# industries ai creates, destroys, and just barely touches dario amodei keeps talking about how it's going to replace all developers by the end of yet another year, while the demand for software and its developers keeps rising. ben evans complains that he's being bullied for saying that ai is, "only as big as the internet", while some pundits are appalled that it's not as big as the industrial revolution. every other person is automating away their marketing team -- first with custom GPTs from openai, then with n8n workflows, and now with claude code. the ai pundits talk about grandiosity, solving all disease and poverty. alternative AI bro brainwash is about "replacing my marketing team in three hours". while on the micro level, everybody's busy tweaking their freaking prompts and fixing edge cases. but i don't see much discussion in the middle. i don't want to talk about what's going to happen with this solar system in a hundred years. i'm also tired of only thinking about how to solve yet another failure mode of my AI engines. so i started thinking about the macro level -- the level of entire industries and economies. and i came up with this non-exhaustive list of the industries made possible by artificial intelligence. let's debate! ## the eight industries **1. compute and storage** -- the boom of the data centers, undeniable beneficiaries of the AI summer we're living through. **2. developer tools** -- coding agents and software factories. better said, it's the industrialization of the software development process. **3. enterprise productivity** -- the automation of business processes across order management, accounting, support, procurement, and others. task automation growing into a promised jobpocalypse. **4. personal assistants** -- some people used custom GPTs for it. now openclaw & family are hyped, eventually rising to the level of os1 from _her_. **5. unmanned warfare** -- as exemplified by the battlefields of Ukraine, this is becoming increasingly autonomous, with some operations conducted without human involvement on the ground. soon they will be conducted without any direct human involvement in the command & control centers either. **6. physical industries** -- as we get more fatigued by the endlessly overwhelming stream of ai content, more people get drawn to real events, physical encounters, and physical products. this includes film photography; the printing of photographs, magazines, and books; and proof of provenance -- that a given song or image has been created by humans, or better yet, existed before generative ai came into play. **7. household robotics** -- and by that i don't mean humanoid. at some point manufacturers will understand that a fridge, washing machine, and kettle aren't very humanoid at all, while bringing tremendous value and changing the quality of life of people substantially. **8. simulation of senses** -- today best exemplified by virtual reality, and in the future vastly developed by world models. these use cases will be tied heavily to the entertainment industry, but simulated environments don't have to be. there could be an os-complete navigation environment for operating systems, and sandboxes for testing ai agents. there will be simulated environments of entire companies and entire economies that allow us to model their behavior predictably. ## the disruption will be different, not smaller some of those eight industries already exist and are at different stages of development. i believe all of them will be multitrillion-dollar industries in their own right. some are interconnected. but as you can see, they don't necessarily replace existing industries -- they create new ones. of course, they will penetrate existing industries like manufacturing, insurance, finance, bpo, and professional services to an immense degree. before an **insurance company**, there still needs to be capital, sales, an insurance product, and real-world claims management. ai-driven document processing and robotics will help with some of that, but organization, capital allocation, and repairs are complex, individual, and too low-value to introduce ai immediately. they will eventually be transformed, just as computers themselves impacted every industry without exception. a **bakery** uses a POS system to track how much sugar they have. some transistors to accept card payments. it sends emails from the phone to order more dough. despite all of that, fucking bread is still the fucking bread. you can't make an agentic bread. you can't make an agentic bicycle. you can't really make an agentic insurance either. services, in that sense, are much more malleable to agentic impact -- and a lot of the service industries will indeed be consumed by ai service providers. **banks** existed for six hundred years and lived through mnay technological revolutions. they survived because the core value proposition of a bank has nothing to do with technology -- it's their business model. it's the supply and demand of capital, which is free from the technological substrate it runs on. well, with better tech, it can run faster, at bigger scales. but there are still deposits, and there are still loans, and nothing in the agent economy is changing that. because of this, i believe the disruption that ai promises to the global economy will not be smaller than we think. it's just going to be different than we think. it's always different