It’s a thesis about where the money in AI actually goes once we transition to scaled-up AI workloads and how that might look different from today’s expectations.
Seems like they’ll need to new benchmarks for non-SOTA models to companies can “pick” a task from x model (and the benchmark will validate the model’s ability to complete said task)
If deprecated models are incredibly cheap compared to SOTA, and more spending shifts to the vendor of the deprecated model, wouldn't some of the increased demand be offset, or potentially more than offset, in a decrease in price per token (assuming a reasonably competitive market?)
I think the one missing part of the write up is the value add that a harness provides….I think you are paying for that as much as the model itself. My firm uses its own harness and the usabilty is muchmuch worse
Thanks, the other point I was thinking about was on-prem inferencing taking share from hyperscalers. Consumer traffic should stay with hyperscalers but for Enterprises there is a big value proposition for doing your inferencing locally (along with training smaller/specialized models as you noted)
I think you’ve laid out the hyper-bull case for semiconductors, but framed it as a bear case.
From first principles, if the efficiency of each unit of compute (GPU, CPU, memory, networking) rises dramatically, then the economic value per unit of semiconductor rises with it. That is not bearish for semis, but the opposite.
Take your accountant example. If it previously took 10 units of compute to perform the work of one accountant, and now it takes only 1 unit, that single unit has not become less valuable. It has become 10x more productive. If the output value remains tied to the labor being replaced or augmented, then the semiconductor unit capturing that output has far more pricing power, while its production cost does not rise proportionally. That is how margins surge.
This is why I think the key distinction from your “Scenario 2” matters. Semiconductors are no longer just picks and shovels. They are becoming the scarce technological layer of digital labor. The chip companies own the technological bottleneck; the hyperscalers are largely just 1 out of many segments that distributes, operates, and rents that capacity.
Hyperscalers may own the installed base, but that installed base depreciates, becomes obsolete, and must be refreshed to stay competitive. The semiconductor layer determines the performance frontier, the cost curve, and the rate at which digital labor improves. In an economy increasingly powered by AI agents, owning that bottleneck is far more valuable than merely renting it out.
The barrier-to-entry point is also crucial. Many companies have become AI infrastructure operators, AI clouds, or inference platforms. In contrast, very few companies capable of producing frontier GPUs, CPUs, memory, networking, packaging, and the associated full-stack hardware systems at scale. The rental/operator layer is far more likely to be competed down and commoditised.
Don't take my word for it, here are the current facts: Half of Nvidia's AI revenue now comes from non-hyperscalers.
Jensen Huang: "Instead of five or six or seven companies representing the revenues associated with our first category, the second category is hundreds, thousands of companies, and in the future would be hundreds of thousands of companies. A large number of companies with smaller installations. And that category is going to continue to grow at incredible pace.”
If AI efficiency improves, the world does not need fewer semiconductors. It gets more use cases, more deployments, lower unit costs, and a much larger addressable market for digital labor. Demand explodes. That is Jevons paradox with pricing power attached.
The hyperscalers need access to the semiconductor layer to participate in the AI economy. But semiconductor companies do not need any hyperscalers in the same way, and have a much more diversified customer base as compared to hyperscalers' semi suppliers.
Semi are the kingmakers of this new economy, because they control the scarce input that everyone else operates on top of.
Thank you @zende. Very solid take. The most irreplaceable infrastructure vendors command more pricing power than the hyperscalers. In the two scenarios Rihard lays out, the hyperscalers (the infrastructure owners, as opposed to infrastructure suppliers) are better off in the first scenario, where capex tails off, right? The infrastructure suppliers are highly dependent upon continued capex spend, whether to expand net capacity or merely to replace outdated existing infrastructure, right?
Correct me if I'm wrong, but it seems that some of the infrastructure suppliers (an elite few) are far more irreplaceable than the hyperscalers, whereas the hyperscalers may be less cyclical than these elite infrastructure suppliers, because the infrastructure suppliers depend on the first derivative of capacity demanded whereas the hyperscalers may still be reasonably healthy even when the first derivative of capacity demanded drops to zero, so long as capacity demanded still corresponds to a healthy level of utilization of existing capacity.
One way to play this. When you think the chances of an impending flatlining/pullback in AI demand is high, you rotate into the hyperscalers (to be crouched more defensively, if you don't have enough conviction to rotate out of AI entirely). When you think the flatlining/pullback will soon reverse, you rotate back into the infrastructure suppliers.
