AI eats the world

Benedict Evans –– December 2020 AI eats the worldBenedict Evans November 2025 www.ben-evans.com

Benedict Evans –– November 2025 2The next platform shiftEvery 10-15 years, a platform shift reshapes technology PCsMainframesWebSmartphonesGenerative AI

Benedict Evans –– November 2025 3Who is affected, and how much?What happens in a platform shift?The New Thing! All tech innovation, investment and company creation switches overInside tech New gatekeepers, new value capture New and bigger markers Outside tech Is this a new tool, new revenue, or an existential threat?

Benedict Evans –– November 2025 4Source: Apple, Google, GartnerMicrosoft OS share of global computer unit salesMicrosoft dominated the PC era, but when the centre of gravity shifted to smartphones it became irrelevant Dominance is won and lost 0%25%50%75%100%19751980198519901995200020052010201520202025

Benedict Evans –– November 2025 5Source: GartnerGlobal PC unit sales shareThe ‘first’ was not the winner in PCs, browsers, search, social or smartphonesEarly leaders often disappear 0%25%50%75%100%197519801985199019952000AppleAll othersIBM-compatible

Benedict Evans –– November 2025 6For every new platform, we forget how many ideas failed and how unclear everything wasHow will the new thing work? We don’t knowInternet AOL, Yahoo, Pointcast, Flash, plugins, portals… Sun? Netscape?Mobile internet i-mode, J2ME, WAP, mesh, DVB-H, keyboards… Nokia? RIM?Generative AI? Browsers? Agentic? Voice? MCP? Wearables? GEO?

Benedict Evans –– November 2025 7When things are exciting, people get excited Noise, hype, anti-hype

Benedict Evans –– November 2025 8This often brings bubblesPeople draw straight lines on log scale charts They forget that exponential growth is generally an exponential curve And always say “this is different” The trouble is, they’re generally right – every bubble is different! But it can still be a bubble

Benedict Evans –– November 2025 9Source: Rosenfeld HCMSTUS heterosexual couples who met online, by year of meetingThe internet has gone from the New Thing to a basic part of daily lifeBut when the dust settles, the world has changed 0%10%20%30%40%50%60%199520002005201020152020

Benedict Evans –– November 2025 10SaaS means the typical large enterprise in the USA now uses 4-500 appsNew platforms mean new tools (and new revenue)Mainframes One appPCs Dozens of appsSaaS 4-500 apps*LLMs What can we automate next?*Source: Productiv

Benedict Evans –– November 2025 11One way this platform shift is different, though For PCs, the web or smartphones, we knew the physical limits of what could happen next year With LLMs, we don’t know how much better this could get “The race to AGI is afoot” Sergey Brin “AGI needs multiple further breakthroughs” Demis Hassabis

Benedict Evans –– November 2025 12Another platform shift, or more?We know this will get better, but we don’t know how muchPCsMainframesWebSmartphones?Generative AI

Benedict Evans –– November 2025 13How will this work?How will this be useful?Where is the distribution, value capture, and value destruction?If this is ‘only’ as big as mobile or the internet, that seems like enoughSo how will the new thing turn out?

Benedict Evans –– November 2025 14Inside tech

Benedict Evans –– November 2025 “The risk of under-investing is significantly greater than the risk of over-investing” Sundar Pichai, Q2 2024 15“The very worst case would be that we have just pre-built for a couple of years” Mark Zuckerberg, Q3 2025

Benedict Evans –– November 2025 16“This is a huge new market, a huge threat to all our existing businesses, and we can’t miss it”Three years of FOMO in Big Tech“This is a huge opportunity and threat We can’t miss it”“It keeps getting better with more capex”How much better? How much capex?

