The Capital is Misaligned and the Crash is Coming
My strong disagreement with the sentiment on Wall Street.
I am going to say two things that sound contradictory and are not.
The first is that AI is the future. Large language models are the most significant advance in the history of software, and the industry has not yet absorbed what that means. The second is that the current AI buildout is a bubble — not because the technology is overhyped, but because the capital being deployed is massively misaligned with where the technology is actually going. Both claims are true simultaneously. That simultaneity is what neither Wall Street nor the hyperscaler executives are prepared to face, because the implication is that the entire cloud-software business model of the last fifteen years is about to be dismantled by the technology the same industry is pouring trillions into building.
I made the technical and philosophical case against the scaling-to-AGI thesis at length in Why I’m Betting Against the AGI Hype — the compound probability analysis, the architectural barriers, the category error at the heart of the AGI project. That piece is the foundation of the argument I’m making here. I won’t rehearse it in full. The relevant conclusion for this essay is narrower: the scaling paradigm is not going to deliver the transformations its current valuations require, and the capital allocation built on the assumption that it will is historically misaligned.
I want to explain why I think this. I am not a detached observer. I spent a career in the technology industry, I have built software at scale, and I use these tools daily — I have several Claude Code projects running right now. I know what these models are capable of. I also know what they are not. And when I look at what they are, and where they are going, and then I look at the investment thesis that is supposedly supporting the valuations, the gap between the two is the biggest mispricing I have seen in my professional lifetime.
⁂
What LLMs actually do to software.
Start with what has already happened, because the industry has been remarkably reluctant to describe it plainly. Large language models, when properly deployed by a competent engineer, collapse the cost of producing custom software by at least an order of magnitude. I am not speculating about this. I am reporting it from the inside. Tasks that used to require a six-month engagement with a development team can now be done, by one engineer with a good model and clear specifications, in a week. Sometimes in a day. Sometimes in an afternoon.
This is not a claim about the models being smart. It is a claim about the economics of software production. Large language models — as I argued in LLMs Are Universal Translators, Not Universal Thinkers — are translation engines. They translate between symbol systems. They translate business requirements into code. They translate specifications into implementations. They translate natural language descriptions of workflows into the software that executes those workflows. That translation function, applied to software production, is what collapses the economics of custom software. The cost of a line of working code has collapsed. The cost of translating a business requirement into a functioning system has collapsed. The cost of integrating two systems that were not previously integrated has collapsed. The cost of producing a reasonable UI on top of a reasonable database has collapsed. All of the tasks that made enterprise software expensive to build — the reason companies paid Salesforce forty thousand dollars a seat per year instead of building their own CRM — are tasks the models now do for the engineer in real time.
Think about what this means for the software-as-a-service business model. SaaS exists because, historically, the cost of building and maintaining bespoke software was higher than the cost of renting generic software from a vendor. You paid Salesforce because building what Salesforce does was more expensive than paying Salesforce to do it. You paid ServiceNow because standing up your own ticketing and workflow system would have taken a team of engineers a year and you could rent ServiceNow in a week. You paid Workday, Zendesk, HubSpot, Monday, Atlassian, and the hundred other SaaS tools in the average enterprise stack for the same reason. The generic product, however imperfect, was cheaper than bespoke.
That equation is in the process of inverting.
Consultancies are already figuring this out. A mid-sized consultancy with a team of competent engineers using modern tooling can now build a company a custom CRM that fits that company’s actual workflow — not a generic CRM configured into a shape that approximates their workflow — for less than three years of Salesforce licensing fees. And the custom CRM runs on the company’s own infrastructure, does not charge per seat, does not extract telemetry for training on competitors, does not impose an upgrade cycle, does not get worse when the vendor decides to reorient its roadmap around a new executive’s priorities, and can be modified at any time by anyone with the model access to modify it. The economics are not close. They are not going to be close. They are going to diverge further as the tooling improves, which it is doing on a timescale of months.