As an infrastructure supplier, the problem with having your revenue driven by the first derivative of capacity demanded is that when AI capital spend peaks, your revenue growth doesn’t just decline. It goes negative. This revenue path may not be reflected in your current valuation.
GOOGL’s $80B equity raise is instructive. FCF is tapped out. Debt markets are tapped out. There is only so much stomach for equity dilution.
What’s the only salvation for continued increase in AI capital spend? Large revenue (and ultimately FCF) increases from that capital spend.
Yet, it’s now apparent that a sizable proportion of current AI demand is not sustainable. Lab subsidization of customer usage is shrinking. Enterprise token spend is being rationalized. It goes on and on.
Exactly how much this will impact the slope of capacity demanded is anyone’s guess, but it’s certainly not going to help lift internal cash flow generation, which capital spend will have to increasingly rely upon as external funding gets more difficult to come by.
It’s really perilous when the market is pricing in continued revenue growth for the infrastructure suppliers, which in turn depends upon continued increases in capital spend by the hyperscalers (and other infrastructure owners), right when the sources of that capital spend are drying up.
The scenario whereby first derivative of capacity demanded drops drastically is when either a) AI progress / adoption stalls or b) AI is too expensive to garner mass adoption. Both of which is the exact opposite of the scenario laid out in this article and also opposite of what is happening at the moment.
In the event that AI infrastructure becomes much more efficient, valuable and productive, yes hyperscalers will benefit from the temporary arbitrage gain of their current asset value rising, until the next replacement cycle within a few years. Whereas semi companies will experience a structural value rerating because they are the only few in the world who can manufacture these newly minted digital labor. Hyperscalers in essense are rentors of the asset (with leasing contracts of a few years) and will benefit temporarily, whereas semis are owners that benefit permanently due to the surge in fundamental value of their products.
An imperfect (but good enough to illustrate this point) example is the analogy of air travel: if it were to remain expensive and reserved only for the wealthy, its TAM would be much smaller than it is today. But there is a structural difference between companies that manufacture the main asset: jet engines (semis) and airline companies involed in large scale fleet management and rental of asset (hyperscalers). Former produces the asset that needs renewal every few years, while latter merely rents (with renewal cycles) and distributes it. Fundamental difference in barriers to entry, competition and margins between both groups of companies.
Great points. Conditions for when capacity demanded stalls/drops are not only when algorithmic breakthroughs at least temporarily halt growth in capacity demanded (which is what I thought Rihard may have had in mind) but also, as you pointed out, when AI progress/adoption stalls (e.g., tokenmaxxing dies out) or AI’s value-to-cost proposition is too low to garner mass adoption.
You’re absolutely right that the infrastructure suppliers are largely the owners of the economic value created by their products, whereas the hyperscalers are renters of that asset value. Where the hyperscalers are owners instead of renters are in the data/service gravity that binds their more loyal customers to them. This is a force less strong than what the N of 1 or N of 2 infrastructure suppliers own, but it still commands some economic rents. The neoclouds can win customers that have standalone AI use cases, but some enterprises will opt for co-location of their AI deployment with their existing applications (and data) in the cloud.
Over the long run, I would prefer to own N of 1 or 2 assets that are supported by very moat-y characteristics, but to the extent that I am willing to time the market, the hyperscalers would have appeal when I feel the first derivative of capacity demanded is potentially poised for a fall (a concern you seem to have shared with me) but do not have enough conviction yet (or enough confidence in my timing instincts/nimbleness) to rotate out of AI entirely. The better relative performance of AMZN today vs. a high proportion of infrastructure suppliers seems to support the premise that the operators of the infrastructure have generally greater down-cycle resilience than the suppliers of the infrastructure, which matters less for those who are concerned less about volatility.
Really great example about aircraft makers vs. owners. Being the asset operator isn't great if barriers to entry aren't high enough. (Also isn't great if fixed-cost economics often drive you to pauper yourself in pursuit of marginal revenue.)
A trading firm being able to build and operate its own AI infrastructure reflects how much of a moat hyperscalers really have.
Agentic AI drastically lowering software’s barriers to entry, enabling all companies big or small to manage and orchestrate their own compute.
