AI Is a Platform Shift, Not an Economic Singularity

Benedict Evans’s deliberately unpopular position is that AI is “as big a deal as the internet or mobile, and only as big a deal as the internet or mobile.” The qualifier is doing the work. Evans is not minimizing AI. He is rejecting two mistakes at once: treating it as a narrow product cycle, and treating it as so unprecedented that earlier platform shifts have nothing to teach about adoption, labor displacement, distribution, and value capture.

His preferred comparison is the internet in 1997: exciting, immature, unevenly adopted, and surrounded by confident predictions that would mostly miss the eventual winners. In that frame, asking whether OpenAI has already won or whether Anthropic has the lead is like asking in 1997 whether Excite or Yahoo would own the internet. The point is not that every present company will fail. It is that the most important uses, interfaces, business models, and distribution channels may not yet exist.

Evans described the present as a wide distribution of belief, use, and capability. Inside tech, some people have reorganized their work around AI tools and “don’t use Google anymore.” Outside tech, many people who use AI do so every week or two, if at all. Even among teenagers and young adults, he said survey data in his presentation showed daily generative-AI chatbot use around 15% to 20%, with another roughly 20% using weekly and a large majority not using at all. Adoption is real, but habit formation is uneven.

That unevenness is tied to the “jagged frontier”: AI works very well for some things, unreliably for others, and it is often hard to know in advance which is which. Software developers are, in Evans’s analogy, like accountants seeing VisiCalc in the late 1970s. A spreadsheet that recalculates a week of work in 30 seconds is obviously transformative if your job is built around those calculations. But a lawyer or journalist looking at the same spreadsheet might see something clever but peripheral. The word processor mattered more to them, and it came through a separate path.

That is how Evans sees AI now. Software development is already in a “before Claude Code and after Claude Code” moment. Other professions are picking up the tools, experimenting, and often remaining puzzled about what exactly applies to them. A law firm, for example, faces not only the question of what a model can do, but how to use it without submitting hallucinated work, how it affects associate hiring, and how to redesign process around it. Those are organizational questions, not just model-capability questions.

A slide Evans used to frame the uncertainty quoted William Goldman: “No-one knows anything.” It listed failed or uncertain platform ideas around the internet and mobile — AOL, Yahoo, Pointcast, Flash, VRML, portals, i-mode, J2ME, WAP, JOYN, DVB-H, decks, and preloads — and then placed generative AI beside similar open questions: browsers, MCP, voice, app stores, wearables, and GEO. The visual point was not that all current AI ideas will fail. It was that new platforms generate many plausible theories before the durable architecture becomes clear.

The most consequential labor question, in Evans’s telling, is not “What percentage of this job can AI do?” but “Is this a task, or is this the job?” He argued that much current analysis of AI exposure confuses these two levels. Some jobs really are mostly a task that can become a button. His example was the elevator attendant: when elevators became automated, the act of driving people to floors was replaced by pressing a button. In that case, the task was the job.

But many roles are not like that. Writing code line by line may be a task inside software development, not the whole job. Making a PowerPoint deck may be a task inside consulting, not the whole job. Getting the SKU may be a task inside retail, not the whole job. Evans used Amazon as an analogy: if you know exactly which microphone stand you want, Amazon is very good at getting you the SKU. If you do not know which microphone you should buy, that is a different problem.

The same applies to AI coding tools. Claude Code may write the code, but Evans asked: what code do you want? What feature? For which customer? With what product strategy? Through what go-to-market motion? Those questions are not eliminated by making implementation cheaper.

Lenny Rachitsky noted that, contrary to a simplistic view that AI should eliminate consultants, leading AI companies are investing in professional services, consultants, and forward-deployed engineers. Evans said this is not surprising if one understands how companies actually change. Large enterprises do not have idle teams waiting to redesign workflows, connect vertical systems to horizontal systems, analyze churn, rethink stores, build new products, or train employees on new processes. Those are projects, and companies hire Bain, BCG, McKinsey, Accenture, Infosys, ad agencies, architects, and other professional-services firms because they do not keep all those capabilities permanently on staff.

Evans’s view is that AI deployment inside companies is itself such a project. Someone has to identify which workflows can be changed, which systems must be connected, what new processes must be built, how employees should use them, and what the political constraints inside the organization are. AI may change the tools used in that work; it does not remove the need for the work.

He was especially dismissive of claims that a model can simply generate “a McKinsey deck” and therefore replace McKinsey. Even if Claude could make a polished 75-slide presentation, Evans argued, the deck is not really what clients pay for. They pay consultants to walk through the company, understand why something has not already happened, map internal politics, talk to customers, and discover what is actually going on. The PowerPoint is the artifact; the job is the investigation, synthesis, and organizational intervention behind it.

