A first principle: Be humble in looking far ahead as technology races forward, expanding the range of choices for economy and society. Back in 1875, perhaps you could have predicted the advent of the automobile. But the traffic jam? Sprawling urban forms locked into automobile dependence? The decline of public transit? Suburban segregation by class and race? These were all unintended consequences, what analysts call the “emergent properties” of a system nobody foresaw.

There are plainly limits to our ability to pierce the veil and defeat the dark forces of time and ignorance. And I dare not try to predict what the onrush of AI technologies will bring to society beyond the next decade. That said, suppose we project what society and technology will be like in just a decade, and then look back at today. What might we see?

Start with the relatively easy piece of the puzzle: will investors in today’s platform behemoths that are betting so heavily on artificial intelligence be disappointed?

I expect that the near-monopoly digital platform giants – Google, Facebook, Amazon, Microsoft, Oracle, Apple and company – are, with the possible exception of Apple, likely to come up short.

That doesn’t mean the data centers will be reminders of the “dark fiber” left stranded after the dot-com crash in 2000 before its long recovery. The computational capacity can always be used to render ever-higher-resolution virtual worlds for computer games, or some emergent high-value task. But they will almost surely not justify the investment in terms of earnings per share.

Why don’t the executives of today’s platform monopolists see this, too?

I think they do. I suspect they haven’t pulled the plug because they all see themselves in the same position as Microsoft back in 1995 when Bill Gates wrote his “Internet Tidal Wave” memo, demanding that Microsoft strain every nerve and muscle to make Office and Windows the best ways to write to and read from the internet at no additional cost. This time, it is the coming of AI now playing a role parallel to the internet and the web browser in the late 1990s.

Back then, Marc Andreesen, the founder of Netscape, predicted that users would shift their window on cyberspace from the graphical user interface operating system of a PC to the tab of a web browser, famously adding (quoting 3Com founder Bob Metcalfe) that the internet would reduce Microsoft Windows to “a poorly debugged set of device drivers.”

That putdown is not fair: Microsoft did as Bill Gates had demanded. What Microsoft objected to was the notion of anybody making money by disrupting its platform monopolies built on Windows and Office. So Microsoft’s managers deployed a slew of programmers to write a superior web browser (Internet Explorer) that was integrated into Windows and Office, and then gave it away – precisely so that Netscape would never make a dime on its competing Netscape Navigator or gain a toehold on Microsoft’s turf.

As then, so now, are the platform giants throwing vast sums at challengers real and imagined. Thus I expect that by 2037, today’s platform monopolists will have spent fortunes to make it easy and cheap for us to use the AI they provide. Need a natural-language interface between you and any database or information store or set of bureaucratic requirements? You’ll get that automatically for little or no money, the way Microsoft defensively added freebees like Internet Explorer and much more to its operating system.

Need any sort of very big-data, high-dimension, flexible-function classification analysis to understand pretty much any problem you face? You’ll get that automatically, too, and for pennies. AI will become part of the background, like the air we breathe.

Back in the mid-1990s, we had one software platform monopolist: Microsoft. Now we have six: Google, Facebook, Amazon, Microsoft, Apple and Oracle. They are all sitting pretty with high profit margins because they face little effective competition for their core businesses. But they do fear AI and the disruption its build-out might bring, just as Microsoft feared the internet in the late 1990s.

In particular, the platforms fear that somebody will come up with a better AI that can thrive sitting between their systems and their users. That AI would absorb the users’ natural-language queries and be capable of analyses far superior to what they could get directly from the platforms. And that sooner or later the platform monopolists would be left to beg the AI provider for the value delivered.

Note that it does not matter whether this threat is real. What matters is that all of the platform monopolists think it is possible. Their primary goal will shift to making sure that as little as possible of the value of AI services passes through any market at all, but rather is rolled into their existing revenue streams or given directly away as user surplus. Someone else who makes money by selling AI services to users of their platforms is someone who has the money and the market position to mount a competitive challenge.