than we think, because we humans live in a world of a certain structure -- a collection of subsystems that work within a global societal super-system. but we forget that systems can change in different ways. maybe some constants change: - how fast money flows on average, - how many images a person consumes on average. important, but not the most important ones. - maybe the nature of value accumulation changes. - maybe a feedback loop breaks, or a new one is created. for example, automated claims appraisal in some property and casualty insurance has existed for over a decade now -- an example of a quicker feedback loop. did it change the cost structure of insurers' claims divisions? for sure. did it make the division disappear completely? no. did it change the nature of the insurance business such that insurance companies disappeared or changed? not really. most customers still have no idea whether claims appraisal by their insurer is human-driven or automated -- maybe until they get an insurance case, which by definition *must* be a rare event, and then they are happy. considering most people don't even think about that while buying a policy, and that most people will never have the specific claim they're insured against, on the global scale this simply does not matter. that's what's going to happen in most industries. it's not the first time we're going through automation -- mechanical automation is common in manufacturing and automotive production processess. cognitive automation as well, mostly in ops and backoffice. batch jobs and backend scripts already automate invoice creation or payments, which is a cognitive job. workflow tools for process automation have existed for at least twenty years. did they eliminate accountants or customer support agents? they changed the nature of the job to some extent. b ut you get my point: in most industries the impact of ai is substantial and powerful, and still incremental. it doesn't make or break the industry -- it optimizes it. ## why these eight are different my thinkingh with the eight industries -- computing and storage, industrialization of software development, enterprise productivity, personal assistants, unmanned warfare, physical industries, household robotics, and simulation of senses -- is that they're new. and i actually think about these parts somewhat chronologically. computing and storage has been happening for quite a while; developer tools are developing rapidly; enterprise productivity has a gazillion of tools, but the roi still needs to be shown, especially with recently increasing token prices; unmanned warfare exists and is being developed, but we're still a couple of years away from real drone swarms. ### compute and storage compute and storage have existed for a while, but with the current buildout of data centers, if you extrapolate how much compute and storage we'll need to serve every person with intelligence on top, we'll need ten thousand times more. we're a long way from saturation, and a lot of optimization will be done because it becomes economical. this is the industry that has exploded the hardest, and the actual numbers back it up. the four major hyperscalers collectively spent roughly $413 billion on data centers and ai infrastructure in 2025 -- more than double their combined spend in 2023, and combined capex rose by 84% from 2024 to 2025 alone, going from $224 billion to $413 billion. spending on data centers is projected to grow 55.8% in 2026 and surpass $788 billion, per gartner. zooming out, goldman sachs estimates roughly $7.6 trillion of cumulative capex between 2026 and 2031 across compute, data centers, and power, and mckinsey-cited figures put the build-out at $3 trillion to $8 trillion, about $5.5 trillion at the midpoint. so the "ten thousand times more" framing is about long-run saturation; the near-term run-rate alone is already a multi-trillion-dollar trajectory. > [placeholder for chart: data center / ai capex growth 2022–2030] -- see the goldman sachs "tracking trillions" charts at https://www.goldmansachs.com/insights/articles/tracking-trillions-the-assumptions-shaping-scale-of-the-ai-build-out and the bloombergnef build-out tracker at https://about.bnef.com/insights/commodities/ai-data-center-build-advances-at-full-speed-five-things-to-know/ compute, storage, memory, wiring, data center services, and everything connected to it has been exploding for a couple of years, and those hardware industries are the ones that have seen the most revenue from the ai boom so far. ### developer tools dev tools are the second industry reliving a big boom, and the industry itself is changing -- as exemplified by the (frankly nonsensical, but telling) sixty-billion valuation around cursor. for context, anysphere/cursor went from $100 million in ARR in january 2025 to $500 million by june, past $1 billion by november 2025, and $2 billion by february 2026, and on june 16, 2026 it was announced that spacex would acquire cursor at a $60 billion valuation, placing it under its xai subsidiary. the broader market matches the frenzy: ai coding tools generated $12.8 billion in revenue in 2026, more than double the $5.1 billion in 2024. but, compilers used to be developer tools at some point too. nobody thinks of compilers as developer tools anymore, although they still are. IDEs are developer tools, and nobody really thinks about that. kubernetes, command-line