Hooray! Post-Fordism Is Finally Here!
AI, Monopoly Capitalism, and Open-weight Models
While I want this substack to mostly be about show-and-tell, I am still an economic historian. Often I am drawn into conversations about “what AI means” and so I thought it would be useful to be clear.1
As I see this debate, this question of our age, there are two main questions that history can shed some light on.2
Is AI a complement or a substitute for labor? That is, will it increase demand for and the productivity of workers, or decrease it?
Will AI be controlled by the few or be accessible to the many?
A Complement or a Substitute?
Consider a some of the most important technologies of the past 200 years.
When I am asked about what automation might look like, I inevitably discuss agriculture. Roughly all of our ancestors were farmers and approximately none of us today are. Yet we still eat bread made from wheat. That shift is possible because of automation.34
The mechanical thresher, used to process wheat, was a substitute for the most backbreaking work of the harvest. But it also enabled more land to be cultivated, and that land was cultivated more efficiently, allowing for greater harvests. Mechanization of the farm, like the thresher, turned the American Midwest into the breadbasket of the world.
Those displaced farmers found work on railroads, moving all that. And those jobs, according to people at the time, were a kind of liberation from the raw animal labor of threshing. On net, it created demand for more workers at better wages in work more fit for people than beasts. For those that remained farmers, they found other higher-value work to be done. On a farm, there is always more work to do.
The failure, then and now, is to think farmers were only threshers. That was one part of their jobs. Today, our work, for most people, is also a bundle of tasks. Workers then and now could and can focus on parts of their job that are of higher value. And in a new economy, new tasks in new industries will be created. Many of the jobs that we do today (web designer, UI expert) were simply unimaginable in 1850. That is a good thing.
Consider now the assembly line. I’m sure you all know about the staggering increases in productivity that come from the division of labor. If you take my class in industrial history, you would learn deeply about the story of the automobile. With the assembly line, and no other change in technology, car assembly went from 12 and a half hours to about 30 minutes (once they worked out the kinks). Did this reduce the demand for workers? No. It reduced the price of cars. And that increased the demand for workers, who eventually could demand even higher wages through unionization.
It is important here to realize that better tools don’t make us get paid worse. They generally make us get paid more. Why? Because the tool, without the person, is useless. Even for today’s most cutting-edge AIs, that is true. It can code, but it can only code what I imagine it to code. It can draw, but only what I imagine it to draw. That is true for AIs as it was true for the thresher.
So, I would offer that AI will create more growth, more abundance. In the long run, all growth comes from higher productivity.
I would add one more piece to this story. Economic inequality has worsened since roughly 1970. It has worsened, therefore, not in the industrial era, but the digital era. I have argued elsewhere that this happened because for decades we did not use computers as tools of automation but as glorified typewriters (and then as televisions). Our productivity did not increase, especially to justify the expense of computers. Economists have debated for decades now over the lack of increase in productivity that came with the “digital age” of computing, but it is simple. We don’t use them as computers. Now we can.
For the first time now, normal people with their normal problems can use their computers to solve and automate their problems. AI can write code. AI can automate their tedium. The digital age did not bring any gains because it had no yet arrived. We were living through the last gasp of the industrial economy.
It is now here.
This technology will unleash unimaginable productivity gains. It will level the playing field between coders and the rest of us. Coders will lose their jobs, to be sure, but for the rest of us, the bundle of workplace tasks will become much better.
And truthfully, the demand for real computer scientists will probably increase in the era of vibe-coding. Computer science itself is a bundle of skills, of which coding is just one. The more important skill – software and data architecture – will only increase in demand as the usefulness of software expands.
Monopoly or Market?
Isn’t all this talk about AI well and good, but professor, won’t it be expensive? Won’t large corporations like OpenAI just monopolize it? This fear of a “monopoly“ is that it would become a price-gouger, a tax on the rest of the economy. I do not think this fear is reasonable.
The threats of monopoly are overstated, not simply because such monopolies in the last century never persisted for more than a few years (either because of regulation or competition), but because, on the ground now, AI doesn’t seem like it is heading in that direction.
More than people realize, the software is in a competitive market, and at the same time, the hardware is coming down in price.
Software
Most people have heard of ChatGPT some have heard of Claude, and they both cost money to use. But there are a range of AIs that are open-source and can be freely used. Just go to huggingface.com (which is a silly name) and you can download hundreds of different large language models. Now, is the current version of ChatGPT better than most? Sure. But it isn’t actually better than the best open weight large language model, DeepSeek-V4-Pro-Max.
To run this model, you need very powerful hardware. Less powerful models (with lower memory requirements) can run on my Mac, as I wrote about earlier this week.
The larger point is this: the open-source models are already nearly as good as the closed-sourced models. And the open-source models of today are much, much better than the closed-sourced models of a year ago. So if you are willing to sacrifice a few months of progress, you can have the best models for free right now.
Even the best agentic coding tools can rely on open-source models. Open-source groups are pushing back on other kinds of possible monopoly moats, like agentic memory. Perhaps some other feature of AI will emerge that enables a monopoly, like a god-like artificial general intelligence (AGI), but barring that apotheosis, competition seems to be the most likely path.
Hardware
The question then is who has access to the chips? Right now, the GPUs you need to run the big models, like the new DeepSeek, are tens of thousands of dollars.
Here we come back to whether or not you think hardware will be a barrier to using the LLMs.