Are investors plowing an estimated $5 trillion by 2030 into AI and data centers likely to be rewarded for their big bets? Will those investments yield the $250 billion return needed to justify the plunge in conventional investment terms? Unlikely. And it’s especially unlikely to yield bright financial futures for OpenAI, Anthropic and Elon Musk’s xAI, which have to raise money on the capital markets in order to fund their own training, inference and data center build-outs.

OpenAI’s original plan was not implausible. The company assumed its overnight success in building ChatGPT and gaining a user base was a sign it had a durable lead in LLMs that it could build on. Remember, it was OpenAI’s apparent superiority at building increasingly better LLMs that led Microsoft to woo it, saying that the partnership would force Google to “get up and dance.” That turned out not to be true.

Now, OpenAI has moved on to Plan B. With ChatGPT dominating consumer visibility, it sees itself as the closest thing to a mass-market standalone AI business. But what is market share worth if “pretty good AI” becomes something everyone gets for nothing, like the toy in the Crackerjack box? OpenAI counters with the example of Google and its world-changing search engine, arguing it can also create a very profitable business on the foundation of a killer app. If this is indeed what OpenAI is betting on, it needs to start selling ads. Lots of ads.

I think it’s likely that OpenAI in its for-profit corporate form (as opposed to its research foundation form) is likely to be broken up for parts and sold off – the fate of Netscape when it proved unable to make money competing with Microsoft’s free Internet Explorer browser.

Google did, of course, survive, winding up with a very large part of Microsoft’s treasure hoard. But by the time Google came along, Microsoft no longer saw competing browsers as a threat to its survival, and certainly was afraid to tangle again with antitrust regulators after its narrow escape from breakup in 2002. Hence Microsoft’s search-engine business was assigned the mission of merely keeping Google contained, rather than (in Bill Gates’ words) “choking off their air supply.”

A parallel shift to détente between the platform monopolists and one or more of the standalone AI startups is a possibility. But I see no signs of that emerging yet.

Right now, Anthropic appears in a better position than OpenAI. It is headed for a richly valued IPO by the end of 2026, based on two strengths. First, its enterprise business is already on the cusp of profitability, offering the best-in-class agentic programming assistant, Claude Code, and agentic-enterprise workflow assistant, Claude CoWork.

Competition from nearly free, very good general-purpose chatbots damages the willingness to pay for a general-purpose chatbot from Anthropic. But that competition matters less in determining the willingness to pay real money for an AI system that enables highly paid engineers to complete projects in half the time.

The economics of data centers is highly uncertain. Researchers at nonprofit EpochAI estimate that you can operate a data center which requires one gigawatt of power for less than $10 billion a year. And right now, Anthropic is paying $15 billion a year to rent xAI’s Colossus I one-gigawatt data center – apparently at quite a profit for xAI. If demand holds up, Anthropic will be more than covering its costs.

Anthropic therefore may yet become at least another Oracle or Bloomberg, with its product being not just another LLM producing tokenized natural-language output like ChatGPT or Google’s Gemini, but rather a trusted system for performing high-value work. The key is making Claude part of the standard toolkit of software companies, banks, pharmaceutical firms, consultancies and governments.

The second part of the Anthropic success hypothesis is that true frontier excellence in software coding assistants is a superpower, because superiority at coding translates directly into superiority at training the next generation of LLMs to be even better at coding and other tasks. With luck, this leads to a virtuous circle of recursive self-improvement in which the AI model, working much faster than any group of humans could, explores many more possible paths for making much better versions of itself than any group working “by hand” could. That means the next generation of Anthropic models would have an even bigger lead – and not just in coding but in knowledge work and information management in general.

Perhaps there is a third reason to declare advantage, Anthropic. If Anthropic gets through its IPO successfully, it may be able to pull off a détente with the big platforms using its investment in new data center capacity as an important bargaining chip.

Not every roads lead to Fort Knox, though. Just as it proved easier to copy what had been thought to have been ChatGPT’s secret sauce in designing LLMs, so it has proven easier than expected to copy the elements of an AI harness that keeps an LLM on task. Plus, it has proven much easier than I expected to make a small AI model behave like a large one via what’s called “model distillation.”

