Carriers Spend $200 Billion a Year. The Money Went Up the Stack

Global mobile operators watched data traffic rise by a thousand or two thousand times over fifteen years. It is a trillion-dollar industry. They spend about $200 billion a year on capital expenditure. Do they make money? A bit, Evans says, but not the money that matters. They did not get to do Uber. They do not run your banking. They do not do YouTube. All of that sits with other people. The value went up the stack.

Jacob Effron, a managing director at Redpoint, brought Benedict Evans onto episode 91 of Unsupervised Learning to talk through exactly that kind of ledger. Evans is a former a16z partner who now writes independently; his newsletters and annual slide decks circulate widely in the industry. The conversation is about who captures AI's money, what the model companies might actually be worth, and why consumer AI apps still feel stuck.

Effron frames Evans as having a subdued take only by current standards: as big as the internet and mobile, not a civilizational break on the scale of the Industrial Revolution. Evans's method is to stop arguing about which wave was larger and to study what happened the last five times a world-changing technology showed up. The mobile-network story is the one he keeps returning to, because it is a warning about mistaking a huge, expensive infrastructure layer for the place where profits land.

Is AI bigger than mobile? Was mobile bigger than PCs? Evans does not think you can quantify those comparisons in a useful way. You can say the web was bigger than PCs in one crude sense: when Marc Andreessen launched Netscape there were perhaps 75 to 100 million PCs in the world, and now there are five billion smartphones. Platforms have gotten larger. Debating which one wins the size contest, he says, has little percentage. What is interesting is what happened the last five times.

A computer-science PhD at a large company recently told him that big firms had never before had to deal with employees building tools and systems for themselves. Evans's first thought was: have you heard of shadow IT? Have you heard of Excel? Yes, this is different. It is always different, and always bigger. That is not a reason to pretend nothing like it has happened. Wave your hands and say this is like electricity if you want. Fine. What happened with electricity? People before electricity were not dumb. They scratched their heads and tried to work out what to do with it.

He had just written about token pricing by going back through semiconductors, mobile, fiber, and operating systems. None of those analogies have predictive power. You cannot prove AI will follow mobile by arguing it is close to mobile. You can still learn the range of possibilities. Semiconductors had Rock's Law: the cost of a cutting-edge fab doubles every four years, and foundation models look a little like that. Mobile networks have marginal cost, which a lot of people in tech do not understand. Double the traffic and you have to build more base stations. That looks like the compute crunch of the past six months, when everyone got a new phone, started watching video, and the networks fell over and had to rewrite pricing.

Each shift also compounds. Mobile internet did not wait for the internet; the internet did not wait for PCs; PCs did not wait for semiconductors. Saying AI is as big as the internet can mean the internet was ten times PCs, so maybe this is ten times the internet. The genuine difference is that we do not know the physical limits. In 2010 you did not know the next iPhone's shape, but you knew it would not have a retinal implant, a one-year battery, and a five-dollar price. In 1995 you did not know how the internet would work, but you knew PCs cost a couple of thousand dollars and telcos were not about to give every household fiber next month. Today there is no good scientific account of why the models work so well or what happens next. Evans likes a Cuban Missile Crisis trading legend: a rumor that the missiles had launched, everyone selling, one veteran buying, because if the rumor was true they were all dead anyway and if not the stocks were cheap. If AGI really arrives, middle-class unemployment is not the problem. If it does not, go back to analyzing enterprise software. A lot of AI talk, he says, sounds like a slightly drunk philosophy student at two in the morning. You cannot deduce God from first principles. Meanwhile, build some stuff and see what people do with it.

Once you put the mobile-network lesson next to AI, the question becomes whether the model layer is Windows or a carrier. Sam Altman has said OpenAI wants to be like Windows. Evans's objection is simple: LLMs do not have network effects, so they do not look as if they will turn out like Windows. He also rejects the other extreme. TSMC has a monopoly on the cutting edge of semiconductors and sits among the ten biggest companies on earth. Last year its net income was half of Apple's. It is a great, profitable company. You do not write apps for TSMC. Uber does not write apps for TSMC. Windows is a great business. AWS is a great business. None of them owns the whole stack.

