Excellent! I kept waiting for the final thought. The more it increases productivity, the more in will increase the value of land. Production, minus wages & interest = rent. Thanks for writing it.
From the China angle, your consumer-surplus world looks more reachable there than in the US, on three of your four candidates.
Take the labs. The pricing power competition is eroding in America has largely gone already in China. Open weights are the baseline: DeepSeek, Qwen, Kimi, GLM and MiniMax all ship strong models anyone can run, and the price war has pushed inference close to zero. Under export controls no Chinese lab has a path to a durable frontier lead anyway, so the sensible move is to give the model away and compete for ecosystem share. Commoditization is the default equilibrium, squarely in your pens world.
Hardware is your strongest petroleum case, but the rent lands on NVIDIA, ASML, TSMC and SK Hynix, all outside China. There the chokepoint is a pure cost, with no domestic gatekeeper to hold the public hostage, and Beijing is pouring money into commoditizing the layer (Ascend, SMIC, domestic HBM).
Energy is the American bottleneck; China's constraint is chips. It keeps adding generation fast enough to keep power cheap, with the grid state owned. The resource rent that looks most like oil is the one China is least exposed to.
That leaves land. A productivity boom capitalizes into land, and in the US that means private owners in San Francisco and Seoul collecting the windfall. Urban land in China is state owned, and households hold time-limited leaseholds, so the state already sits where your private landlord sits. What it does with that rent stays an open question, but the starting point sits far closer to the Land Value Return you want than anything in America.
So China looks more likely to reach the consumer-surplus world: three of your four rents are structurally weaker there, and the fourth sits in public hands.
I will say the land analysis is a little more complicated than it appears at first glance, because although all the land is nominally state-owned in China, China has been extremely resistant to implementing an actual property tax, and local governments have gotten used to financing themselves with “land leases” that are functionally equivalent to land sales. Meanwhile the US is a bit of a patchwork on that front — California has done the most to undermine its property tax, enabling huge private capture of land rents, while in states like Texas the property taxes are actually quite high, leading to more public land rent recapture.
That said, the Chinese government could change the arrangement anytime, and is on a much firmer legal footing to do so should they desire.
This is just a fantastic explanation of free markets vs rent seeking. It would be tremendous if you turned these articles into short-form TikTok videos. The youth really need someone to explain these concepts simply.
You touch upon regulatory rents, but in this post you do not consider the biggest and most pervasive regulatory rent of all: intellectual property monopolies. Intellectual property creates long lived (even if ultimately temporary) monopolies, restricts competition (due to the monopoly), creates property rights over what is non-rivalrous, and restricts typical common law property rights over what is rivalrous for everyone else. This regulatory rent is pervasive throughout all industries (over 90% of fortune 500 assets are non-tangibles) and this includes both the AI and AI Hardware Manufacturing industries
IP law in the US is converging on a human-authorship requirement for protectable works, meaning work done by general AI models is not protected. I don't think frontier AI labs realize that, even if they are the first to AGI, all of AGI's new discoveries and inventions will default to the commons under existing law. I've written about this issue here:
I like Harbinger to self-assess fixed improvements when combined with an auction for occupancy rights to a location but am not familiar with it in the context of IP. What are your thoughts?
Yeah, I think what regulators are seeing in AI (at least the foundational models) is a natural monopoly akin to the utility and rail industries. Foundational models require astronomically high upfront costs, with lower (though still high) operating costs. The high barrier to entry gives them pricing power that allows them to collect a windfall. Then the question would be 'is AI like internet service providers'? I actually read a policy paper from the Hamilton Center on Industrial Strategy that concluded that AI foundation models were natural monopolies, but the social cost to the monopoly was too low for regulation to be a significant benefit.
I mean the question is, even if it requires astronomically high upfront costs to produce a frontier model, once it’s produced, so far evidence shows us that they are rapidly distilled. Also, China can afford high upfront costs too. Is there a plausible story here for a frontier model taking the lead and durably keeping it without having to constantly plow forward?