Benedict Evans –– November 2025 17Source: Companies, company guidance. Includes capital leases * Amazon does not break out AWS capex directlyCapex ($bn) ~$400bn in 2025 for the big four alone (for comparison, global telecoms is ~$300bn)FOMO drives a capex surge01002003004005002010201120122013201420152016201720182019202020212022202320242025eMicrosoftAWS*AlphabetMetaOracleCoreweave

Benedict Evans –– November 2025 18Source: Companies, company guidance. Includes capital leases * Amazon does not break out AWS capex but reports it as ‘the majority’Capex ($bn) ~$350bn in 2025 for the big four alone (for comparison, global telecoms is ~$300bn)Planned 2025 growth almost doubled – in 202501002003004005002010201120122013201420152016201720182019202020212022202320242025eJan est for 2025MicrosoftAlphabetAWS*MetaOracleCoreweave

Benedict Evans –– November 2025 19Source: US Census. * Excludes mini-storage ** Excludes compute NB: data after July 2025 delayed by US government shutdownUS construction value (2025 $bn, seasonally adjusted annual rate)US data centre construction overtaking officesA new investment cycle0255075100125Jan 1995Jan 2000Jan 2005Jan 2010Jan 2015Jan 2020Jan 2025OfficesRetailWarehouses *Data centres**

Benedict Evans –– November 2025 20Source: Companies Quarterly revenue ($bn)Trying to build a new Sun Microsystems (though China and hyperscalers’ own chips are coming up behind)Nvidia can’t keep up0102030405060200020052010201520202025IntelNvidia

Benedict Evans –– November 2025 21Source: Companies Quarterly revenue ($bn)TSMC unwilling/unable to expand capacity fast enough to meet Nvidia’s bookNvidia can’t keep up (neither can TSMC)0102030405060200020052010201520202025IntelNvidiaTSMC

Benedict Evans –– November 2025 22Source: Schneider Electric industry surveyMain constraints to data centre construction, USA (February 2025)US power demand growth is ~2%, and AI might add 1% that’s hard to build fast (this is not an issue in China)US power backlogs becoming a major issue 0%25%50%75%100%Utility powerAccess to chipsFibrePermittingWater availabilityAccess to natural gasLandPublic oppositionExtremely significantVery significantSomewhat significant

Benedict Evans –– November 2025 23“It’s been almost impossible to build capacity fast enough since ChatGPT launched” Kevin Scott, Microsoft CTO

Benedict Evans –– November 2025 24“We now expect the FY26 growth rate to be higher than FY25 ” – Microsoft “Capex dollar growth will be notably larger in 2026” – Meta “We expect a significant increase in 2026” – Alphabet

Benedict Evans –– November 2025 25Source: CompaniesGlobal data centre capacity estimates, H1 2025 (GW)Some very large numbers (although some of the ‘bragawatts’ may be more performative than real)Data centre capacity triples? For $3tr? $5tr? More? 0100200300Omdia/IEABloomberg BNEFGoldman SachsMorgan StanleyBainBCGMcKinseyCurrentEst. of 2030 plans

Benedict Evans –– November 2025 26Annualised AI capex aspirations are a similar magnitude to mature global capital-intensive industries“Three trillion dollars!”Global telecoms ~$300bnOil & gas ‘upstream’ ~$540bnUS private sector ~$20.1trGenAI $500-750bn annual?Source: ITU, IEA, ONS

Benedict Evans –– November 2025 27Source: Companies, consensus estimates. Calendar year.Annual Capex and Free Cash Flow, 2010 to 2025e ($bn)Big Tech cashflow has surged since the Pandemic, and most of that growth is going on AI capexThe hyperscalers can afford it…-50050100150200Alphabet AmazonMetaMicrosoftFCFCash capex

Benedict Evans –– November 2025 28Source: Companies, consensus estimates. Calendar year.Annual Capex and Free Cash Flow, 2010 to 2025e ($bn)Capital leases are not new, but they’ve got a lot bigger The hyperscalers can afford it… up to a point-50050100150200Alphabet AmazonMetaMicrosoftFCFCash capexLeases

Benedict Evans –– November 2025 29Hyperscalers add leases and debt, while some analysts suggest Oracle’s cloud capex might be >100% of revenue Up to a point

