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.
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 will write my comment in 2-3 parts over the weekend as I get more time. Here is the first part: I think the thesis is directionally right around SaaS, but it overgeneralizes by treating “SaaS” as a single thing.
I’ve been implementing enterprise-level solutions across ERP, CRM, and custom applications for Fortune 50 companies, the three largest federal departments, and many other environments for over 25 years as a consultant. At this point, I run the IT organization and have moved away from consulting, so I do not have much skin in the game beyond maintaining the system well—whether that system is built in-house or provided by a vendor. From where I sit, that operational perspective matters, because it changes the question from “can this be built?” to “can this be sustained, governed, secured, and improved over time without creating a bigger problem than the one you were trying to solve?” and what I’ve seen is not just a reduction in coding effort, but a structural shift in where the hard work lives. We went from technical teams that were often twice the size of the business analyst team to environments where that ratio is reversed in the last 10 years. That matters because in many enterprise contexts, the bottleneck is no longer writing software. It is understanding the business deeply enough to implement, govern, and maintain it correctly.
That is why I think the application's complexity matters much more than the article allows.
Modern SaaS is already highly configurable. If a company is willing to standardize workflows and business processes around leading practices, many SaaS products are relatively easy to maintain with a much smaller internal team than was required in the past. In those cases, the value is not just that the vendor wrote the code. The value is that the vendor has already packaged years of process assumptions, operational patterns, controls, upgrades, and support into something that can be run at scale.
Where I agree strongly with the post is that AI is compressing the cost of building and modifying software, and that this will pressure parts of the SaaS market. But I do not think the conclusion should be that SaaS broadly gets dismantled. I think the correct conclusion is that some categories are highly exposed, while others remain durable for much longer.
ERP is the clearest example. In ERP, you are not just building screens and workflows. You are combining finance, procurement, inventory, manufacturing, reporting, controls, permissions, auditability, and, often, regulatory requirements into a single operating system for the company. Even if AI makes it much easier to generate code, that does not mean it makes it easy to own and maintain institutional correctness across that system. For mid-sized and large companies, I think broad in-house replacement of ERP cores is still a long way off.
By contrast, for other categories—CRM, ServiceNow-like workflows, internal case management, departmental apps, lightweight operational systems, and various custom extensions—the economics are already shifting. In those areas, in-house or semi-custom solutions may soon have an edge, if they do not already.
The other missing variable in the post is that replicability is not the same as advisability. Yes, most SaaS software can be replicated. That does not mean most companies should do it. If you want to maintain a large internal team to continuously upgrade the system, handle security, satisfy audit and compliance requirements, manage integrations, and preserve business continuity, you can. But I would not recommend that path for most companies. Business and domain knowledge are critical, and most IT departments lack an end-to-end understanding of the solution from both the business and technical sides to own that responsibility safely.
So I think the strongest version of the thesis is not “SaaS is going away.” It is something more like this:
AI will unevenly compress the moat of software categories whose main value came from packaging relatively standard workflows into configurable products. The more a system depends on cross-functional integration, process discipline, controls, compliance, and deep domain knowledge, the more durable the incumbent platforms will remain.
Put differently, here is how I would segment it:
- Simple departmental tools — High vulnerability
- Easier workflows
- Limited control burden
- Easier to replicate or replace with AI-assisted custom solutions
- CRM / case management / ticketing — Medium-high to high vulnerability
- Often configurable and reproducible
- More exposed where differentiation comes from workflow fit rather than deep platform complexity
- Strong candidates for AI-generated or bespoke solutions
- Especially where the process is specific to the company
- Service management platforms — Medium vulnerability
- Replaceable in parts
- Still valuable where governance, ecosystem, and enterprise controls matter
- ERP core systems — Low vulnerability in the near term
- Deeply coupled processes
- Significant audit, security, compliance, and business continuity requirements
- Hard to replace wholesale, even if pieces around them become easier to build
- Regulated industry systems — Low to medium vulnerability
- Validation, traceability, and risk management matter as much as functionality
- Replacement cost is not just technical; it is organizational and regulatory
- Large cross-enterprise platforms — Medium vulnerability
- Difficult to replace end-to-end
- More likely to be surrounded, extended, or selectively displaced than fully rebuilt
So yes, I agree with the article that AI changes the economics of software. I also agree that many current assumptions about value capture are too broad and too bullish. But I think the real story is not that all SaaS is threatened equally. It is that application complexity, process standardization, control requirements, and domain knowledge determine where AI breaks the model first.