This is also why Evans rejected simplistic “percentage automated” scoring of professions. He criticized attempts to break down every occupation into components and assign a percentage that AI can perform. In his view, that repeats the old expert-systems error: trying to describe a messy, contextual activity through an exhaustive hierarchy of formal steps. “You can’t look at a senior partner at a law firm and say, well, 17% of their work could be automated,” he said. “This is bullshit.”

Evans’s presentation put the issue in a more concrete form with a slide titled “Which will this be? Is this a job or a task?” The example was Philippine IT-BPM outsourcing: roughly 2 million jobs and 8% of GDP, according to the visible slide text, based on skill and income arbitrage. The question was not whether AI can perform pieces of that work. It was whether the arbitraged activity is itself the job, or whether the job contains other judgment, process, trust, and organizational work that remains.

The deeper uncertainty is that exposure often appears from unexpected angles. In 1997, a forecaster might have said newspapers would benefit from the internet by saving on printing costs, while taxi drivers had little to do with the internet beyond possible online booking. That would have missed both the damage to newspapers and the transformation of taxis through Uber. Evans used the same logic for “safe” jobs such as personal trainers: perhaps an iPhone camera and an AI-generated routine could watch a workout and correct form. That may be wrong, he said, but it illustrates why neat exposure categories are unreliable.

Evans’s labor-market argument is not that AI will cause no pain. It is that pain and displacement are normal features of major technological change, while the new jobs and new forms of demand are hard to see before they exist.

Every major technology automates away work, he said, and then creates new work through lower costs, new demand, and new capabilities. In 1800, he noted, most people were peasants and a central worry was crop failure. Since then, economies have repeatedly automated jobs and created other jobs. The disappearing jobs are easier to name because they already exist. The new ones often sound implausible beforehand: “railway engineer” is not meaningful before the railway exists.

That does not make the transition painless. Evans explicitly acknowledged “frictional pain and dislocation,” lost jobs, hollowed-out towns, and people harmed by the shift. His point is that the long-run pattern is not mass permanent idleness, but a reallocation of work around new capabilities. He placed AI inside that pattern unless there is a specific reason to believe it is different.

Rachitsky raised the more acute version of the fear: AI-lab leaders warning that entry-level white-collar jobs may disappear. Evans cautioned against treating AI-lab CEOs as authorities on labor economics. He said he is interested in Dario Amodei’s views on where models may go over the next six to twelve months, but less interested in his views on labor-market value, comparative advantage, and employment structure. Running an AI lab does not automatically confer expertise on those questions.

On the claim that AI adoption will be much faster than previous technologies, Evans partly agreed. ChatGPT can reach enormous user counts because the internet, smartphones, broadband, and cloud infrastructure already exist. Netscape could not get 900 million weekly active users at launch because there were only tens of millions of PCs in the world. AI stands on the shoulders of previous platform shifts.

But faster consumer adoption does not mean instant enterprise transformation. Evans mocked the idea that every big company will buy ChatGPT and fire staff two weeks later. Enterprise software cycles are slow; large companies do not “tear out SAP” overnight. It may take three, five, or ten years for enterprise systems and workflows to change sector by sector. Even when the enabling technology exists, the actual companies often arrive much later. Evans pointed to SaaS companies founded just before ChatGPT and asked how many could have been built in the previous 15 years. The missing ingredient was often not technology, but someone realizing that a specific industry problem existed and could be solved in a particular way.

The historical evidence he returns to is not that automation reduces work in a simple way, but that cheaper tasks often expand the surrounding market. Evans cited charts showing that the number of accountants rose through the 20th century and again in the 21st, despite adding machines, punch cards, mainframes, databases, ERP, cloud, spreadsheets, and PCs. Similarly, before modern tooling, developers wrote much more from scratch; now operating systems, libraries, IDEs, APIs, and platform services provide huge amounts of code. That did not reduce software engineering to a tenth of its former size.

Evans used an IBM advertisement from the 1950s to make the point more sharply. The ad described an IBM Electronic Calculator, before such machines were called computers, as “like having 150 EXTRA Engineers.” Evans compared that pitch to today’s claims around AI coding tools: many companies are effectively promising “150 extra engineers,” and the promise has appeared before.

The IBM slide matters because it makes Evans’s “this is totally different, just like last time” claim concrete. The visible ad copy said the calculator could speed through intricate computations so quickly that, on many complex problems, it was like having 150 extra engineers. It also said valuable engineering personnel, then in critical shortage, would no longer need to spend “priceless creative time” on routine repetitive figuring. Evans’s point was that the structure of the pitch is familiar: automation is sold as extra labor, while the economy later reorganizes around what the automation makes possible.