interfaces -- all of them are. ai agents, and especially software factories, bring us to a level never seen before, and it has qualitative implications. there's one clear trend: developer tools will converge. we'll keep climbing the layers of abstraction. the new hottest shift is software factories -- i'm working on such a factory for my self, but also on the **factory of factories**: software factories that build other factories. the model makers pride themselves on being general (which is factually not true). anthropic became successful lately precisely because it prioritized software development and programming above all other use cases. many of those other use cases simply followed as the model became more intelligent at programming, but many didn't -- the current new wave of hatred for ai writing is an example. claude is a shitty writer. these factories will run over specific repositories. even two repositories by the same person might need different best practices, different testing frameworks, different deployment pipelines -- because that's how software is. that's the reason there are three hundred CRMs systems, and why you can't just take one open-source one and plug it into your company. but with modern AI capabilities it becomes possible to build factories of factories: highly specialized and immensely powerful swarms of AI agents that can industrialize the process of software development one repository at a time. combine that with what i call a **foundation** -- an orchestration layer on top of the factories -- and one person with a properly designed and evolving foundation could build a whole family of software products simultaneously. it's still going to be expensive and take time, but it can run fully automated end to end: from the first idea of the user, through initial implementation, feedback loops, and subsequent improvement. i do not think it's a threat to the entire job of the software developer, but an inevitable change nonetheless. ### enterprise productivity enterprise productivity is a big and interesting field, and it has existed for dozens of years, because companies have always wanted to be more productive. you can say the invention of computers and their implementation owes itself to the demand of businesses -- which the name of "IBM" itself suggests. more recently, businesses were automating with ERP, CRM tooling, batch jobs, backend scripts, workflow-based automation, RPA, and now AI agents. there is nothing new under the sun. but the development of AI presents a different opportunity: the **total automation of an enterprise**. i have a simple framework to understand the levels of enterprise automation, which i'll share in a later article. the state of total automation means that any and all tasks in an enterprise are being executed by ai agents -- all and any document work. in that scenario, though, it makes little sense to talk about ai agents or workflow automation in the present sense. if all of the business processes are replaced by an ai, what is left are just databases of transactions, without a need for interfaces and without the need for separate systems. i believe every piece of company software will eventually converge to one **company brain** that oversees the activities of the entire organization, and everything runs autonomously. so enterprise productivity isn't only the workflows that connect systems -- it's the emerging industry that will eventually replace all of the enterprise software through total automation. with organized ai agents you'd be able to automate entire companies -- maybe not today, maybe in two years or five. there will be not just toys, but entire real organizations with revenue, run entirely by AI. and in large organizations, still too complex for AI to grasp or run effectively, we'll automate entire processes within departments. ### personal assistance personal assistants are another kind of productivity tool, but for individuals and not enterprises. while some people swear by openclaw and similar agents, i find them fairly useless. as a person, i do have some repetitive tasks. but they tend to combine repetition with subtle variation. that's very hard to handle with agents. they're not good with edge cases -- and edge cases are all there is for personal assistant systems. we are very far away from a fully functional system that can actually listen to you and do things right on demand, and remember your preferences across different contexts, like os1 did in the movie. https://search.brave.com/search?q=os1+did+in+the+movie a few years ago there were a few attempts -- the rabbit r1 and the humane ai pin -- to introduce those assistants to the public, but they failed miserably because the models lacked the capability for long-term memory. that might be solved with continual learning, but i'm not sure. no amount of databases and context management will truly replace it, because a naked LLM will never be good enough at picking up that context anyway. so for personal assistance to work, i believe there needs to be a fundamental change in the architecture of the ai models compared to what we have today. however, once it's technically possible and scalable, these assistants can grow very rapidly. ### unmanned warfare unmanned warfare has exploded in recent years. take ukraine: FPV drone production went from a few thousand units in 2022 to more than two million in 