Part 2: I also think the model discussion is being framed too much as “cloud vs local,” when the more important shift may be toward a more heterogeneous stack: frontier models in the cloud, smaller specialized models in private or edge environments, and orchestration layers that route tasks based on cost, latency, privacy, and reliability requirements. Some workloads will run locally, some privately, and some in the cloud. The key distinction is not just the hosting location. The question is whether the workflow actually requires a large, centralized, general model or whether a narrower model, a tuned system, or a hybrid approach is good enough.
That matters because it changes where value accrues. If the future is increasingly hybrid and specialized rather than dominated by a small number of centralized general models, then distribution, integration, workflow fit, governance, and operational deployment may matter more than raw hosting position alone.
Apple has every incentive to push AI toward edge computing and on-device inference, because that aligns directly with its strengths: hardware integration, control of the platform, privacy positioning, and monetization through device sales and ecosystem lock-in. Hyperscalers have the opposite incentive. They want AI to remain cloud-centric because their economic model depends on centralized compute, recurring usage, and infrastructure rents. But incentives do not fully determine outcomes. Edge constraints are real, and the cloud will remain the better architecture for many workloads. Another thing, less as one architect pushing another and more as different players pushing the stack in the direction that best fits their business model.
Part 3: Before I jump to the last part, I wanted to say thank you for writing this post, and it made me think deeply about this topic more than ever before.
Regarding the bubble question, I think my response below overlaps with several of the points you mentioned in your post. I would distinguish between a genuine technological wave and an overheated investment cycle. The two are not mutually exclusive. As you also said, AI can be real, important, and ultimately transformative while still generating pockets of excess in which capital, expectations, and narrative run ahead of durable value capture. My reading of past bubbles, especially the railway bubble of the 1800s, is that this is often how major technology cycles unfold. The central paradox is that the technology is already useful, but making it broadly useful beyond a relatively narrow set of high-value applications may require enormous spending on compute, power, data centers, integration, security, workflow redesign, governance, and human oversight before the returns become clear. In that sense, the technology may be real even if parts of the investment case are overheated.
The first area of possible bubble behavior is hyperscaler and frontier-model spending. That spending may eventually prove justified, but once capital deployment moves materially ahead of demonstrated demand and monetization, the conditions for overbuild are in place. The risk is not only that demand disappoints; it is also that usage grows while economics remain weaker than current spending implies because pricing compresses, competition intensifies, or customers capture more of the value than suppliers do. There is also a reflexive element to the cycle. Hyperscalers and frontier labs may feel compelled to keep investing because progress sustains the narrative, and the narrative sustains the capital, strategic relevance, and customer confidence needed to keep progress going. Part of the capex cycle may therefore be self-reinforcing rather than purely demand-led.
A second area is the startup ecosystem built on top of foundation models. Some companies will build real businesses through distribution, workflow integration, proprietary data, or execution. Still, many may prove to be thin wrappers with limited differentiation, weak pricing power, and deep dependence on upstream model providers they do not control. Closely related is the broader belief that because coding has become easier, building durable software businesses has become easy. It has not. AI-assisted coding can radically accelerate prototyping, but a prototype is not a production system. Architecture, security, integration, testing, governance, maintenance, and product judgment still matter, especially inside large organizations.
A third area is the widening infrastructure complex forming around anticipated AI demand: semiconductors, networking, cooling, power, construction, real estate, and data center supply chains. Whenever capacity is built primarily against expected future demand rather than realized utilization and durable returns, the risk of overinvestment rises. Bubble behavior often spreads outward in precisely this way: investors stop buying only the core theme and begin buying everything adjacent to its buyers.