The clues here: a substantial number of very solid “distilled” models coming out of China, plus the reality that Apple has apparently managed to drastically improve its Siri on-device AI models via distillation of Google Gemini.

The technology continues to look miraculous. But extraordinary profits commensurate with the scale of investment require passing the point where value increases exponentially while unit costs do not, with some secret sauce allowing the technology leader to translate that into sustainable profits. Getting there doesn’t look easy.

Advances in AI these days seem to be coming from:

• larger models, but with those advantages over “distilled” competitors growing no larger
• more data, but there are no more data to be had on any large scale
• thinking with insight further down the forking paths of future possibilities.

Note that none of this points toward a way to generate the sort of economies of scale that protect the digital platform monopolies. Run four times as large a model through five times as many branches looking one more step ahead and you have spent 100 times as much in electricity, in cooling and in chip depreciation.

Looking back from 2037, perhaps the most likely outcome is that everyone will agree that the underlying economic story was always straightforward. AI intelligence was always unlikely to remain a simple stand-alone commodity, sold by the token at very high margins, because the platform-monopoly companies benefitted too much from making and keeping intelligence cheap. And so the economic “rent” or surplus generated by this marvelous technology migrated elsewhere.

Surely some migrated upward into profits for companies producing enormously capable frontier models. Surely more migrated outward into enterprise workflows and proprietary data. Apple hopes it will turn out that some surplus migrates downward into extraordinarily efficient distilled Gemini models that run on iPhones. And some surplus, but probably not a lot, will migrate back to the platform monopolists.

Was there time for OpenAI to become a platform of its own before the big digital platforms turned chat into a free feature? I do not see that as a likely outcome for 2037. Was there time for Anthropic to become a trusted enterprise integration provider before Microsoft’s harnesses running on whichever cloud or local LLM you chose became frontier-level? I do not see that as likely in 2037.

There are still plenty of imponderables with respect to what we will see when we look back from 2037. Did Elon Musk pull another rabbit out of his hat, mixing up the competitive landscape with Grok? Did Claude’s recursive self-improvement ship come in, and did that make Claude indispensable rather than merely impressive? However this all goes, it will seem obvious in 2037 that using machines to search all human wisdom written down since the pyramids were built was destined to become essentially free.

AI and Human Productivity

As explained in detail above, I am definitely a Debbie Downer with respect to the prospects for platform monopolists or AI frontier “labs” making colossal fortunes from LLMs and other modern advanced machine-learning models. But that does not mean I am inclined to dismiss the socioeconomic impact of the underlying technologies.

I believe we will look back in a decade and marvel at the massive gains in terms of user utility that AI technologies will have delivered. For example, children who are far from rich will find that a personalized AI tutor is there to be spun up at a moment’s notice, a virtual Socrates for any and every discipline at every and any level. They will be able to learn algebra, physics or history at a pace and depth that equals what I got from Sidwell Friends, the much-celebrated private high school in Washington. The cost? Near zero, once the infrastructure is in place.

Health care will be another big winner. Patients with next to no ability to pay will find that their symptoms can be classified by an AI diagnostician that has read every medical journal, every case study and every treatment protocol, mapped the symptoms to every symptom configuration, and can recommend the needed tests and the array of effective treatment protocols that should follow. And at least the incremental cost of the best pharmaceutical treatments will also be close to zero, provided the systems (and the right financial incentives) are in place.

When contemplating the potential benefits of AI, there is an understandable tendency to ignore what we don’t (and perhaps can’t) know: how it will affect creativity. But we do know that where the question comes down to trying out variation after variation to solve a problem, access to AI will get us there quicker.

Such gains will not show up in productivity statistics any more than Wikipedia’s impact shows up in productivity statistics. For, as Google’s chief economist, Hal Varian, noted a decade ago, economists’ national accounting techniques do not have a way to handle the value of goods that no one pays for. A technology that saves you an afternoon at the library or answers a question a consultant would once have billed for generates no receipt and so generates no recorded output. But the value is real enough.