2024, and monthly capacity jumped roughly 900% in a single year, from about 20,000 units a month in summer 2024 to over 200,000 a month in 2025. by late 2025 ukraine was producing around 4 million drones per year -- more than any nato country, and likely more than the entire alliance combined, built out from a pre-war base of roughly ten manufacturers to several hundred producers. https://www.kyivpost.com/post/55897 a lot of Ukrainian defenders started using drones as the means of asymmetric combat, with great success. there are many attempts at AI-powered swarm intelligence; more basic perception tasks such as object recognition have worked on such devices for a while. > police drones will not always shoot rubber bullets while we're still trying to figure out how the simplest and dumbest drone swarms should coordinate -- and they fail at the slightest sight of a cloud or a branch covering the view -- unmanned warfare will likely not stop at drones. it's switching to other types of mobile units: robot dogs, planes, and surface vehicles on water (which has, by the way, already happened). we will soon have armies of small drones and big drones, robotic hunter-killer dogs, tanks, planes, and aircraft carriers handled entirely with robots and ai. it also matters enormously **who controls them**. think about police drones: whether they're run by a friendly government or a non-friendly one changes everything. and you need to understand that the police drones will not always shoot rubber bullets. ### physical industries you think going to a live event is stupid and shooting film is outdated? think again. think through the time when maybe we won't need to work -- or the other way around, we won't have much to work on at our computers. a lot of office workers spend their entire time at a computer. what if in ten years all of this work is automated? literally all of it. what are those people going to do? where are they going to go? what are they going to hope for? i don't know, but if they don't have a resource -- a computer that feeds them -- they will gather together in smaller spaces and sing. and someone else who lost their job a month ago due to automation will come and listen, and throw five euros into their hat. after another year they'll get tired of singing for stuff thrown into a hat, and they'll start charging for it. and other people will still come, because they need a resource. and they'll be stonemasons, and they'll paint, and they'll raise each other's children. and for fuck's sake, maybe that's good, as we have lost any sight of human connection in the last thirty years. in that world, canon or kodak might finally move their ass and make a proper modern film camera: modern equipment and lenses, advanced autofocus, extreme sharpness, versatility of shutter speeds -- a quick, simple, physical interface, yet fully automated. as simple to use as a smartphone camera, but still with that deep touch of reality only a film camera can offer. you might think i'm a film photo geek -- and i am -- but there's a reason for it. there's a reason people still buy thirty- to fifty-year-old cameras that sometimes barely work. they spend around one euro per single photograph, without printing, just to see it on screen, and 1-3 weeks between the shot and seeing an image. and they spend time buying the film, loading it, bringing it to get developed and scanned, and picking it up again. it's literally incomparable how much easier it is to buy a good modern digital camera, which is going to be a hundred times faster and sharper than your old film piece of bullshit. the camera itself might be much more expensive, but using it is much cheaper. the point is not about film cameras. the point is that the physical world is not the same as digital. meeting someone physically is not the same as calling them over video. it's just not the same, and with AI we will feel it more and more. ### household robotics household robots are shit -- that will change. but real dexterity, as we've discovered, is actually hard; i give that ten years. the other thing is they're creepy. humanoid robots are creepy as hell. who would ever want a humanoid robot? it's literally a killing machine in your house. i love using paper notebooks, but i really don't love taking pictures of them and scanning them. why isn't there a small robot on my desk that does that? or the same for scanning documents -- every entrepreneur in Germany knows that scanning your jahresabschluss is a drag; you spend an hour photographing and synchronizing and doing all the jazz. why isn't there a robot that cleans my shoes? that does my dishes or laundry? people keep talking about that, but you don't need an uber-dexterous humanoid thing to do it. it doesn't need to be that universal either, because most kitchens are fairly similar. give it a couple of learning experiences -- and accept that it's going to break a couple of plates before it learns -- and it's not that bad. but i want those robots to be secure. i don't want them observing me while connected to the internet, watching how i am on the toilet (seen some reports of this happening). that's the problem nobody wants to solve, because many companioes want to steal your data. and there's no good business model to avoid that. funny enough, that's often the problem -- the business model, not the technology. we don't know how to sell household robots. we