There is also a softer but important bubble risk in enterprise behavior. Many companies are funding pilots, copilots, demos, and AI initiatives because they feel they must show momentum, not because they already have a clear path to scaled deployment and measurable return. That can produce a kind of pilot theater: visible activity with little impact on production. The same is true, to some extent, of the consulting layer around AI transformation. Some of that work will be valuable, but some of it will monetize urgency before the operating model is mature.
Another area of overconfidence may lie in the technical thesis around LLMs themselves. The question is not whether they are powerful; clearly, they are. The question is whether LLMs alone can carry us all the way to robust, production-grade intelligence in critical applications. I am skeptical. Hallucination, traceability, and robustness remain serious constraints where correctness matters more than fluency. At the same time, I am not persuaded that “world models” alone provide a clean answer. The more plausible path seems to be hybrid systems that combine LLMs with retrieval, structured data, symbolic methods, domain-specific models, tool use, verification layers, and tighter workflow constraints. If that is right, then part of the market may be overpricing a simpler and cleaner technical story than reality will support.
The economics will also depend heavily on model specialization and deployment architecture. We still do not know how valuable AI will be, or how much will require massive centralized models, and how much can be handled by smaller, domain-specific systems running locally, on-premises, or at the edge. If useful workloads migrate meaningfully toward specialized models and edge deployment, inference economics could change materially, dependence on frontier providers could weaken, and more value could shift toward devices, chips, integration, and workflow ownership. If that transition comes faster than the current capital spending assumes, parts of the present investment narrative could deflate. If it comes more slowly, the case for centralized scale and continued infrastructure concentration becomes stronger.
Two broader uncertainties could reshape the entire landscape. One is open source. If open models continue to close the gap, they could compress pricing and reduce the strategic leverage of frontier providers, shifting value toward deployment, integration, and customer ownership. If they do not, concentration around a small number of frontier firms may prove more durable than many expect. The other is geopolitics. Cyber risk, export controls, defense demand, model misuse, and national-security concerns could all influence where models are built, how they are deployed, how open ecosystems remain, and whether parts of the spending cycle are driven as much by strategic logic as by ordinary commercial return.
More broadly, the risk here is not always a classic valuation bubble. In many places, it may be an expectations bubble, a monetization bubble, or a time-horizon mismatch: long-term demand may turn out to be real, but adoption, pricing power, enterprise readiness, and operational maturity may arrive much more slowly than current capital commitments assume. Many firms believe they are AI-ready because they possess data, when in practice, data quality, permissions, governance, and process maturity remain serious bottlenecks. In regulated or high-liability domains, compliance, auditability, and accountability may slow deployment far more than current enthusiasm suggests.
So yes, I think bubble-like behavior is possible, but I wouldn't describe it as a single AI bubble. It is more likely a set of overlapping pockets of speculation: hyperscaler capex, frontier-model funding dynamics, wrapper startups, second-order infrastructure plays, enterprise pilot theater, consulting layers, overconfidence in AI-generated software, and an overly simple belief that scaling LLMs alone will deliver dependable intelligence. The common pattern is not fake technology. It is real capability accompanied by capital, expectations, and extrapolation that may be moving faster than sustainable monetization, operational reality, and the actual system architectures needed to make AI reliable in the real world. And that is why one of my favorite lines feels like the right way to end: “Reality always wins. Your job is to get in touch with it.” The AI industry will meet that reality, too. When it will happen is unknowable, but many of today’s companies will likely not survive the reckoning. As with the dot-com bubble, the hype will fade; the durable businesses will remain.
You just wrote a note about capitalism. This is another of its signatures. Periodic economic crises caused by overproduction, which is generated by universal competition among private firms with a license to amass gigantic wealth by owning the society’s means of production. It’s fancier this time because AI and other sophisticated technology is involved, but it’s the same old collapse.
Unregulated capitalism will consume itself and tend towards monarchy. I believe this to be true. But I don't think the mere existence of a capitalist economy need necessarily be unsustainable. It is a proper function of government to regulate markets towards providing positive social utility. To allow a frontier of creativity through competition. But also to ensure an economic safety net. So that everyone has a house, and healthcare and healthy food to eat and a chance to do their best and contribute to society. We have enough wealth in this country to do all this, and its time for a new deal for the American people.