Enterprise SaaS is going to experience what the record industry experienced when MP3s got compressed enough to download. Except the SaaS companies are, in many cases, larger than the record labels were, and the dependence on their products is deeper, and the customer lock-in is stronger — which means the reckoning will take longer to arrive and will be more violent when it does.
Where inference is going.
The second piece of the mispricing is the assumption that large-scale language model inference will continue to run in hyperscaler data centers forever, consuming the electricity and renting the compute that justifies the current capital expenditure on GPU buildouts. This assumption is not going to hold, and anyone paying attention to the hardware roadmap already knows it.
Watch what Apple is doing. The Neural Engine in Apple Silicon has been quietly accumulating capability for years. The A-series and M-series chips have been engineered specifically to run transformer models efficiently on-device. The combination of algorithmic improvements — quantization, distillation, sparse attention, mixture-of-experts architectures — and hardware acceleration on custom silicon is making local inference viable for models that, two years ago, required a cluster of H100s to serve. Within the next several iterations of the hardware — I am not making a ten-year prediction, I am making a short-horizon prediction — Apple will be running a local Siri on your phone that is as capable as ChatGPT or Claude is today. For free. Without sending your queries to a server. Without training on your data. Without rate limits. Without surveillance.
And Apple will do this not because they are benevolent but because they have figured out that local inference is an enormous competitive advantage against Google and Microsoft, both of whom depend on cloud-based AI for revenue and would be structurally incapable of matching a local-first offering even if they wanted to.
Once the local version is good enough — and “good enough” is a rapidly moving target that the local side is catching up to faster than the cloud side is improving — the economics of cloud inference collapse for a large class of consumer use cases. Why would anyone pay for a subscription to a chatbot when their phone has one that is equivalent and free and private? The cloud providers will still have a market — for frontier-model research, for specialized enterprise applications, for use cases that genuinely require scale — but the mass consumer and small-business market will migrate to local inference, and the revenue models that justified the current GPU buildout assumed that mass consumer and small-business market as the foundation.
The hyperscalers are building data centers to serve a market that will, within a short horizon, have migrated to a substrate they cannot charge for. The capital expenditure is going to produce useful infrastructure — no argument there — but the revenue that was supposed to justify the expenditure is going to be a fraction of what the investment thesis requires.
⁂
Why this is a bubble without being a fraud.
This is the part that is genuinely hard to hold, and the reason most observers are failing to hold it. A bubble does not require the underlying technology to be fake. It requires the capital to be misallocated relative to where the technology’s value will actually land. The dot-com bubble was a bubble despite the internet being real. The railroad bubbles of the nineteenth century were bubbles despite railroads being real. The question is never whether the thing is real. The question is whether the capital deployed into the thing will earn a return commensurate with the capital deployed.
The current AI buildout fails this test not because the technology is a mirage but because the technology is going to do in-situ business model destruction to the companies building it. The value added by LLMs is enormous. The value that will be captured by the companies currently spending hundreds of billions on training runs and data center buildouts is much smaller than the total value added, and in many cases the companies will lose value overall — because the technology they are building is going to destroy the business models of their own other products.
Microsoft is pouring capital into OpenAI while simultaneously presiding over an enterprise software portfolio — including Office, Dynamics, and Azure SaaS services — that AI will cannibalize. The returns from the OpenAI investment would have to exceed the destruction of the rest of the Microsoft portfolio for the capital allocation to be rational. I do not believe they will. I do not think anyone doing the math inside Microsoft believes they will, either. But the alternative is to not invest in AI, and that is not an option any CEO can take, because the short-term market would punish the refusal more severely than the long-term market will punish the miscalibration.
