By Thorsten Meyer
The confident forecast at the center of this industry is that intelligence becomes abundant — cheap, ambient, seeping through the economy like electricity until it is simply everywhere and nearly free. I think that forecast is basically right. And I think almost everyone drawing comfort from it has misread where the comfort actually lives.
Because “intelligence becomes a commodity” is not a soothing statement. It is a statement about where value leaves. When something becomes abundant, it stops being the thing you get paid for. The money, the power, and the scarcity all migrate somewhere else — and the whole game, for anyone building in this decade, is being early to where. This is my map of where the value goes when the intelligence itself no longer holds any.
The forecast is right: intelligence becomes a commodity, cheap and ambient like electricity. But “commodity” is a statement about where value leaves. The whole game is being early to where it goes instead.
▲ Opinion & analysis · not investment adviceWhen the crude is cheap, value moves to the refinery, the trusted name on the deal, and the buyer who can only drink so much. Same shape here.
When a capability becomes abundant and free, we stop exercising it. Some of that is fine. Some of it hollows us out.
knowing which wishes are worth making — and being a person who can still tell.
The commodity nobody wants to admit they're selling
Start with the honest version of the abundance thesis, because the polite version hides the sting. If raw intelligence — the ability to reason, write, code, analyze — is on a curve toward being priced like a utility, then the frontier labs are, whether they enjoy the framing or not, commodity producers. Extraordinary ones. But commodity producers sell a fungible good into a market where the buyer does not care which molecule of crude, which kilowatt-hour, which token they received.
I live at the receiving end of this. On my own infrastructure I route work to whichever model gives me the best answer per unit cost, and I switch without sentiment the moment a better trade appears. That is exactly how you treat a commodity. It is not disloyalty; it is what a commodity is. And the deep tell that the producers know this too is the language of "every point on the curve" — the ambition to offer the best intelligence-to-price trade at every latency and every price band. That is not the language of a luxury good. That is the language of a refinery.
So the first move in reading the future correctly is to stop asking which lab has the smartest model — they will trade that crown monthly — and start asking what does not become a commodity when the intelligence does. There are three answers, and they are the whole map.
Hewlett Packard Enterprise ProLiant Compute DL360 Gen12 w/one Intel Xeon 6530P Processor, 1P 2x32GB-R 8SFF NS204i-u v2 MR408i-o 2x1000W PS (HPE Smart Choice P89997-005)
- Model Type: Enterprise 1U Rack Server
- Processor: Intel Xeon 6530P, 32-core, 2.3GHz
- Memory: 64GB DDR5 RAM
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What stays scarce, part one: the fleet
The first thing that does not commoditize is the ability to produce the commodity at scale. When crude is cheap, the durable advantage is not owning a barrel; it is owning the refinery, the pipeline, the port. When intelligence is cheap, the durable advantage is the compute fleet — the physical capacity to turn electricity into tokens, and the supply chain that lets you build more of it faster than anyone else.
This is the single most under-appreciated inversion in the whole story. Everyone stares at the models, which are the part racing toward zero value. The scarce thing is underneath: the chips, the racks, the land, the power, and above all the rate at which you can add more. A frontier model is a depreciating asset that a competitor can match or distill in months. A gigawatt of energized, chip-filled, cooled datacenter capacity is a physical fact that takes ten thousand workers eighteen months to build and that no amount of algorithmic cleverness conjures out of nothing. The moat was never the intelligence. The moat is the means of production.
And this is precisely where my concern as a European sharpens into something specific. If the scarce, value-holding layer of the entire AI economy is physical production capacity — fabs, high-bandwidth memory, and gigawatts of power — then a region that consumes intelligence but does not produce the means of making it has outsourced the one layer that stays valuable. Being a brilliant user of abundant intelligence is a fine life. It is not sovereignty. Sovereignty lives in the fleet, and the fleet lives, overwhelmingly, elsewhere. That is the strategic fact I cannot stop turning over.
Seagate IRONWOLF PRO 16TB 2PK
- Storage Capacity: 16TB per drive
- Number of Drives: 2-pack
- Warranty: 5-year limited warranty
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What stays scarce, part two: the human in the loop
The second thing that does not commoditize is stranger and more reassuring, and I have become convinced it is real rather than sentimental.
Even in a world of superhuman intelligence on tap, people keep choosing the human. Not out of nostalgia — out of something structural. We are hardwired to care what other people care about. The value of a piece of art collapses into the signature not because the brushwork is better but because you want to know there was a person behind it. You read a novel to meet a mind. And in business, the thing customers actually want is not the smartest possible decision — it is a someone who is accountable for the decision, who can be trusted, praised, blamed, and held responsible. Nobody wants an AI CEO. Not because the AI would decide worse, but because accountability is a human-to-human relationship and cannot be delegated to a system that no one can look in the eye.