Capitalism can be productive, and it can become extractive when its untethered from the pursuit of the common good. A lot of businesses contribute to the common good. Some businesses grow to a certain size, and they then capture entire markets, and the politicians and then the government and we get what we have today. I think we can have a world where people are free to start a business, grow a business and succeed. And that doesn't have to be incompatible with making sure everyone is housed, educated, and given opportunities to grow and lead a meaningful life.
I have no argument with this but I’d revise the standard discourse around the idea of restraining capitalist prerogatives through regulation, because such framing endows a capitalist economy with an aura of being the default, which only afterwards is to be rationally controlled, but to which society must defer in the first instance as if it were a law of nature. It is not a law of nature, and is not entitled to be assumed. To do so is to preemptively disenfranchise the commons and subordinate it in stature to the pursuit of private gain - the great right-wing “achievement” of the last 45-50 years.
We should start from the proposition that certain public goods are not for sale - end of story - and then capitalism can have the remainder so long as it does not infringe on those public goods.
What those public goods are is debatable. I think it includes basically everything we need to live decently: housing, food, education, and health care right off the bat, but also libraries, parks, art, access to nature, and to the democratic process. One can think of others, but it must include limits on the quantity of wealth any one person can own or control, not just because it’s just and protects the commons, but because it is ludicrous to assert that any one person deserves more than that limit no matter what innovation they might have contributed. Private wealth beyond that number just means that other people’s contributions are not being acknowledged, or are being stolen. I think that limit should be set much lower than most: $20,000,000 (that number should be uncontroversial - alas, it’s not).
The real problem with Silicon Valley and its technologists is that all this discussion is framed from their perspective. The Consumer is so left out, the Consumer has become the problem. There is no evaluation of what people want, just how do we make people want whatever new things we can invent.
That leads to Yarvinism where “The Masses are Asses” and you look at the Consumer more as potential BIODIESEL.
Capitalism has two ways of working out. One is the Star Trek Replicator/Holideck World where all your physical needs can be met, letting everyone live lives at the Top of the Maslow Pyramid. And the other is a machine that is a giant Skinner Box that deals everyone reward and punishment reflective of how well they keep the machine functioning as judged by the Architects who become our Masters.
Elon Musk wanted X to be the Everything App. Now it’s a place you can post ideas and get personal insults yelled at you by literal Nazis. But you can also make porn of anyone you like/despise, especially underage girls, courtesy of MechaHitler. I mean @Grok.
The idea of Musk and Peter Thiel torturing/ feeding us hamster pellets for their amusement in pursuit of their Vision and calling that Freedom. Instead of gross, old democracy. 🤮!
This is how you get Luigis and guys throwing Molotov cocktails into toilet paper factories. Which means you need to build more PAIN into the Machine.
Agreed. This is not my wheelhouse, but it def helped me have a better understanding, a clearer picture for what I have innately been feeling about this whole expansion.
If you've done any research on how the data centers are being built and what it takes to support them, it's very clear this is unsustainable. There's an egregious amount of waste and excess capacity being built. You have multiple players building what is effectively the same thing, it's redundancy at scale.
AI and LLMs should have been a public/private partnership like nuclear power or the space race. The implications for our society were just as important, probably even greater. We knew this back in 2017. Instead, we have allowed a "winner takes all" system to take root like we did with the Internet. There are way too many people chasing far too few resources and no referee.
Physics says you can't dissipate the heat on these new servers effectively without massive amounts of coolant and power. Using water efficiently is a requirement, we can't change the weather patterns of the world, no matter how many billions of dollars we have. Using oil, glycol or other coolants have their own complications. The latest Blackwell chips from NVIDIA will only have 75 seconds to determine if you have a problem that will cause a catastrophic heat cascade that cooks the server to slag. If a heatpump fails (and it will, entropy is a thing), you need to make a very fast decision on what to do. Is it a false positive? Do you interrupt your 99.99% uptime? What's the useful life of the incredibly expensive server? 2 years? Less? Is that built into your cashflow forecast? It's worth noting that Physics also says you can only move light so fast through fiber optic cable, there are limits we can't upend with more.