Hmm, I don’t know enough about distillation and how much better the frontier model is from a distilled model, but like you said with China’s open-weight models, they do decrease the pricing power of the frontier models. If the AI models become commodified, then I guess the hyperscalers/cloud-providers would be the natural monopolies in this analogy? But I agree with the thesis of the article, the biggest issue will be the ground rent landowners will extract from the increase in productivity.
Yeah I don't doubt that a moat can hold up for a few years, maybe even a decade or longer. But I don't think they hold up forever under sustained competition. Someone needs a thesis for how they perpetually fend off competition (the ASI ascension thesis is basically this)
I misspoke, it was the RAND Corporation not Hamilton Center on Industrial Strategy. But here's their conclusion:
"Application of the natural monopoly criteria to the status quo foundation language model market (as of January 2024) indicates that the current case for a natural monopoly is relatively strong. This conclusion is based on the observations that the current generation of foundation models is reasonably homogeneous, economies of scale are high, costs are largely sunk, and network effects and economies of scope are present."
Though in all fairness, there are newer studies that claim Reinforcement Learning AI, doesn't have the same features.
This is a great piece, but I think the petroleum analogy could use some refinement. Petrostates can maintain consistently incredible profits from oil because they have very low production costs and because they collude through OPEC to restrict the supply of oil. The remaining oil demand is met by high-cost countries like the US and Canada and here the profits are much smaller. On average, oil companies are more profitable than pen companies, but it is also much more volatile, so this extra profit is mostly just a risk premium. Some years oil companies will be very profitable, and some years, such as 2015 and 2020, they will lose tons of money.
Excellent! I kept waiting for the final thought. The more it increases productivity, the more in will increase the value of land. Production, minus wages & interest = rent. Thanks for writing it.
From the China angle, your consumer-surplus world looks more reachable there than in the US, on three of your four candidates.
Take the labs. The pricing power competition is eroding in America has largely gone already in China. Open weights are the baseline: DeepSeek, Qwen, Kimi, GLM and MiniMax all ship strong models anyone can run, and the price war has pushed inference close to zero. Under export controls no Chinese lab has a path to a durable frontier lead anyway, so the sensible move is to give the model away and compete for ecosystem share. Commoditization is the default equilibrium, squarely in your pens world.
Hardware is your strongest petroleum case, but the rent lands on NVIDIA, ASML, TSMC and SK Hynix, all outside China. There the chokepoint is a pure cost, with no domestic gatekeeper to hold the public hostage, and Beijing is pouring money into commoditizing the layer (Ascend, SMIC, domestic HBM).
Energy is the American bottleneck; China's constraint is chips. It keeps adding generation fast enough to keep power cheap, with the grid state owned. The resource rent that looks most like oil is the one China is least exposed to.
That leaves land. A productivity boom capitalizes into land, and in the US that means private owners in San Francisco and Seoul collecting the windfall. Urban land in China is state owned, and households hold time-limited leaseholds, so the state already sits where your private landlord sits. What it does with that rent stays an open question, but the starting point sits far closer to the Land Value Return you want than anything in America.
So China looks more likely to reach the consumer-surplus world: three of your four rents are structurally weaker there, and the fourth sits in public hands.
This is an excellent analysis!
I will say the land analysis is a little more complicated than it appears at first glance, because although all the land is nominally state-owned in China, China has been extremely resistant to implementing an actual property tax, and local governments have gotten used to financing themselves with “land leases” that are functionally equivalent to land sales. Meanwhile the US is a bit of a patchwork on that front — California has done the most to undermine its property tax, enabling huge private capture of land rents, while in states like Texas the property taxes are actually quite high, leading to more public land rent recapture.
That said, the Chinese government could change the arrangement anytime, and is on a much firmer legal footing to do so should they desire.
This is just a fantastic explanation of free markets vs rent seeking. It would be tremendous if you turned these articles into short-form TikTok videos. The youth really need someone to explain these concepts simply.
I wish you’d considered the nonzero chance that AI is our successor species…
Yes, that's the other possibility. Surplus goes to the AI agents themselves.