His supermarket example made the same argument in a different domain. A slide in Evans’s presentation showed average SKUs per US supermarket rising roughly fivefold after grocery barcodes launched in 1974. Barcodes and automated inventory management did not merely make it cheaper to manage the same store. They made it possible for supermarkets to stock far more products.

The most personal version of this was Evans’s own work as an analyst. To build a chart of supermarket SKUs before the web, he said, he would have needed to know the Food Marketing Institute existed, discover that it published the relevant data, find a library that held the reports, make phone calls, perhaps travel, and possibly spend hundreds of dollars. In his current workflow, he said, the same exploration took about two hours in Google. That does not just save time; it changes how many speculative questions an analyst can afford to ask.

Benedict Evans refused to give a confident forecast about AGI. His reason was not modesty as posture, but a claim about the absence of theory: “We have no theory of what human intelligence is. We have no theory of why these models work so well. We have no theory of how much better they will get.” In that environment, he said, people are largely “vibes forecasting.”

What can be observed, in his view, is definitional drift. He cited Larry Tesler’s line that “AI is whatever machines can’t do yet.” Once machines can do something, people relabel it as ordinary software: image recognition, sentiment analysis, voice recognition. AI is like “technology” in that it often refers to what is newly working, not what has become infrastructural and invisible.

Evans sees the same slippage happening around AGI. Some definitions now frame AGI as the ability to perform a meaningful percentage of economically valuable work. But that is a very different claim from possessing human-level general intelligence, consciousness, or personhood. By the economic-work definition, he argued, a 1970s IBM mainframe could perform work previously done by people. Databases and enterprise systems also perform economically valuable work. That does not make them alive.

He made the same point about “superintelligence,” whose relation to AGI he said has become unclear. Last year, in his telling, superintelligence sounded like something very powerful but short of “actual AGI.” Now, in some discussions, people say AGI is already here and superintelligence is the hard next step. Evans’s conclusion is not that the terms are useless, but that one must know what someone means when they use them. There may not be a single correct usage.

More importantly, Evans argued that one does not need to believe any AGI scenario to believe AI is a giant deal. If models stopped improving tomorrow, the existing technology would still be “incredibly useful” and would still roll through the economy over the next decade. In other words, the practical transformation does not depend on settling metaphysical arguments about consciousness or a final theory of intelligence.

One of Evans’s sharper claims is about where value might accrue if the market develops the way it appears to him now. He does not assume the foundation-model companies will capture most of the economics simply because they are powerful today. His conditional thesis is that model companies do not appear to have strong network effects, and without network effects or durable product differentiation, it is hard to see why they should have lasting pricing power.

Lenny Rachitsky summarized the implication directly: over time, the margins of companies such as OpenAI, Anthropic, and others may get squeezed, while larger opportunities may sit in the application layer. Evans accepted that as a deterministic version of his argument, with caveats. If models remain competitive with one another indefinitely, and if users perceive them as broadly substitutable, then the economics may look less like an operating-system monopoly and more like commodity infrastructure.

Evans used telecoms as a cautionary analogy. Mobile networks are enormously complex and valuable. Evans said global mobile revenue is about a trillion dollars a year, with mobile capex around $200 billion annually. He also said mobile data consumption has risen exponentially, roughly 1,500 to 2,000 times since 2010 by his estimate. Yet in his telling, telecom stocks have gone nowhere in 25 years because the industry became a low-growth, low-margin utility. The cool stuff happened above the network, in products and services built by others.

He applied the same skepticism to Sam Altman’s metaphor of selling intelligence by the meter, like electricity or water. Evans’s response was that utilities are not famous for attractive margin structures. The television company does not pay a percentage of the monthly bill to the electricity company; Bosch does not pay a percentage of a washing-machine sale to the power utility. If AI becomes metered infrastructure, that does not automatically mean the infrastructure owner captures the consumer surplus of every AI-powered product.

The key strategic question is whether the chatbot is the whole user experience. If users simply go to a general chatbot and ask it to do everything, model companies may own the interface. But if AI must be embedded in thousands of specific applications, workflows, interfaces, and businesses, then those applications cannot all be built by the model labs. Microsoft did not build every Windows application. Cloud providers do not capture the business value of every SaaS company running on their infrastructure.