Today everyone has access to Wikipedia, which is a truly wonderful compendium of human knowledge and one that is almost always reliable. If, three decades ago, you had wished for access to something like today’s Wikipedia at your fingertips, your best option would have been the Encyclopaedia Britannica, which weighed 125 pounds and ran around $5,000 at today’s prices.

When knowledge was scarce and only available for a price, it showed up in GDP. But when the same service is free, the measured economy registers not a gain but a hole. Wikipedia thus made the world richer and the national income accounts poorer. AI, deployed the way the platforms intend to deploy it, will do the same thing across law, medicine, tutoring, translation, coding and research.

Yes, But…

These things are not magic. Modern advanced machine learning models are not artificial general intelligences. They are not autonomous agents with their own goals, desires or capacity for independent planning. And I see no path for them to become so.

They are accurate emulations of the next word in a conversation you are having. The intelligence behind them comes from the millions of human minds whose thoughts were encoded in their training data. Each small unit of text (or token) processed is a choice made by a human. And the aggregate is intelligent in the same way that Google search or the University of California library is intelligent – as a pattern-recognition engine built from human knowledge, not as a new form of mind.

They can thus be superb as search-and-summarization engines over the store of human knowledge. They can try a thousand things in the time it takes you to try one, and as long as there is a gate at the end of what it tries that rejects the bad and passes the good, they can work miracles.

But we are not facing an existential threat from machine superintelligence in the next decade. The models are tools for augmentation, not replacement of human cognition. They excel at pattern matching, information retrieval and generating plausible text. But they lack genuine understanding or intentionality, save that the human who used the word it chose had understanding and intentionality in the conversation captured in the training data.

With that clear, turn to the questions of how it is likely to transform incomes, labor productivity and organizational structure over the next decade.

What People May Be Willing to Pay For

The benefits from natural-language LLMs are going to be very high – and close to being obvious. The old joke and the old major friction with information technology was always that you could not ask the computer to “do what I mean.” Now you can.

But actions in the economy are incentivized by money flows. And when we look back on the decade from the vantage of 2037, what will we see that people have been willing to pay for? So far, the evidence on AI’s productivity impact is mixed but suggestive, with major real benefits concentrated in a few areas.

Customer service and support. Contact centers deploying AI chatbots report 30-50 percent reductions in human-agent workload for routine inquiries, with customer satisfaction either unchanged or slightly improved for simple queries but worse for complex, emotionally nuanced issues.

Software development. The big one so far – although contested. AI coding copilots have demonstrated 20-40 percent productivity gains for experienced developers, primarily by reducing boilerplate code generation and debugging time. But these gains are concentrated among senior engineers. Junior developers sometimes actually lose ground because they rely on AI suggestions without fully understanding the underlying code.

Content creation and knowledge work. Writers, analysts and researchers report remarkable two- to three-fold speedups in draft generation, literature reviews and data synthesis. But the quality of outputs is variable and human editing remains essential, often consuming as much time as the original AI generation. AI does not replace even a single human with more than a typing-pool level of autonomy.

Scientific research. In specialized domains like materials science, drug discovery and protein folding for a variety of niches in biotechnology, AI has already accelerated progress by factors of 10-to-100. The protein-structure prediction example is illustrative: what once took a PhD student a decade can now be done in hours. But these are narrow applications; general scientific discovery remains human-driven.

What’s Worrisome

All that said, the higher knowledge-worker productivity promised by AI is now generating the anxiety of the moment. We elite knowledge workers are – or at least fear we are – about to be in the bulls-eye of Schumpeterian creative destruction.

For two centuries, we have been the beneficiaries of technological progress. The steam engine shredded the lives of the handloom weavers, but it created the engineers, the managers and the corporate executives. The automobile destroyed the horse-and-buggy industry, but created opportunities for auto workers, suburban developers and highway contractors. The computer turned typists and data-entry clerks into endangered species, but it created software engineers, data scientists and product managers.