Benedict Evans –– November 2025 30OpenAI joins the clubAnnounced commitments for 30GW+ of capacity at $1.4tr Aspiration for 1GW/week of new construction at $20bn/GW = ~$1tr annually… Equivalent to 2/3 of total current global base, every year Source: OpenAI, October 28 2025

Benedict Evans –– November 2025 31“Circular revenue”Without its own cashflows, OpenAI partners with Nvidia, Oracle, Softbank, petrodollars… OpenAI is buying Nvidia chips with Nvidia’s cashflow… Which comes from the hyperscalers… and using Nvidia’s cash to turn AMD into an Nvidia competitor, and pay Broadcom to make its own chips…

Benedict Evans –– November 2025 32Nvidia has $72bn of TTM FCF* and TSMC can’t keep up with demandUse your excess cash to buy demand, FOMO and platform lock-in OpenAI has mindshare and expensive stock, but no platform, infra or moatSwap your paper for hard assets and market positionOracle is a cash-generative legacy business losing share to cloud and now to AIGear up and burn your way into the new thing?What would you do if your company was sitting on a bubble?Rational actors? *Source: Nvidia

Benedict Evans –– November 2025 33After three years, lots more science and engineering, but no real clarity on the shape of the marketYes, but where has all this money got us?Models still improving Far more models, China, OSS Lots of new acronymsNo apparent moats No clarity on product or value capture

Benedict Evans –– November 2025 34Source: ArtificialAnalysisModels by aggregate benchmark score Every week – new models, new (problematic, gamed, saturated) benchmarks, new acronyms Far more models0255075 Nov 2022 Nov 2023 Nov 2024 Nov 2025OpenAIAll other

Benedict Evans –– November 2025 35Source: LMArena, Artificial AnalysisBest scoring model relative to leader, October 2025Dozens of (saturated) benchmarks to choose from, but on the most general, the leaders are very close Models converge and leaders change weekly90%95%100%LMArena TextOther: WesternOther: ChinaOpenAIGoogleAnthropicMMLU-ProNB truncated axis

Benedict Evans –– November 2025 36Source: Reuters InstituteWeekly active users of generative AI tools as share of the population, June 2025The models may be close to commodities (especially for general use), but market position is not, so farTech versus brand versus distribution?0%10%20%30%USAUKFranceDenmarkJapanChatGPTGoogle GeminiMeta AIMicrosoft CopilotGrokClaude

Benedict Evans –– November 2025 37Source: OpenAI. NB: round numbers reported at scheduled events ChatGPT global weekly active users (m)800m weekly users, but apparently only 5% are paying – and why announce WAU and not DAU?“Everyone is already using this!”02505007501,000 Dec 2022 Jun 2023 Dec 2023 Jun 2024 Dec 2024 Jun 2025 Dec 2025

Benedict Evans –– November 2025 38Source: Deloitte. Surveys conducted in June. Defined as “Use of a purpose-built GenAI tool, like ChatGPT, Microsoft Copilot, Google Gemini, etc”Consumer generative AI chatbot use: USASo far, many more people use chatbots occasionally than make them part of their daily livesStill more experimentation than daily use0%10%20%30%40%50%60%202320242025DailyWeeklyMonthly202320242025Consumer generative AI chatbot use: UK

Benedict Evans –– November 2025 39Source: Consumer surveys conducted as of date shownHow many people use generative AI chatbots in the USA?Surveys are early, scattered and inconsistent, but an engagement gap seems clearMost data shows the same picture0%10%20%30%40%50%Pew, Aug 2024CCIA, Nov 2024Blick et al RPS, Nov 2024Bain, Dec 2024Hartley et al, Dec 2024YouGov, Mar 2025Hartley et al, Jun 2025Reuters Inst., Jun 2025Deloitte, Jun 2025Bain, Sep 2025Daily activeWeekly active

Benedict Evans –– November 2025 40Why do most users of ChatGPT only use it a little bit?Is this just early? Or a harder problem?How many use cases are an obvious easy fit?Who has a flexible job… And consciously looks for ways to optimise?For everyone else, do you need to wrap this in tooling and product?