That is why AWS did not swallow the entire tech industry and why every app on an iPhone is not made by Apple. The right people to build enterprise go-to-market for creative software are not the right people to build the substrate. You could define the model as something that invents ten new enterprise apps, builds them, and figures out distribution. That, Evans says, is Anselm again: if you define it as able to do absolutely everything, then yes, it can. You have not won the argument. You have defined it away.

The competitive picture is, in his view, a negative signal dressed as a positive one. There are something like three to six companies making frontier models, and they are all kind of the same. Someone is ahead this week, someone else next month. That xAI could flunk out and then jump almost back to the top of the leaderboards is, paradoxically, bad news for xAI, because it suggests the feat is not that hard if you are willing to spend a couple of billion dollars and hire the right people. Search, social, and mobile had barriers to entry. This does not, not yet. Semiconductors shrank from dozens of cutting-edge firms to one, with two a step behind, as each generation got harder and more expensive. If scaling continues, expect the frontier set to shrink. If the curve slows, everything else changes. SK Hynix's CEO recently said supply and demand stay out of whack until 2030. Well, Evans says, he would say that, wouldn't he? It is good for his business.

He draws a Pareto curve. At one end, dictation already works for free on your phone. At the other, some jobs need the absolute cutting edge and will pay for it. Where does Amazon need to sit to get an ROI on recommendations or review summarization? There will be players all along the curve. The cheaper you go, the more commoditized it gets, perhaps with real-time bidding across thirty neoclouds. A paper could still land that makes a model ten times larger for ten times the price, or prices could collapse even as everything scales. How much of the value the model can capture is a separate question from how good the model is.

Evans keeps coming back to one word: jagged. A model can do something that takes a person seventeen or twenty hours and fail at something that takes five minutes. Intelligence is not linear. Charts that say Opus 8.6 can do 93 percent of a first-year law-firm associate's work drive him crazy, because you cannot measure the associate that way and you cannot measure the model that way either. It is the old expert-systems mistake: people thought you could write down how you tell a cat from a dog. You cannot. Geoff Hinton said a decade ago to stop training radiologists. The narrow problem, Evans says, is that machine learning could not do what Hinton thought it could. The broader problem is that that is not what radiologists did anyway. Clients do not buy a PowerPoint. The deck is the tangible expression of a project. They are buying everything else around it.

Picture an accountant seeing the first software spreadsheet in the late 1970s: life-changing. Picture a lawyer seeing the same grid: cool, I could use that for my timesheet next week, but that is not what I do every day. Today's AI users sit on that split. Evans himself is an outlier. He works alone, with no interns, and has structured the work so it does not produce repetitive tasks. Claude and ChatGPT arrived, and he does not actually have the tasks to give them. Model capability is one problem. Isolating and describing a job you could hand over is another, and it is especially hard with voice and chat. People in the Valley walking around with five Mac minis in a cluster and a laptop open so Claude Code can keep running are not the normal user. They are the 1992 Telnet crowd.

The number of accountants in the United States through the twentieth century is a straight line up and to the right, even as punch cards, adding machines, and mainframes automated the work over and over. That is not, in his reading, a simple Jevons story of doing more of the same. A fifth-year at PwC is not doing 1970, but more. The old work takes ten seconds once a week. Automation has climbed from legs to arms to fingers to brains, which sounds profound until you notice you could have said the same unfalsifiable thing in 1950 about work you could not yet see. The lump of labor fallacy still holds: you can always see the jobs that go away and you cannot see the new ones, which will use skills that have not occurred to you. Trying to score occupations on a radar chart is, he says, horse shit, like physicists predicting a horse race after assuming the horse is a perfect sphere.

Impact will be wildly uneven. Uber did not sell software to taxi companies. In New York the yellow-cab market shrank by something like three quarters while a larger ride market opened. Airbnb is maybe 5, 10, 15 percent the size of hotels and mostly slowed their growth, because half of hotel demand is business, a confidence business. Models are jagged, use cases are jagged, and some of the use cases are physical. The internet did not change what it was to be a journalist and still demolished local advertising, which was not in anybody's job-exposure model. Evans was in the mobile business, talking about location all the time, and nobody saw Uber. Consumer surplus is another quiet outcome: in 1980 a paralegal spent a day in the basement and came back with two precedents; later a database produced twenty, the other side also found twenty, the bill and the pages and the win rate stayed the same. The iPhone is the extreme of that pattern: tens of thousands of dollars of consumer electronics disappeared into one device.