don't know what real use cases people are ready to pay for. we could make some of them happen today, instead of developing universal, dexterous humanoids, which i'm very skeptical of -- which is inconsistent, because i love movies like _blade runner_, so it would be natural for me to think they must exist. and some part of me actually says yes, humanoid robots will exist, if for no other reason than that people love doing things that are like them. chatbots speak kind of like us -- but if we learned anything about chatbots, it's that "kind of like us" is still very different from a human; you recognize AI slop from a mile away. they might fake human speech better and better if you prompt and guide them, but it's hard to imagine they'll actually write like humans, because that's not what they're trained for. an average of all human writing is not human -- exactly because of the human quirks and individuality. it does sound soft and wacky, and it is, but that's the point: it's soft and wacky at its heart, hard to quantify and hard to depict in the data. that's one of the reasons language models have trouble sounding human. they don't really need to. and household robots don't need to look or sound like humans. they just need to fulfill a goal, satisfy a need, solve a problem. just like the dishwasher, the fridge, the washing machine. those are boring, basic appliances that made a revolution in how people live a hundred years ago -- and keep making it happen. it's just that we don't care anymore. that doesn't make them less important; less exciting, for sure, but no less important for quality of life. imagine yourself without a fridge and a washing machine. some people live without washing machines and go to laundromats -- how fun is that? i don't remember any apartment i've ever seen without a fridge. hundred years ago it didn't exist at all, and i have no idea how people preserved food beyond spices and the basement. household robotics will be massive -- as soon as it learns what problem it's solving. ### simulation of senses the ultimate picture might be simpler and much less pleasant than we'd like to hope. i do think that some people, when the simulation of reality becomes appealing enough, will decide to spend all their lives, time, and effort in a simulated reality. i don't think there will be very many, but i'm pretty sure there will be some. some of them will be rich and dictate the rules, and some will be poor and serve the rich in any way, shape, or form possible to survive, to satisfy their need, to pay the bills. there will be an economy for purely simulated reality. probably nobody will spend a hundred percent of their time there, but there will be people working full-time and overtime -- eighty hours a week -- in those environments. that's one part of it, and world models like deepmind's genie are a definite prerequisite. but as i mentioned, that's not the be-all and end-all of simulated environments. the capability to simulate the behavior of a company, of five companies on an oligopolistic market, of the hundred main companies in a country's economy, of the twenty thousand biggest companies in the world -- if you could predict exactly what twenty thousand of the most important companies will do, you'd be able to predict maybe 95% percent of what's happening on the planet. include governments and organizations, and twenty thousand -- or ten thousand, if you like round numbers -- would cover 90% percent of what happens. numbers are speculative, but i believe they will hold. i do not subscribe to conspiracy theories saying that the corona pandemic was designed and manufactured. i do not believe that. however, a lot of world-changing things -- like the war in Ukraine, the war in Iran, the potential war between China and Taiwan -- are designed and manufactured. if you could precisely simulate what all of those actors are about to do, you'd be able to quite literally predict the future, at least on the strategic scale, at least in the midterm. that's also a simulated environment. simulated environments are the area i know and understand the least -- not just because it's complex, but because it doesn't exist yet. still, the industry of the simulation of senses is an interesting one, because i don't mean senses in a narrow way. although even the narrow version would be enough for a multi-billion-dollar industry, senses are everything we experience in the world. just about five days ago, i stopped listening to youtube all the time when i cook, walk, and drive somewhere. it brought me some peace of mind and let me hear my own inner voice. it allowed me to sit now on a bench near the deutsche oper, under a big old tree, drinking a beer, talking to my transcriptors about the future of ai-powered industries, and wondering where we'll end up. ## closing all in all, this is an essay on how new and old ai industries will grow and develop -- in very rough strokes. if we talk about strokes and their precision in descending order, we have maurits escher, then goya, then van gogh, then mark rothko. i have drawn this essay like van gogh right now. it's not a formless blur like rothko, but the industries is far from being precisely mapped out as escher would do it. that's the world mid-2026 we live in. that's my essay. enjoy it, relent it, digest it -- and come back after ten years to tell me i was wrong. --- // 22 jun 2026, berlin