Benedict Evans –– November 2025 41“People don’t know what they want until you show it to them” “You’ve got to start with the experience and work backwards to the technology” Steve Jobs

Benedict Evans –– November 2025 42Where is the value capture for a research-heavy, capital-intensive commodity?So where does a model lab compete?If the models are commodities? And don’t have network effects?Go down the stack Win on scale? (see: aircraft, AWS, chips)Go up the stack Win on network effects & product? (See: software)

Benedict Evans –– November 2025 43Source: Microsoft. Includes capital leasesMicrosoft capex/salesFrom competing on network effects to competing on access to capital?Microsoft’s shift away from network effects?0%10%20%30%40%50%Mar 1995Mar 2000Mar 2005Mar 2010Mar 2015Mar 2020Mar 2025

Benedict Evans –– November 2025 44Everything, everywhere, yesterday (on other people’s balance sheets), before the market slips away For OpenAI, “yes!’ to everythingInfra deals with Oracle, Nvidia, Intel, Broadcom, AMD…Ecommerce integrations, ads, vertical data setsApp platform, social video, web browserRobots Jony Ive Biotech

Benedict Evans –– November 2025 45“There are two ways to make money. You can bundle, or you can unbundle” Jim Barksdale

Benedict Evans –– November 2025 46What’s the right experience? The right distribution? And why isn’t it just the ChatGPT app?OpenAI bundles and unbundles use cases

Benedict Evans –– November 2025 47Source: Y Combinator Y Combinator startups by fieldThe coming wave of AI startups trying to unbundle Google, Excel, email and Oracle… and ChatGPTStartups exist to unbundle use cases0100200300400500W 2015W 2016W 2017W 2018W 2019W 2020W 2021W 2022W 2023W 2024F 2024S 2025AIOther

Benedict Evans –– November 2025 48If models are near-commodities, and we don’t know the right product, where will the value be?Where is the value capture?Best model? Most capital?Proprietary vertical data?Distribution & GTM? Product? UX?Building ‘normal’ software companies?

Benedict Evans –– November 2025 49Outside tech

Benedict Evans –– November 2025 50How do we always deploy new technologies? “What’s our AI strategy?” Well, what’s the pattern?Absorb Automate obvious use-cases Make it a featureInnovate New products, bundling and unbundlingDisrupt Redefine the question

Benedict Evans –– November 2025 51Where is it easy and obvious to use generative AI?So far, most successful use-cases are ‘absorb’CodingMarketingCustomer supportAutomation

Benedict Evans –– November 2025 52AI coding as the new AWS“Vibe coding” as the new abstraction layer, after AWS, libraries, operating systems… A new step change reduction in software creation costs

Benedict Evans –– November 2025 53Source: Accenture Accenture reported new quarterly ‘generative AI’ contracts ($m)Step one: ask your systems integratorHow do you know what to automate?05001,0001,5002,000Feb 2023Aug 2023Feb 2024Aug 2024Feb 2025Aug 2025

Benedict Evans –– November 2025 54Source: Palantir Palantir quarterly revenue by segment ($m)Step two: buy some SaaS from Dr EvilHow do you know what to automate?02505007501,0001,250Mar 2020Mar 2021Mar 2022Mar 2023Mar 2024Mar 2025GovernmentEnterprise

Benedict Evans –– November 2025 55Source: McKinseyAI ‘agent’ use by business function, where Generative AI is already used. June 2025“Agentic!’ is 2025’s buzzword, but deployment takes longer Pilots come first and deployment takes time 0%10%20%30%ITKnowledge managementMarketingServiceProductSoftware developmentHRSupply chainManufacturingDeployedDeployingPilotExperiment