This is a classic bubble dynamic. Every individual actor is responding rationally to their local incentives. The aggregate outcome is irrational. Wall Street is not irrational in some mysterious way — Wall Street is responding to the fact that nobody wants to be the analyst who called the top, and nobody wants to be the portfolio manager who underweighted Nvidia in 2024. The incentives inside financial media are identical. CNBC and Bloomberg are not wrong about the technology. They are structurally incapable of describing a scenario in which the technology is real and the capital is misallocated, because that scenario requires a kind of analytical patience that cable financial news does not reward.
The gap between the two claims — the technology is revolutionary and the capital deployed into the technology will not earn its return — is the gap where most of the money is going to be lost.
⁂
What the industry knows but cannot say.
The thing I want to say, because I have been inside these conversations and I know what is discussed in private versus what is said in public, is that many of the people building the current AI infrastructure know that the economics do not work. They know that SaaS is going to be disrupted by their own technology. They know that local inference is going to eat the consumer market. They know that the capital deployed is not going to earn its return on a cycle that matches the current valuations.
What they cannot say, in public, is what the alternative is. Because there is no obvious pivot from we are building the future of software to we are building infrastructure for a market that will partially disappear by the time the infrastructure is operational. The former supports the stock price. The latter does not. And the executives of these companies are not paid to tell the truth about the market. They are paid to maintain the conditions under which the current valuations are defensible for as long as possible, and then to be gone with their compensation packages intact before the reckoning arrives.
This is not a conspiracy. It is a structural feature of how publicly traded companies behave when their valuations depend on a narrative that insiders have reason to doubt. The narrative is maintained not because anyone is lying but because no one is paid to contradict it, and everyone is paid to extend it. The only people who lose by the extension are the retail investors who bought the story, the employees whose stock options will be worth a fraction of what they were promised, and the communities whose energy infrastructure is being strained to serve an industry that will not be at its current scale in ten years.
The pension funds will take the hit. The 401k plans will take the hit. The municipal bondholders who financed the data center power upgrades will take the hit. The hyperscaler executives will be fine. This is the pattern, and it is the same pattern every bubble produces.
⁂
The revolution inside the bubble.
And here is the thing I want to leave you with, because it is the hardest to hold and the most important. The bubble does not invalidate the revolution. The revolution is real. The technology is going to transform the production of software, the operation of businesses, the texture of everyday digital life. The fact that the current capital allocation is misaligned does not mean the technology is a fraud. It means the market has, as markets do, built the wrong infrastructure for the transformation that is coming.
When the bubble pops, the technology will still be here. The GPUs will still exist. The models will still run. The software produced with them will still work. What will not exist is the specific set of business models that the bubble was built to justify. SaaS will be dismantled. The cloud-inference revenue thesis will be revealed as a transitional artifact. The hyperscaler capex will be marked down. The companies that survive will be the ones that figured out, before the rest, that the value of AI is not captured by controlling the models but by deploying them — at the level of the engineer, the consultancy, the local handset, the specific enterprise workflow.
The future belongs to the people and organizations that understand that AI is not a platform business. It is an infrastructure dissolvent. It dissolves the platforms that used to collect rent on the production of software. And the capital that is currently being deployed to build new platforms on top of AI is making a category error — building fortresses out of the same material that is designed to dissolve fortresses.
The political stakes of this are larger than the financial ones, and I’ve made the case at length in Garbage In, Garbage Out. The philosophical architecture being built around AI — the framing of intelligence as pattern recognition, of governance as optimization, of human judgment as an inferior computational substrate — does not disappear when the bubble pops. The financial reckoning will arrive first. The philosophical reckoning will take longer, and it is the one that matters more. When the GPUs are marked down and the valuations corrected and the hyperscaler earnings revised, the deeper question will remain: what kind of political order did we build while the bubble was inflating, and who gets to dismantle the parts of it that shouldn’t survive?
⁂
What to watch.
If you want to know whether I am right, watch three things.
First, watch consultancy revenue over the next several years. If custom-built enterprise software starts eating into the SaaS revenue curves — if companies start reporting that they are canceling subscriptions and moving to in-house systems built by smaller teams — that is the leading indicator of the model I am describing.