About 10 to 15 percent of people are daily active users of chat products, and even they are using them once or twice a day, not all day. Another 20, 30, maybe 40 percent are weekly or monthly. From social products, Evans says, weekly active user is bullshit. Using a thing once or twice a week is not Silicon Valley's dream of changing computing. It is a useful new tool sometimes. Software development is the exception: product-market fit, unquestionably. Everyone else is somewhere between the accountant and the lawyer. In 1999 he sat in a meeting trying to get a Palm V onto the internet over GPRS through an infrared port on a Nokia. The room said it was so cool. It did not actually work. Nobody would really do that for another ten years.

Most people, he wrote in a column the night before the taping, are three things. They are not tool builders. They do not see the problem the tool would solve. And they are not in a position to build the tool even if the first two were false. Being a great salesperson is not the same skill as designing sales-enablement software. Being a great editor is not the same as inventing a new way to edit video. Every great software pitch makes you think you would never have come up with that. Yet the industry talks as if a random middle manager, handed Copilot this afternoon, will dream it up. A lot of the data is regulated. Fifteen hundred people touch it. It has to go into a system of record with money attached. You cannot let random people replace SAP. So the real world is a mush of CSVs, Excel, Tableau, big iron, and four or five hundred SaaS apps inside every large company. AI shuffles those layers. Workday, he jokes, is the ultimate test for AGI: can it navigate Workday?

Enterprise rollout has had a script. Step one, give everyone a Copilot. Step two, oops. Step three, run pilots. Lots of pilots, about half deployed, they work fine, one pain point at a time. It feels like putting all the catalogs on the website as PDFs: printing bills collapse, more customers download them, and it is still not the internet. Evans's reliable conference laugh line is that if you say "digital transformation" three times, an Accenture partner appears in a puff of smoke. Accenture has something like 800,000 people. That number may fall by half. No company keeps a standing army whose only job is to reimagine processes. So vertical software firms with forward-deployed engineers and consultancies with software engineers are climbing the same mountain from opposite sides. A forward-deployed engineer, he has not quite found the joke for yet, is an Accenture body-shopper who got a job at OpenAI.

Computer use, the hoped-for consumer unlock, sounds geeky to him. Most people's main device is a phone. In the nineties a frozen PC meant crawling under the desk, pulling the plug, and hoping you had not lost more than an hour. The new version is telling Cowork to clean the inbox and watching it delete the inbox. Use cases have to be invented. Flickr, Instagram, and TikTok were not grassroots consumer behavior. He was in the a16z pitch for Instacart, when nobody believed the economics. The founder did not have a car. He put up a bare-bones site that emailed him orders, took an Uber to the supermarket, shopped, and took another Uber to the door, just to see if the thing was real. Sora, Evans files with consumer social, not with AI: you need a new mechanic and a new feeling, and it has to be more than a gimmick. TikTok found that. Snap found two or three. Drones and 3D printers are the warning. One Christmas everyone bought a drone; three days later they had seen the roof of the house. Hobbyist case, not consumer case. Midjourney was a month of play. Virtual try-on could be a product: look at my Instagram, look at this brand's lookbook, make ten videos of me in different looks. That would be insane. It would also be a product. David Bowie told Jeremy Paxman in 1994 that the internet would change the world. Paxman said it was just strangers in chat rooms. The mistake was never the chat rooms. It was the network. AI as an enabling technology will improve fraud detection and back-office processes you have never thought about. Better products will arrive looking nothing like ChatGPT. Judging it by one slightly wrong answer is like calling the internet slow in 1995.

If he were running a lab, would he change the focus? Evans has a lot of sympathy for Sam Altman, which he admits may be a unique and controversial statement. The plates keep moving: devices, lawyers, power for data centers, funding, whether to make chips. It is like being Bill Gates in 1980 and having to build PCs, enterprise software, and broadband at once. Incumbents have a clearer play. Google sprays AI across every product, mostly successfully. Microsoft does not have the native capability, so it buys, rents, and spreads Copilot everywhere. Apple fumbled badly two years ago; a beta of the new Siri, he says, is about as good as ChatGPT was six to nine months earlier. OpenAI and Anthropic have no legacy business, no distribution, no infrastructure, only cutting-edge science, and they have to build the rest.