Benedict Evans –– November 2025 56“Why did our AI pilot fail?” That’s a CTO question, not an AI question Not everything works? Welcome to techSecurity, privacy, IPR, error rates, legalData integration & legacy systemsFinding the right solution for the right peopleThe same issues as deploying any new tech

Benedict Evans –– November 2025 57Source: Morgan Stanley CIO SurveyCIO expected timing for first LLM projects in production, September 2025A quarter of CIOs have launched something – but 40% don’t plan anything until at least 2026The future can take time0%25%50%75%100%Already deployed at least one projectH2 20252026Later or no plan

Benedict Evans –– November 2025 58Source: Goldman Sachs CIO Survey Enterprise workloads in public cloudCloud is old and boring – but still only 30% of workflowsBut the future always takes time0%10%20%30%40%50%60%2013201420152016201720182019202020212022202320242025TodayExpected in 3 years

Benedict Evans –– November 2025 59Source: FMIAverage SKUs per supermarket, USAUPCs, barcodes and databases let retailers manage 5x more SKUsSometimes ‘automation’ alone is a big deal010,00020,00030,00040,00050,00060,00019501960197019801990200020102020Grocery barcodes launched in 1974

Benedict Evans –– November 2025 60But… “We’ve all seen lots of AI presentations now, and we’ve deployed a bunch of stuff. Is that it? What’s next?” F100 Retailer CMO, summer 2025

Benedict Evans –– November 2025 61What comes after automating the obvious, easy things?What next?Absorb Automate obvious use-cases (20 years more deployment to come!)Innovate? New products, bundling and unbundlingDisrupt? Redefine the question

Benedict Evans –– November 2025 62What can LLM automation unbundle? What things did we not realise were bundles?Where might we look for change?Online distribution unbundled physical assets What do LLMs unbundle?Internet unbundling created new aggregators How do LLMs do that better?

Benedict Evans –– November 2025 63“AI gives you infinite interns”

Benedict Evans –– November 2025 64We have no indication that error rates will go away, so where’s the human in the loop?How do we use automation that makes ‘mistakes’?Do ‘errors’ matter?Can you automate verification?Is human verification efficient?How much do we need to wrap the LLM in software?

Benedict Evans –– November 2025 65The Jevons paradox – applied price elasticityWhat do ‘automated interns’ change?Do you do the same work with fewer people? Or more work with the same people?Was employing lots of people your moat?What becomes possible when you don’t need millions of people to do that?

Benedict Evans –– November 2025 66Source: LandesUK population and steam engine labour unit equivalent (m)Steam engines gave Britain the equivalent labour of (very roughly) 5x its total population by 1900 300m interns? Jevons paradox at work050100150200250300184018701896PopulationSteam engine equivalent

Benedict Evans –– November 2025 67All recommendation systems today work by driving, capturing and analysing user activity Where’s the human in the loop?Pre-internet Human editors, for both retail & media Physical assets as moatInternet All of us are mechanical turks feeding algorithms Network effects as moatGenAI? Can an LLM do this better? Can an LLM do it without needing a user base?

Benedict Evans –– November 2025 68Source: Companies, WPP Media NB: ‘China’ represents spending in China, not global spending by Chinese companies Global ad revenue, 2024 ($bn)Brands spend a trillion dollars a year to talk to consumers – plus rent, shipping, marketing, returns…Value to capture!01002003004005006007008009001,0001,100Google SearchMetaAmazonAll other ex-ChinaChinaYouTubeGogole Network

Benedict Evans –– November 2025 69“Advertisers that activate AI Max in Search campaigns typically see 14% more conversions” Google, Q2 2025 “Our new AI recommendation model drove 5% more ad conversions on Instagram and 3% on Facebook” Meta, Q2 2025

Benedict Evans –– November 2025 70Ad asset creation costs ~$100bn globally: now add 10-20x more assets and unlock cheap video for everyoneAbsorb the new thing, automate what you know