Second, watch Apple‘s on-device AI story. The moment Apple ships a local Siri that matches cloud-based chatbots for the average consumer use case, the cloud-inference consumer revenue thesis dies. It will not happen in one product cycle. It will happen over two or three. And when it happens, the revenue curves for cloud AI services will bend.
Third, watch what the hyperscalers do with their data centers when the AI training demand plateaus. The buildout assumes continuous exponential growth in training compute. When the curve bends — which it will, because there are only so many engineers to train and only so many use cases to serve — the infrastructure will need to find other uses, and the other uses will pay substantially less than the training demand was paying.
The reckoning will not arrive dramatically. It will arrive quarterly, in earnings reports that slightly underperform expectations, in revised guidance that explains away the shortfall, in write-downs that are framed as strategic realignments. It will arrive the way all bubble reckonings arrive: in the gap between the story the industry told itself and the revenue the industry actually produced. The story was always too big for the revenue to catch up. The revenue is what the market eventually prices on.
⁂
AI is the future. The current AI buildout is a bubble. Both are true. The industry cannot say the second part out loud because the current valuations depend on pretending it is not true. Wall Street is cheering the buildout because Wall Street is paid to cheer the buildout. Financial media is narrating the story because narrating the story is their job. None of these actors are lying. They are performing the roles their incentives assign them.
The question for anyone trying to understand what is actually happening — as an investor, as a technologist, as a citizen, as a participant in the broader economy — is whether you can hold the two claims at once. The technology is real and the capital is misaligned. The revolution is coming and the revolution will destroy the business models of the companies currently being valued on the assumption that they will capture it.
This is what a technological revolution looks like from inside the bubble that accompanies it. The bubble will pop. The technology will remain. The people who understood the difference will be the ones who built what comes next.
The capital is misaligned. The technology is not.
agree, as a tech person watching this bubble exponentially inflate, even before the step up in hype about AI and LLM's pre-2022. Inference at the edge is the future, for most use cases its better, cheaper. And yet their trying to hypnotize/distract the retail investor with more and more datacenters-when its getting harder to build one by the day. And with supply chain disruptions from the middle east it will be more impossible to even maintain our grid let alone add capacity; consider how many more weather disasters we will have and the demand to replace equipment worldwide. We already have multiyear waiting list for transformers.
I have two remarks on this article.
First, the rent of generative AI must be captured by those whose jobs it is replacing, not by investors who already had the capital to allocate and happened to do so correctly or by executives who flew the coop before their bad business decisions came crashing down. We have failed consistently to do this in the past and each time it has been a step on the road toward the current late-stage capitalist dystopia that we've been in at least since Kropotkin penned "The Conquest of Bread".
Second, despite this article, I am still of the view that generative AI must be seared from existence. It is having two incredibly negative effects. First, it is devaluing the real talent, real skill, and real effort that creative people have put into their abilities at writing and/or visual art, to the point where they now need disclaimers that no AI was involved in their production. (Further, there are copyright concerns around the training data for LLMs and graphics generators, and I think all such data used must be licenced from rightsholders.) Second, the proliferation of AI-generated articles on various topics threatens to poison something that has made humanity as successful as we have been, which is intergenerational knowledge transfer. Trained experts across numerous fields have repeatedly demonstrated that when LLMs produce "information" about their fields, the LLM output is very far from reliably correct. But with little way to distinguish between text written by an LLM and text written by an actual expert, the casual reader, including future generations needing to learn a given topic, has no way of knowing which knowledge is valid and which is not.
It's little use having a technological revolution that serves only to enrich those who are already rich and threatens to destroy the circumstances that permitted the revolution to occur in the first place.
Generative AI delenda est.
https://journa.host/@jeremiak/113811327999722586
https://existentialcomics.com/comic/540