What OpenAI did in the second half of last year, in his telling, was ask what you can put on top of a commodity foundation model. A browser? A social video app? An app store, another app store, a third app store? Shopping? Ads? Everything. Anthropic had less money, more focus, more worry about evil outcomes, and went narrow on coding. They stumbled into the one place that works. Now everyone rushes toward coding, like the scene in *Life of Brian* where the crowd starts throwing rocks. He reminds listeners that coding is a pretty small industry, a remark that makes people in tech upset. This is also the first platform shift that got more expensive rather than cheaper, and that has marginal cost. You cannot ship a consumer app, get a hundred million users, and invent the revenue model later. That is why the story is enterprise. Coding flipped from kind of working to actually working at the beginning of this year. Whether that was inevitable is a different question.

OpenAI's recent app launches tell him one thing: writing the code is no longer the hard part. He calls the result a disastrous, chaotic, confused mess, with Copilot and Work and chat in a tangle of pop-ups and hidden modes. Hiring a wave of Meta executives felt cargo-cultish. In the Pacific in World War II, Americans built airfields, handed out cans and saucepans, then left. Islanders climbed the control tower, lined up on parade, and ran a flag up the pole, because that is what the Americans did when the planes came. They had imitated the forms without the cause. The people you want, Evans says, are the ones who were at Meta when it was still Harvard, not the ones who arrived when five billion people were using it. MySpace was first. Google had Wave and Buzz and all the data. Winner-take-all still required execution. The week before the taping, OpenAI gave up on the browser. His annual deck has a slide of 1995 companies, concepts, and acronyms that failed. Presume a bunch of what we are building now will not work. Presume MCP may be replaced. Goldman's CIO survey, as he remembers it, still has only about 30 percent of enterprise workflows in the cloud after twenty years. Diffusion is slow.

When Evans pulls the conversation into something you can use on Monday, three questions sit in front of any AI use case. Is there enough training data? Is validation scalable and repeatable? Do you actually want the way most people would probably do this, or do you want something deliberately different? Coding and math pass because you can run the answer a million times. Complex consulting projects fail all three. AlphaGo could play moves no one had played, but it had a verifier: did the move win? A brand-new enterprise strategy does not come with that loop. You can get a music model to make more songs that sound like Taylor Swift. Ask it to invent punk, or hip-hop, and you are asking for something the average of the training set cannot give you. Most enterprise work is not inventing hip-hop. The further you walk toward "not the obvious good answer," the more the model is the wrong instrument.

Martin Sorrell's version of this, from advertising, is that you do not start by making the ads. You automate the boring back-office crap, almost literally the faxes to Google to book the ads. Consultancies use the same systems to crush customer research from a week to a day: same spend, faster delivery, Evans's consumer-surplus point again. People outside an industry always invent the wrong first use. People inside it point at the unglamorous one. The same humility applies to industries that will not move. He presented on machine learning to a European water company. The CEO said innovation might be one of next year's top five priorities. In tech that sounds insane until you remember lead pipes, Huawei gear to rip out, a drought, a leaking dam, an earthquake, a rate case. Caterpillar still makes large pieces of metal. An aggregates company still digs sand, puts it on trains, and delivers it. The internet did not change that much. AI may not either, and you will not know in advance which industries those are.

Watch daily actives, not weeklies. Ask whether a user is having an accountant moment or a lawyer moment: life-changing, or useful for the timesheet. Discount the 93-percent radar charts. Accountant headcount rose for a century while the content of the job turned over completely. And remember Anselm: if you define AGI as necessarily inevitable and necessarily going to kill everyone, you have proved nothing a unicorn's existence would not also prove. John Perry Barlow's Declaration of the Independence of Cyberspace thought the network would end war because people would finally understand each other. Some of that millenarian poetry was true. Some of it was not. Meanwhile they built software.

Evans closes with a 1999 story he had not used in the token essay. He was at a bank looking at the IPO of an electronics-and-software e-commerce firm, formerly a catalog company, started in a bedroom at sixteen, a hundred million pounds of turnover, talking about floating for seven or eight hundred million pounds, insane even then. On the train back from Birmingham everyone argued about the founder and the board. A veteran banker named David Tate opened one eye and said it was a low-margin reseller, one-time sales, then went back to sleep. Sometimes, Evans says, it is that simple.