future where intelligence is autility like electricity or water and peoplebuy it from us on a meter” Sam Altman Commodity infra rarely captures value up the stack Mobile networks are a trillion dollar industry, but all the use-cases and value-capture are built by other people And so far, models do seem to be commodities For most general and consumer use, models are very similar, and crucially there are no network effects What will equilibrium look like? What will you get if you win the capex war? A provisional thesis This is all very early, and these might be the wrong answers, but hopefully they’re the right questions Deployment “Everyone is already using this!” OpenAI reports 900m+ weekly users (but only 5% are paying) Consumer use is a mile wide and an inch deep Less than 1,000 prompts in a year means it isn’t a daily essential for at least 80% of users, so far Experimentation versus daily use Glass half-empty/half-full - lots of people using this sometimes, but fewer make it a daily habit The ‘capacity gap’ - usage versus potential The same for work - rapid growth, but how do we go from ‘I used it last week’ to daily essential? Imagine you were anaccountant seeing the firstsoftware spreadsheets Now imagine you were alawyer: “very cool, but…” How do wealwaysdeploy new technologies?“What’s this for?” Well, what’s the pattern? What works first? First, automate the obvious use cases - innovation takes longer What’s working first? Where is it easy and obvious to use generative AI? A generational change in software development“ChatGPT, make me a chart of product-market fit” “We’re seeing more and more exampleswhere one or two people are buildingsomething in a week that would have Benedict Evans–– May 202639 Writing code isn’t the hard part What should the code be doing, and where does it fit into the market? What gets bundled and unbundled? “For half of my jobs I tell clients who useExcel to switch to a database, and theother half are the other way around” The best answer to this question might be “BREATHE”What does AI do to software? More! But then what? “People don’t know what they want untilyou show it to them” “You’ve got to start with the experienceand work backwards to the technology” How do you build ‘AI software’? How do you turn an amazing raw technology into something people can use? How do you know what to automate? Step one: ask your systems integrator Automation takes a lot of manual labour PE roll-ups and GTM partnerships with outsourcers and strategy consultants Benedict Evans–– May 202646 Everyone has a pilot Enterprise software takes time, and come with early disappointments Startups exist to u