Agentic AI at its core is the democratisation of tech, which will result in rapid decentralisation. One of the worst dynamics that can happen to the hyperscaler business model
Seems like they’ll need to new benchmarks for non-SOTA models to companies can “pick” a task from x model (and the benchmark will validate the model’s ability to complete said task)
Great read! We have similar takes!
https://lifeinthesingularity.com/p/glm-52-proves-ai-comes-for-all-moats
If deprecated models are incredibly cheap compared to SOTA, and more spending shifts to the vendor of the deprecated model, wouldn't some of the increased demand be offset, or potentially more than offset, in a decrease in price per token (assuming a reasonably competitive market?)
It would probably mean a gap window would emerge, before Jevon's Paradox takes over.
I think the one missing part of the write up is the value add that a harness provides….I think you are paying for that as much as the model itself. My firm uses its own harness and the usabilty is muchmuch worse
Our writing just a few days ago: https://www.uncoveralpha.com/p/the-harness-the-moat-for-ai-model
Thanks, the other point I was thinking about was on-prem inferencing taking share from hyperscalers. Consumer traffic should stay with hyperscalers but for Enterprises there is a big value proposition for doing your inferencing locally (along with training smaller/specialized models as you noted)
I think you’ve laid out the hyper-bull case for semiconductors, but framed it as a bear case.
From first principles, if the efficiency of each unit of compute (GPU, CPU, memory, networking) rises dramatically, then the economic value per unit of semiconductor rises with it. That is not bearish for semis, but the opposite.
Take your accountant example. If it previously took 10 units of compute to perform the work of one accountant, and now it takes only 1 unit, that single unit has not become less valuable. It has become 10x more productive. If the output value remains tied to the labor being replaced or augmented, then the semiconductor unit capturing that output has far more pricing power, while its production cost does not rise proportionally. That is how margins surge.
This is why I think the key distinction from your “Scenario 2” matters. Semiconductors are no longer just picks and shovels. They are becoming the scarce technological layer of digital labor. The chip companies own the technological bottleneck; the hyperscalers are largely just 1 out of many segments that distributes, operates, and rents that capacity.
Hyperscalers may own the installed base, but that installed base depreciates, becomes obsolete, and must be refreshed to stay competitive. The semiconductor layer determines the performance frontier, the cost curve, and the rate at which digital labor improves. In an economy increasingly powered by AI agents, owning that bottleneck is far more valuable than merely renting it out.
The barrier-to-entry point is also crucial. Many companies have become AI infrastructure operators, AI clouds, or inference platforms. In contrast, very few companies capable of producing frontier GPUs, CPUs, memory, networking, packaging, and the associated full-stack hardware systems at scale. The rental/operator layer is far more likely to be competed down and commoditised.
Don't take my word for it, here are the current facts: Half of Nvidia's AI revenue now comes from non-hyperscalers.
Jensen Huang: "Instead of five or six or seven companies representing the revenues associated with our first category, the second category is hundreds, thousands of companies, and in the future would be hundreds of thousands of companies. A large number of companies with smaller installations. And that category is going to continue to grow at incredible pace.”
If AI efficiency improves, the world does not need fewer semiconductors. It gets more use cases, more deployments, lower unit costs, and a much larger addressable market for digital labor. Demand explodes. That is Jevons paradox with pricing power attached.
The hyperscalers need access to the semiconductor layer to participate in the AI economy. But semiconductor companies do not need any hyperscalers in the same way, and have a much more diversified customer base as compared to hyperscalers' semi suppliers.
Semi are the kingmakers of this new economy, because they control the scarce input that everyone else operates on top of.
Thank you @zende. Very solid take. The most irreplaceable infrastructure vendors command more pricing power than the hyperscalers. In the two scenarios Rihard lays out, the hyperscalers (the infrastructure owners, as opposed to infrastructure suppliers) are better off in the first scenario, where capex tails off, right? The infrastructure suppliers are highly dependent upon continued capex spend, whether to expand net capacity or merely to replace outdated existing infrastructure, right?
Correct me if I'm wrong, but it seems that some of the infrastructure suppliers (an elite few) are far more irreplaceable than the hyperscalers, whereas the hyperscalers may be less cyclical than these elite infrastructure suppliers, because the infrastructure suppliers depend on the first derivative of capacity demanded whereas the hyperscalers may still be reasonably healthy even when the first derivative of capacity demanded drops to zero, so long as capacity demanded still corresponds to a healthy level of utilization of existing capacity.
One way to play this. When you think the chances of an impending flatlining/pullback in AI demand is high, you rotate into the hyperscalers (to be crouched more defensively, if you don't have enough conviction to rotate out of AI entirely). When you think the flatlining/pullback will soon reverse, you rotate back into the infrastructure suppliers.