Evans distinguished between a Windows-like platform and an AWS-like platform. Windows had leverage because customers and developers reinforced each other: users wanted Windows because the applications were there, and developers built for Windows because the users were there. Cloud does not work the same way for most buyers. A law firm or engineering firm buying software usually does not care whether it runs on AWS. If foundation models resemble cloud more than Windows, the model provider may be important without being the primary value-capture layer.

He also warned against confusing the present state of price disequilibrium with the steady state. Heavy token spending today may be like a $50,000 mobile data bill in 2010: real, but temporary. The durable question is what happens when capacity, competition, open models, local models, and product architectures settle into equilibrium. In that world, Evans asked, are there three to six companies selling a commodity near marginal cost?

This value-capture argument connects directly to distribution. Rachitsky introduced a related claim: as software becomes easier to build, distribution becomes a more valuable moat. If everyone can launch a product, attention becomes scarcer, incumbents gain an advantage, and startups face more noise. Evans agreed, especially for commodity products and “thin GPT wrappers.”

His analogy was the browser. A browser product is a thin wrapper around a rendering engine: an input box, an output box, tabs, and little else. The last major browser-design innovation, he suggested, was tabbed browsing decades ago. Once a product reaches that kind of platonic shape, differentiation gets harder and distribution matters more.

That is how he reads much of the current AI-assistant market. Sophisticated users can distinguish Gemini, Claude, ChatGPT, and other models, but ordinary users may not perceive much difference. In that situation, distribution and brand become decisive. Google can push Gemini through its surfaces. Meta, even when dismissed by technical observers, can place its AI across its products and achieve usage because the product is “not that bad” and appears everywhere. OpenAI’s strategy late last year, as Evans characterized it, looked like “everything everywhere yesterday”: a search for a flywheel, a sticky use case, and distribution before Google, Meta, Amazon, or Apple default their own AI across existing surfaces.

Apple is the unresolved case. Evans said Apple’s 2024 WWDC presentation contained the most compelling vision he had seen of a personal AI assistant: tool-using, agentic, on-device, integrated across apps, and built around standardized intents. But Apple did not ship it, and Evans noted that nobody else had shipped that complete vision either. The difficulty is real: no prompt injection, no hallucinations, standardized APIs across thousands of apps, and reliable execution is a high bar.

If Apple eventually powers its assistant with Gemini, that would reinforce Evans’s commodity-model thesis. The model would be the underlying capability; the product, distribution, interface, and ecosystem integration would belong to Apple. Android might have Gemini Intelligence, iOS might have Apple Intelligence powered by Gemini, and they would still be different products because the model is not the whole product.

The same logic extends to incumbents more broadly. Evans said one lesson from mobile is that some incumbents missed the shift entirely, while others were barely hurt or were helped. Google, in his simplified account, did not fundamentally suffer from mobile. Meta benefited because phones were better for social: cameras, notifications, constant presence. Amazon’s core business did not change much. Yahoo failed to jump. The lesson is not “all incumbents win” or “all incumbents die,” but that each platform shift affects different companies differently. As former Windows leader Steven Sinofsky put it, according to Evans, incumbents always try to make the new thing a feature. Sometimes they are right.

Benedict Evans described anti-AI sentiment as a “big fuzzy mess” of distinct concerns. Some are tangible. Some are exaggerated. Some are culturally or politically loaded. Many are difficult to measure because the field lacks good data.

On data centers, he separated electricity and water. Electricity-price concerns may be real in some places, though he said they apply objectively in a small number of locations. Water, in his view, is often misrepresented. Data centers use water for cooling, mostly in closed loops, and Evans said he had found a Livermore Lab estimate placing US data-center water consumption at about 0.017% of total US water consumption. If a small town has one well and a data center gets access to it, that is a legitimate local planning problem. But he rejected the broader claim that data centers are a major national water-use issue.

On jobs, Evans said economists do not yet have a clear consensus. There are charts suggesting employment effects and charts suggesting otherwise. Youth employment has slowed, but he said the slowdown appears across people with and without degrees and across fields that do and do not look AI-exposed. That makes attribution hard. Students may be struggling to find jobs, but whether that is because of AI, tariffs, politics, macro conditions, or something else is not settled.

A related problem is that AI companies disclose very little useful usage data. Evans emphasized that, in his view, there is no meaningful daily-active-user number for ChatGPT. Model labs release selective studies and broad figures, while much of the serious measurement comes from academic economists backing into patterns through government surveys, or consultancies and agencies surveying large panels. For a technology this important, he argued, the empirical base is surprisingly thin.