You touch upon regulatory rents, but in this post you do not consider the biggest and most pervasive regulatory rent of all: intellectual property monopolies. Intellectual property creates long lived (even if ultimately temporary) monopolies, restricts competition (due to the monopoly), creates property rights over what is non-rivalrous, and restricts typical common law property rights over what is rivalrous for everyone else. This regulatory rent is pervasive throughout all industries (over 90% of fortune 500 assets are non-tangibles) and this includes both the AI and AI Hardware Manufacturing industries
IP law in the US is converging on a human-authorship requirement for protectable works, meaning work done by general AI models is not protected. I don't think frontier AI labs realize that, even if they are the first to AGI, all of AGI's new discoveries and inventions will default to the commons under existing law. I've written about this issue here:
https://amade.substack.com/p/superalignment-9ca
What are your thoughts on a Harbinger tax? Could that address this?
I like Harbinger to self-assess fixed improvements when combined with an auction for occupancy rights to a location but am not familiar with it in the context of IP. What are your thoughts?
I think it’s a better alternative to most of what we have but for medicine I would personally push for prizes instead of patents.
How will we know if it becomes a windfall? what would be the indications here?
Large and persistently durable profit margins that are extremely resistant to competition
Would you regard the tech giants an example of an natural monopoly? Should we tax them in a similar way?
Yeah, I think what regulators are seeing in AI (at least the foundational models) is a natural monopoly akin to the utility and rail industries. Foundational models require astronomically high upfront costs, with lower (though still high) operating costs. The high barrier to entry gives them pricing power that allows them to collect a windfall. Then the question would be 'is AI like internet service providers'? I actually read a policy paper from the Hamilton Center on Industrial Strategy that concluded that AI foundation models were natural monopolies, but the social cost to the monopoly was too low for regulation to be a significant benefit.
I mean the question is, even if it requires astronomically high upfront costs to produce a frontier model, once it’s produced, so far evidence shows us that they are rapidly distilled. Also, China can afford high upfront costs too. Is there a plausible story here for a frontier model taking the lead and durably keeping it without having to constantly plow forward?
Hmm, I don’t know enough about distillation and how much better the frontier model is from a distilled model, but like you said with China’s open-weight models, they do decrease the pricing power of the frontier models. If the AI models become commodified, then I guess the hyperscalers/cloud-providers would be the natural monopolies in this analogy? But I agree with the thesis of the article, the biggest issue will be the ground rent landowners will extract from the increase in productivity.
Yeah I don't doubt that a moat can hold up for a few years, maybe even a decade or longer. But I don't think they hold up forever under sustained competition. Someone needs a thesis for how they perpetually fend off competition (the ASI ascension thesis is basically this)
Fascinating can you give me the link to the paper. Seems like an interesting read.
Sure!: https://www.rand.org/content/dam/rand/pubs/research_reports/RRA3400/RRA3415-1/RAND_RRA3415-1.pdf
I misspoke, it was the RAND Corporation not Hamilton Center on Industrial Strategy. But here's their conclusion:
"Application of the natural monopoly criteria to the status quo foundation language model market (as of January 2024) indicates that the current case for a natural monopoly is relatively strong. This conclusion is based on the observations that the current generation of foundation models is reasonably homogeneous, economies of scale are high, costs are largely sunk, and network effects and economies of scope are present."
Though in all fairness, there are newer studies that claim Reinforcement Learning AI, doesn't have the same features.
This is a great piece, but I think the petroleum analogy could use some refinement. Petrostates can maintain consistently incredible profits from oil because they have very low production costs and because they collude through OPEC to restrict the supply of oil. The remaining oil demand is met by high-cost countries like the US and Canada and here the profits are much smaller. On average, oil companies are more profitable than pen companies, but it is also much more volatile, so this extra profit is mostly just a risk premium. Some years oil companies will be very profitable, and some years, such as 2015 and 2020, they will lose tons of money.
Think Rockefeller and the consolidation of refining