Historically, chip power has increased even as chip prices fall. That has been the case for decades; it is called Moore’s Law: the number of transistors per integrated circuit doubles every two years for a constant price. For CPUs, there has been some fear that it would no longer be possible, as we are starting to come up against basic physics. For GPUs, which power AI, that is not true. In fact, it is the opposite. In the last decade NVIDIA GPUs have sped up even faster than Moore’s Law. What does this mean? If they only follow Moore’s Law, AIs that currently need $10,000 chips will be exponentially cheaper in ten years—only $300.5 But they will be cheaper than that. GPU operations per dollar are doubling in more like 18 months instead of two years so that $10,000 GPU would only cost $98.
The retort—“but won’t the models be bigger?”—misses what is happening. Models are becoming smaller and more efficient. The models today are the dumbest and largest they will ever be.
Moreover, with the rise of Apple silicon chips with integrated memory, like the M5, we can easily run AI locally on consumer-grade hardware. Those AI aren’t frontier models like DeepSeek-V4-Pro-Max but they are shockingly good. You can take a big model and make it much smaller—“quantize it”—and still retain a large fraction of its capacity. These capacities will only grow in the coming years. You will not need a data center for everyday work.6
Post-Fordist Computing and Long-Tail Digital Markets
Put together we come to a very different picture of what the digital age will be. The industrial age required massive investments to build the factories to make the products that were in demand. In the digital age, in contrast, the factories to build digital products will be made by the AI on your laptop. That is not inequality. That is equality.
The physical products of the Fordist industrial age were made for the mass market. In contrast, the digital products of the post-fordist digital age will be long-tail products. I don’t need to make mass market products; I can make them for a small niche, or just for myself.
Rather than fostering inequality, AI, then, is a great equalizer. To make products for a global market you don’t need a billion-dollar factory. You just need a laptop. That is astonishing.
That said, it will not be all sunshine and rainbows. Will AI solve the inequities of capitalism or its reliance on externalities as a source of primitive accumulation?7 Probably not.
But at the same time, AI is not a normal technology in that it has the potential to radically undermine many of the tendencies to concentrate capital that we have seen in the industrial age. We have been automated out of work before, that is nothing new, but it has always concentrated capital in the hands of the few. For the first time, there is potentially an alternative path forward.
AI will bring the digital age out of the hands of the coders. AI will not widen the gap—it will bridge it. Its ubiquity will mean that AI will be a tool that nearly all of us will be able to use in our daily work, which will make ordinary people more productive and prosperous.
This essay is based on a friendly public debate from fall 2025 between me and my amazing colleague (and AI legend) Rama Chellappa at Johns Hopkins University. He is not in any way responsible for my half-baked ideas. He was far more pessimistic.
Perhaps the third question is energy and ecology. I don’t find this debate particularly interesting because the answer is so clear: disallow data centers from connecting to the legacy grid. Require them to use non-carbon energy sources. This regulation would push AI firms to innovate and lower costs in green technology like nuclear, solar, wind, and geothermal. Data centers can be located anywhere, and should be located in places where green energy is abundant, not suburban Virginia.
Our World in Data, “Share of Agriculture in Total Employment,” based on Berthold Herrendorf, Richard Rogerson, and Ákos Valentinyi, “Growth and Structural Transformation,” in Handbook of Economic Growth, vol. 2B, ed. Philippe Aghion and Steven N. Durlauf (Amsterdam: Elsevier, 2014), 855-941; U.S. historical series from Susan B. Carter et al., eds., Historical Statistics of the United States: Earliest Times to the Present, Millennial Edition (New York: Cambridge University Press, 2006), table Ba814-830, and U.S. Bureau of Economic Analysis, National Income and Product Accounts, table 6.8, “Persons Engaged in Production by Industry,” accessed April 27, 2026, https://ourworldindata.org/grapher/share-of-agriculture-in-total-employment.
The 1840, 1850, and 1860 labor-force estimates include enslaved workers. They should be read as all workers by sector. Susan B. Carter and Richard Sutch, “Labor in the United States, 1800-2000,” Historical Statistics of the United States, Millennial Edition Online, working paper version, University of California, Riverside, 2004, 10-12, https://economics.ucr.edu/papers/papers04/04-03.pdf.
That number is shocking. Here is the math. 5 doublings (2-year Moore’s Law over 10 years): $10,000 / 2^5 = $10,000 / 32 = $312.50 → “about $300”
6.67 doublings (18-month doubling over 10 years): 120 months / 18 = 6.67; 2^6.67 ≈ 101.6; $10,000 / 101.6 ≈ $98.4
A distinction here between “training” compute and “inference” computer would be useful. Training compute enables the creation of new frontier models. That kind of power requires a data center. Inference compute, where you just run a model, can be done locally. Moreover, the “fine-tuning” of a model, where you build a small “adapter” to put in front of a big model, can be done in a few hours on a Mac M4 Max right now.
I am constantly struck by how the latent resources—the externalities—of the digital age (documents, data) compare to the latent resources of the industrial age (coal, aluminum).
Jevons paradox is the deep analogy here: AI is the steam engine and training data is the coal.
The difference with the industrial age is depreciation, as you suggest. When a railroad company goes bankrupt in 1880, those rails still have value. Heck, even the "dark fiber" of Web 1.0 had value. When a data center has all A100s and goes belly up, or demand falters for a while, those chips will not have the same value. This moment is not producing the distressed assets that have been the center of so many industrial fortunes. The process of capital accumulation will look quite different.
I guess on the cultural labor front, I am not as convinced. AI slop is not good culture work. And while LLMs are good for bureaucratic communications (reading and writing), they aren't good for innovative or sustained arguments. My colleagues tell me that they will be, but I'm not convinced yet.
We shall see!
So it's Jevons paradox all the way down? Which I guess somehow also resolves Solow's paradox? Then throw in Huang's Law (where performance leaps 1000x in 10 years instead of the 32x under Moore's Law), which seems brutal for the depreciation schedules on all these new data centers.