Other backlash comes from creative fields. Evans mentioned illustrators, novelists, ebook authors, and the culture war over whether using AI in creative production is acceptable. He also noted “AI slop” and claims that a large share of new podcasts are AI-generated. His tone was sometimes caustic, but he did not dismiss the entire category. Instead he compared the moment to the backlash against social media, only more compressed. Some claims about social were true, some partly true, some false but fiercely believed. AI is producing a similar spread.

The social-media analogy also shaped his treatment of harm. Evans argued that every major information technology creates new ways to ruin people’s lives, deliberately or accidentally. Deepfake nude imagery is not answered by saying Photoshop already existed. A 15-year-old could not previously make explicit images of every girl in a school and distribute them in an afternoon; now that is much more feasible. That difference matters.

He used the UK Post Office scandal to make a narrower but important point: institutional harm from technology is not unique to AI. In his account, a Fujitsu-built point-of-sale system falsely showed cash shortfalls in franchise post offices, and the consequences included prosecutions, bankruptcies, suicides, and lost homes while officials insisted the system had no bugs. Evans called it 1970s-style technology. His point was that every wave of technology creates systems that can harm people at scale, and institutions must be conscious of that without collapsing into panic.

Evans’s most productive strategic frame is not replacement but redefinition. New technologies first let people do old things more cheaply or at greater volume. Then they enable things that were previously impractical. Eventually, they can change the question entirely.

He used recorded music to illustrate this. Global recorded-music revenue, adjusted for inflation, fell by about half from 2000 to roughly 2015, then recovered to about 75% of the prior peak, driven by streaming, according to Evans’s account of the chart. Evans interpreted the first half of that chart as the question, “What if I don’t have to pay $15 to get a CD to get that track?” The second half asks a different question: “What if $15 a month gets you all the music there is?” Spotify is not simply an online music store. It is a different product category made possible by the new distribution and pricing structure.

That is the direction he thinks AI analysis must move. The first stage is “the old thing, but more”: print out your emails, put Flickr on mobile, generate a deck faster, write more code. The more important stage asks what was not possible before, what becomes unlocked, and what business model breaks or opens because of it.

His Uber and Airbnb comparison captures the danger of assuming one pattern. Both are software-enabled marketplace companies often invoked as examples of software eating the world. But Evans said Uber demolished the taxi business in many cities and expanded the market, while Airbnb’s impact on hotels is much more marginal in the numbers he has seen. Airbnb created a large adjacent business and may have slowed hotel growth, but business travel still requires hotels for reasons Airbnb does not solve: late arrival, room service, baths, gyms, predictable proximity to client sites. Sometimes software eats the world; sometimes it only nibbles.

That is why Evans embraces “it depends” as more than a verbal tic. He sees it as intellectual discipline. AI’s impact will vary by task, job, company, industry, customer behavior, distribution, regulation, trust, and workflow. The right posture is not paralysis, but “presume radical uncertainty.”

Evans’s advice to people worried about their careers is blunt: do not stick your head in the sand. He pushed back when Rachitsky summarized his view as broadly comforting. In the long run, he said, broad averages can hide real individual pain. If a law firm hired 100 associates last year and hires 50 this year, the affected people cannot console themselves with a 200-year productivity chart. Professional-services pyramid structures may change in ways that are still unclear.

But rejection is not a strategy. Evans said hating AI may provide moral superiority and a community of people shouting about it, but it will not help someone become employable or effective. The practical move is to “dive into this,” internalize what the tools can and cannot do, and understand how they change one’s work. A candidate interviewing at a law firm that is rethinking associate hiring is unlikely to help themselves by saying AI is nonsense and they refuse to use it.

This does not mean every person must become an AI maximalist. Evans himself is not an unusually heavy AI user in his own work. He described himself as the lawyer looking at the spreadsheet: he can see the technology is transformative, but many of his own tasks involve precise information retrieval, where current AI is weak. He uses AI for proofreading and images, including apartment redecoration mockups. He dictates much of his writing into Apple Notes, which automatically transcribes it, though he noted that at some point voice recognition simply becomes automation and people stop calling it AI.

His own difficulty finding daily use cases is part of the broader point. The chatbot is a blank screen with a jagged edge. Users often do not know what to ask, what will work, or where the tool will fail. The solution is not only better models, but wrappers, use cases, workflows, interfaces, and products that make capabilities legible and reliable.

For children and career planning, Evans declined to offer a deterministic map. Someone entering the labor market in the next year or two faces more uncertainty than a younger teenager, because the near-term hiring structures are unsettled. For a child entering the workforce in five years, the world may have stabilized in unpredictable ways. His broader career advice is that people discover three things over time: the skills they have, the jobs those skills make them good at, and the things people will pay them for. Ideally, a career finds all three; at minimum, it needs two.