Benedict Evans –– November 2025 71Old: “half of AI will be turning three bullet points into emails, and the other half will be turning emails into three bullet points” New: half of AI will be turning three bullet points into 300 ads, and the other half…

Benedict Evans –– November 2025 72Source: BainUS consumer search preference (September 2025)Use so far may be more additive and experimental than substitution (and this includes Gemini)Again, this is early0%25%50%75%100%AllUnder 3030-4545-6060-80OlderAlways GenAIMostly GenAIMixedMostly seachAlways search

Benedict Evans –– November 2025 73Remember how early this is, and how hard it is to know how the new thing will workThe web has been dying since 1997Source: Zeldman

Benedict Evans –– November 2025 74What if recommendations go from correlation to an understanding of what those SKUs really represent?But where are we going?You bought packing tape – maybe you need boxes and bubblewrap?Maybe you’re moving. How about some lightbulbs and smoke alarms?Here’s an ad for home insurance

Benedict Evans –– November 2025 75How do we know what we might want?For 30 years we’ve had infinite product, infinite media and infinite retail Now we have a machine that sees all of it, and sees us What does it recommend?Source: Morioka Shoten

Benedict Evans –– November 2025 76What do you actually want? What are you trying to do? Why? What do you care about?And what gets unbundled?Why did you ask for that? Do you care where it comes from?Logistics, data utility, answers Solve my problemExperience, curation, enjoyment, authenticity

Benedict Evans –– November 2025 77“What’s our AI strategy?” Is this a question for the CIO? CMO? CEO? Accenture? Publicis? Bain/BCG/McKinsey? Is this a new tool or a new industry?

Benedict Evans –– November 2025 78AI eats the world

Benedict Evans –– November 2025 79Source: US Census. Seasonally adjusted ‘Retail’ excludes restaurants & bars. ‘Addressable retail’ also excludes cars, car parts & service, and gasoline stations. ‘Core retail’ excludes grocery from addressable retail.E-commerce as % US retail sales Excluding gas and grocery, 30% of US retail sales are now onlineThe old stuff from before ChatGPT is still here0%5%10%15%20%25%30%35%200020052010201520202025Core retailAddressable retailAll retail

Benedict Evans –– November 2025 80Source: California Public Utilities CommissionMonthly ‘robotaxi’ trips in CaliforniaAfter a decade of promises and tens of billions of dollars, ‘automatic cars’ might be starting to workAnd all the other new stuff 0250,000500,000750,0001,000,0001,250,000Sep 2023Dec 2023Mar 2024Jun 2024Sep 2024Dec 2024Mar 2025June 2025Sep 2025WaymoCruiseTesla

Benedict Evans –– November 2025 81

Benedict Evans –– November 2025 82How many times have we been here before?Source: IBMWe’ve had radical change (and bubbles) before And we’ve also done automation before

Benedict Evans –– November 2025 83‘Automation and technological change’, 1955 Source: US Congress report on automation, 1956

Benedict Evans –– November 2025 84“Automation”Source: US Congress report on automation, 1956

Benedict Evans –– November 2025 85“Automation”Source: Benedict Evans

Benedict Evans –– November 2025 86Source: US CensusElevator attendants, USAOtis launched the ‘Autotronic’ automatic elevator in 1950When automation works, it disappears 025,00050,00075,000100,0001900191019201930194019501960197019801990

Benedict Evans –– November 2025 87“AI is whatever machines can’t do yet” Larry Tesler, 1970

Benedict Evans –– November 2025 88Thank you

Benedict Evans –– November 2025 89What matters in tech? What’s going on, what might it mean, and what will happen next? I’ve spent 25 years analysing mobile, media and technology, and worked in equity research, strategy, consulting and venture capital. I’m now an independent analyst, and I speak and consult on strategy and technology for companies around the world. Mostly, that means working out the right questions. For more, see www.ben-evans.com

Benedict Evans –– December 2020 Thank you Benedict Evans November 2025 www.ben-evans.com