This time around, though, the technology is designed to do the things that we (the knowledge workers) now do. AI can write code, analyze legal documents, diagnose diseases, compose music, write newspaper articles, manage projects and teach classes. Hence we are no longer the likely beneficiaries of technological change, but the victims.

Well, let me hedge a bit. AI will not destroy elite knowledge work, it will transform it. The lawyers who use AI for document review will be able to handle more cases and serve more clients. The physicians who use AI for diagnosis will make fewer mistakes and achieve better outcomes. Academic researchers will be able to explore more questions, more deeply.

But most of the impact will be on entry-level positions – the junior lawyers doing document review, the junior analysts doing financial modeling, the graduate students doing literature reviews. These knowledge workers will find that the traditional pathways to success have been blocked. The result will be a mixed bag, with a small number of elite knowledge workers amplified by AI doing the work that once required thousands. The rest will be forced into roles less vulnerable to productivity amplification via AI.

How large will the new cognitive knowledge-worker elite be? One hears invocations of Jevons’s Paradox, misused to imply that there will be more demand for high-skill high-value knowledge workers when their productivity rises. Just as the coming of the microcomputer, the internet and the smartphone quadrupled the number of software developers and programmers, so should the demand for knowledge workers rise. Or so the argument goes.

Maybe and maybe not. In any event, there will be a real status inversion linked to the diffusion of AI. The old elite, the ones who built their careers on the scarcity of knowledge and the difficulty of access to it, will find their power eroded. The new elite will be those who can leverage AI, who can manage human-AI teams and who can navigate the new landscape of information and creativity.

Labor Displacement

First the good news: AI will not cause mass unemployment in the next decade. It will not even lead to large-scale occupational disruption or increase the divide between good and bad jobs. But it may well trigger those outcomes over a longer term. At greatest risk:

Customer-facing knowledge work. Call-center agents, travel agents, insurance underwriters and claims adjusters. AI chatbots and automated decision systems can already handle 60-80 percent of these cases.

Creative and media production. Graphic designers, copywriters, translators and even some journalists. AI tools can generate drafts, images and translations at a fraction of the cost, though human oversight remains essential for sustaining quality.

Health care. Think coders billing insurance, and radiologists and pathologists analyzing images and slides. AI is already matching or exceeding human performance on many diagnostic tasks.

Survivors in these labor-market segments will morph to become data curators, AI model trainers, reviewers of complex and edge cases, and the like. And presumably, AI will also generate some jobs we cannot yet imagine, just as in the early 1990s we could not see the place for social media managers or gig economy platform operators.

Job polarization is likely unavoidable, a continuation of a trend that began in the 1980s. High-skill workers (those who can design, manage and deploy AI) will see wages rise. Low-skill service workers (caregiving, food service, manual labor) will be largely insulated from AI displacement, but will face wage-depressing competition from an expanding pool of displaced middle-skill workers. Indeed, one can expect middle-skill, white-collar workers to be squeezed from both sides.

The chaos of creative destruction is not just economic, but also social, political and cultural. When one-fifth of the economy is sliced and diced across a generation or even a decade, when entire professions are transformed, when the pathways to success are rewritten, the result is disorientation, anxiety and resentment.

Actually, we’ve been here before. The shift to steam power in the early 19th century created the Luddites, who tried to smash the machines they saw as the source of their suffering. By the same token, the shift to science-driven industry and global supply chains that radically revised production has opened the door to demagogues and fascists offering simple answers to complex labor-displacement problems.

We are already seeing the beginning of the backlash to AI. And perhaps this time it will be more intense because the elite won’t go quietly into that good night. They built their careers on knowledge and credentials, and now those credentials are being devalued.

End Game

Come 2037, will we be able to see the impact of AI on the sustainable pace of technological change? Time for me to hedge once more. Certainly the vibe is that new general-purpose technologies are arriving faster than in the past. But the path is often hard to discern. The economic historian Paul David famously illustrated the point with both the electrification of industry and the role of the PC in office work.