As an infrastructure supplier, the problem with having your revenue driven by the first derivative of capacity demanded is that when AI capital spend peaks, your revenue growth doesn’t just decline. It goes negative. This revenue path may not be reflected in your current valuation.
GOOGL’s $80B equity raise is instructive. FCF is tapped out. Debt markets are tapped out. There is only so much stomach for equity dilution.
What’s the only salvation for continued increase in AI capital spend? Large revenue (and ultimately FCF) increases from that capital spend.
Yet, it’s now apparent that a sizable proportion of current AI demand is not sustainable. Lab subsidization of customer usage is shrinking. Enterprise token spend is being rationalized. It goes on and on.
Exactly how much this will impact the slope of capacity demanded is anyone’s guess, but it’s certainly not going to help lift internal cash flow generation, which capital spend will have to increasingly rely upon as external funding gets more difficult to come by.
It’s really perilous when the market is pricing in continued revenue growth for the infrastructure suppliers, which in turn depends upon continued increases in capital spend by the hyperscalers (and other infrastructure owners), right when the sources of that capital spend are drying up.
The scenario whereby first derivative of capacity demanded drops drastically is when either a) AI progress / adoption stalls or b) AI is too expensive to garner mass adoption. Both of which is the exact opposite of the scenario laid out in this article and also opposite of what is happening at the moment.
In the event that AI infrastructure becomes much more efficient, valuable and productive, yes hyperscalers will benefit from the temporary arbitrage gain of their current asset value rising, until the next replacement cycle within a few years. Whereas semi companies will experience a structural value rerating because they are the only few in the world who can manufacture these newly minted digital labor. Hyperscalers in essense are rentors of the asset (with leasing contracts of a few years) and will benefit temporarily, whereas semis are owners that benefit permanently due to the surge in fundamental value of their products.
An imperfect (but good enough to illustrate this point) example is the analogy of air travel: if it were to remain expensive and reserved only for the wealthy, its TAM would be much smaller than it is today. But there is a structural difference between companies that manufacture the main asset: jet engines (semis) and airline companies involed in large scale fleet management and rental of asset (hyperscalers). Former produces the asset that needs renewal every few years, while latter merely rents (with renewal cycles) and distributes it. Fundamental difference in barriers to entry, competition and margins between both groups of companies.
Great points. Conditions for when capacity demanded stalls/drops are not only when algorithmic breakthroughs at least temporarily halt growth in capacity demanded (which is what I thought Rihard may have had in mind) but also, as you pointed out, when AI progress/adoption stalls (e.g., tokenmaxxing dies out) or AI’s value-to-cost proposition is too low to garner mass adoption.
You’re absolutely right that the infrastructure suppliers are largely the owners of the economic value created by their products, whereas the hyperscalers are renters of that asset value. Where the hyperscalers are owners instead of renters are in the data/service gravity that binds their more loyal customers to them. This is a force less strong than what the N of 1 or N of 2 infrastructure suppliers own, but it still commands some economic rents. The neoclouds can win customers that have standalone AI use cases, but some enterprises will opt for co-location of their AI deployment with their existing applications (and data) in the cloud.
Over the long run, I would prefer to own N of 1 or 2 assets that are supported by very moat-y characteristics, but to the extent that I am willing to time the market, the hyperscalers would have appeal when I feel the first derivative of capacity demanded is potentially poised for a fall (a concern you seem to have shared with me) but do not have enough conviction yet (or enough confidence in my timing instincts/nimbleness) to rotate out of AI entirely. The better relative performance of AMZN today vs. a high proportion of infrastructure suppliers seems to support the premise that the operators of the infrastructure have generally greater down-cycle resilience than the suppliers of the infrastructure, which matters less for those who are concerned less about volatility.
Really great example about aircraft makers vs. owners. Being the asset operator isn't great if barriers to entry aren't high enough. (Also isn't great if fixed-cost economics often drive you to pauper yourself in pursuit of marginal revenue.)
A trading firm being able to build and operate its own AI infrastructure reflects how much of a moat hyperscalers really have.
Agentic AI drastically lowering software’s barriers to entry, enabling all companies big or small to manage and orchestrate their own compute.
Agentic AI at its core is the democratisation of tech, which will result in rapid decentralisation. One of the worst dynamics that can happen to the hyperscaler business model
https://www.bloomberg.com/news/articles/2026-06-04/jane-street-plans-new-data-center-as-compute-power-runs-scarce