This is where I part company with the pure-substitution panic, and it is directly relevant to how I run my own work. The most valuable thing I produce is not the analysis — the models can approximate a great deal of that, cheaply, and will approximate more. The valuable thing is the byline: a named human standing behind a judgment, who has staked reputation on it and will answer for it. As intelligence becomes abundant, that stake becomes more valuable, not less, because it is the scarce complement to the cheap input. The abundance of the reasoning is exactly what inflates the worth of the accountable human wrapped around it.
There is a word missing from our language for the faculty this depends on. It is not "taste," though that is close. It is the specific human judgment — about what matters, about what is worth doing, about where responsibility lands — that these systems struggle with most and that recedes most slowly. Whatever we end up calling it, it is where a great deal of human economic value relocates in an abundant-intelligence world. Not the answer. The judgment about which answer was worth wanting, and the willingness to own the consequences.
GPU Kernel Engineering for LLM Inference: CUDA, Triton, and Flash Attention Optimization for High-Throughput AI Production Systems (AI Infrastructure, Hardware & Compiler Engineering Series)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What stays scarce, part three: attention, and its ceiling
The third scarce thing is the one that quietly bounds the entire boom, and almost nobody prices it: human attention.
The demand for intelligence is described as uncapped, and at a low enough price that is nearly true — every builder, myself included, will drag the slider toward "spend more compute thinking on my behalf while I sleep" further than seems reasonable, because the marginal idea is worth more than the marginal token costs. But there is a ceiling hiding in that sentence, and it is not made of silicon. It is that a human being can only absorb, direct, evaluate, and act on so much. If the models get smart and efficient enough to build everything we can think to ask for, and our attention cannot absorb more than a surprisingly finite amount of their output, then even uncapped demand meets a wall — the wall of what a person can meaningfully use.
I raise this not as a bearish note but as a locating one. It tells you where the next real bottleneck sits. Not in the models, not even entirely in the fleet, but in the interface between abundant machine cognition and finite human bandwidth. Whoever solves that — how a single person meaningfully directs and metabolizes the output of a thousand tireless agents — captures enormous value, because they are relieving the one constraint that abundance cannot relieve on its own. The scarce resource at the end of the abundance story is us.
Local LLM Inference Optimization: A Comprehensive Guide to Quantization, Hardware Acceleration, and Efficient Private AI Deployment
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The atrophy question I actually lose sleep over
There is a cost in all this that does not show up on any balance sheet, and it is the one I think about most, because it is the one that compounds silently.
When a capability becomes abundant and free, we stop exercising it. That is not a hypothetical; it is the entire history of tools. Some of that is fine — I do not mourn the arithmetic I no longer do by hand. But there is a version of this that hollows us out: where we outsource not just the labor of thinking but the muscle of it, and wake up a decade later having lost the ability to understand the things that matter, dependent on a system we can no longer evaluate. The danger is not that the machine becomes too smart. It is that we let ourselves become too soft to check its work — and thereby hand it, by default and inattention, exactly the concentration of decision-making power that the optimistic version of this future was supposed to prevent.
This is why I build the way I build — local-first, running my own models on my own hardware, staying close enough to the metal to actually understand the stack I depend on. Not because it is always cheaper; it often is not. But because the alternative — total dependence on a few distant, abundant-intelligence utilities I neither control nor comprehend — is precisely the cognitive and civic atrophy I am trying to avoid. Keeping capability distributed and keeping my own understanding sharp are the same act. Abundance makes both harder and both more necessary.
Where this lands
The abundance thesis is right, and its comforts are almost all misplaced. Intelligence becomes a commodity, which means the value drains out of the intelligence and pools in three places that stay scarce: the physical fleet that produces it, the accountable human who stands behind the judgment, and the finite attention that has to absorb it all. The winners of the next decade are not the ones with the smartest model in any given month. They are the ones who own the means of production, who make themselves the trusted human at the center of the loop, and who solve the human-bandwidth bottleneck that even infinite intelligence runs into.
And the quiet imperative underneath all three, at least for me: do not let the abundance of thinking talk you out of doing your own. When the machine can grant almost any wish, the scarcest and most valuable thing left is knowing which wishes are worth making — and being a person who can still tell.
Opinion and analysis from the perspective of a local-first builder and post-labor economist; part of an ongoing series on the economics of abundant intelligence. This piece engages widely-discussed theses about AI abundance, commoditization, and human complementarity; the framing, the three-scarcities map, and the conclusions are my own. Not investment advice. Point-in-time as of 5 August 2026.