Since the mid-1870s we have seen that it takes roughly 30 years for a wave of new general-purpose technologies to wash over the economy. Each saw a rough doubling of potential productivity, but with qualitative changes as important as any quantitative benchmark. Each saw roughly one-fifth of the economy fundamentally dislocated, with enormous wealth and prosperity for those who adapted well and catastrophe for those in the bulls-eye of Schumpeterian creative destruction who try to hang on to the old ways.

This rhythm of technological revolution and associated creative destruction was not a new thing as of 1875. The previous century had been rocked not only by the steam engine and automated textile machinery, but also by a host of displacing innovations ranging from railroads to mass print media to cheap, reliable ocean shipping that linked the wheat fields of North Dakota to the flour mills of continental Europe.

Roughly one in seven jobs were destroyed in the process. Those displaced did find other ways to eke out a living, but there was massive scrambling of status and lots of uncertainty along the way. Thus the prosperous handloom weavers of England did not vanish, but most were pauperized and forced into the ranks of the urban poor. Their children perhaps found better opportunities as the rising tide of productivity was translated into higher average wages. But there was no happy ending for most of the direct casualties of progress.

The overturning of hand-textile work, however, took 100 years, not 30. And that pace of dislocation was arguably more manageable than the succession of 30-year waves of technological change and creative destruction after 1875. Even this period was a time of confidence in progress, of the belief that change could be managed largely peacefully. Only the yet-faster pace thereafter that brought us Leninism, Nazism, Neoliberalism and Trumpism was able to thoroughly disrupt the mixed-economy, social-democratic order.

My guess, however, is that the AI wave will be the most disruptive to date. The difference this time is that knowledge work is in the eye of the coming storm. We knowledge workers fear we are about to find our skills about as relevant to the contemporary economy as the specialized skills of Britain’s textile labor aristocracy after 1775. Their protests yielded the ineffectual army of General Ludd. Today’s knowledge workers are capable of making a lot more noise than the hand-loom weavers.

If I am right about the pace and focus of AI disruption, it means that, on the one hand, the productivity upside from AI will be smaller than hoped. It may not translate into much more than the 2 percent average annual growth in productivity sustained since 1875 or so, rather than the 6 percent enjoyed in the contemporary transformation of China into a middle-income country.

One can hope that the century and a half of experience in dealing with the disruption associated with a 2 percent pace of technological change will give us greater capacity to ride out the latest wave. If I am wrong, if this next technological wave transforms economic production as rapidly as some hope and many fear, we are in for a wild ride.

Defending the Better Angels of Our Nature

Even the AI Pollyannas these days are inclined to agree that a warning is in order. The potential gains in terms of human capabilities are just that: potential. We are quite capable of screwing this up.

Consider the difference between Wikipedia as an information source that makes it easy to find out what you need to know about the world, and the rage-bait slop that we see today on a typical Google search-results page or a Facebook feed. Imagine that you could time travel back to 2006 and show Mark Zuckerberg or Sergey Brin what their websites are showing us now, and then imagine their horror. The metaphor of frogs being slowly boiled comes to mind – but at risk of mixing the metaphor, boiled in lots and lots of money.

As an old-school weblogger, I am perhaps not objective on this issue. But there’s no disputing that information technologies are edging closer to becoming our masters rather than our servants. In the world of easily accessible information we yearn for, content providers seek to inform, educate and entertain, and are rewarded by being paid by users who judge that they do it well. But in the world we live in, we are at risk from algorithms optimized not for informing or educating or even entertaining, but rather for harvesting eyeballs and eardrums to be sold to advertisers and demagogues.

The hope is that, come 2037, we will look back and conclude that, just as email spam and malware were purged from our computers by proper security scrubs once it became in producers’ interests to do so, so, too, did attention-hacking fall victim to artificial intelligence’s ability to provide people with an info-butler that stands between them and those who would want to hack their attention to their detriment.

Economists see lots of user surplus from AI in our future. But all the time and attention that people spend using AI systems may not make them happier or better able to achieve their goals. Much depends on whether they are using AI systems to enhance their reach – or whether they are being used as fodder by systems of attention-hacking slop.