The Abilene data center in Texas won't have the power it requires until 2027 at the earliest. Other gigawatt sucking data centers need similar levels of power, something we haven't invested in as a society. It takes YEARS to fabricate power turbines for generators, the connection equipment, etc. That power has to come from somewhere. So are consumers going to go without heat, cooling, and light so a few people can make machines that take our jobs? Elon better get cracking on those security robots, because that sounds like a French Revolution level of casus belli for the masses.
The chips are effectively all manufactured in Korea (for RAM) and Taiwan (processors). There's a war that has completely upended those nation's access to energy (LNG and the right kind of oil may not be available at ANY price soon). What happens when they don't have power to do the laser lithography in Taiwan? The debt burden on the data center is still burning even if it's an empty warehouse without any computer hardware generating inbound cashflow. Sounds like a debt restructure in the making. That will have knock-on effects for equity investors who suddenly will look at the actual economics.
Every LLM company loses money at the margins on every prompt. They charge less per token than it costs to create the compute for the token. You cannot brute force the cost down. The entire pricing structure for the usage of LLMs is a trojan horse. They are going to charge more for these, they don't right now because they are subsidized by investors. At some point the investors need a return and the vacuum becomes a blower. For that to happen they need to substantially increase pricing, which begs the question of whether or not it's better to just have a human doing this work.
The tech companies will spend every last dollar to the point of folly because they think the winner here will be like the winner of Search, Social Media, etc. The reality is far different. The only barrier to entry is the cap-ex for the servers and we already have open competition here. LLMs will be a utility.
More importantly, novel thought is not built into the design and cannot be engineered into the machine. You cannot brute force AGI with an LLM. LLMs are great for remembering what we know as a society (most of the time, when they aren't hallucinating). They struggle with the black swans or projecting events that haven't happened before. They might make connections based on their training routine, but that's not the same as a human doing "what if". A human mind is not a matrix math engine.
All of this is to say that LLMs are incredibly useful. They have a point of diminishing returns. We are investing past that point. It's the Railroads all over again.
When the ‘sub-prime mortgage’ market imploded supposed experts testified anywhere and everywhere they could expressing astonishment that the bankers would have kept writing all those loans to all of those people who clearly could not service the debt. They feigned being dumbfounded that the bank execs would have done that.
I said, “I know why they did it: If the annual bonus dangled out in front of me if I just kept the balls in the air for one more year were that big I woulda kept writing NINJA loans until all those shoes dropped too.”
I don't understand the "it will change everything" argument made in this piece because it doesn't address hallucinations. How can LLM code ever be good enough? What I'm hearing from software devs is that they're pulling back, becoming more and more aware that they're producing more/faster but the mistakes are harder to catch and fix and their people will over time, lose their skill/understanding.
I find enormous value in LLMs, don't get me wrong, but not for coding. Immensely useful in answering conceptual questions.
I'm not going to read your post. The first few paragraphs tell me what I've been thinking for awhile. There's no reason I should be right but we agree fundamentally, for me the why doesn't matter.
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
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 will write my comment in 2-3 parts over the weekend as I get more time. Here is the first part: I think the thesis is directionally right around SaaS, but it overgeneralizes by treating “SaaS” as a single thing.
I’ve been implementing enterprise-level solutions across ERP, CRM, and custom applications for Fortune 50 companies, the three largest federal departments, and many other environments for over 25 years as a consultant. At this point, I run the IT organization and have moved away from consulting, so I do not have much skin in the game beyond maintaining the system well—whether that system is built in-house or provided by a vendor. From where I sit, that operational perspective matters, because it changes the question from “can this be built?” to “can this be sustained, governed, secured, and improved over time without creating a bigger problem than the one you were trying to solve?” and what I’ve seen is not just a reduction in coding effort, but a structural shift in where the hard work lives. We went from technical teams that were often twice the size of the business analyst team to environments where that ratio is reversed in the last 10 years. That matters because in many enterprise contexts, the bottleneck is no longer writing software. It is understanding the business deeply enough to implement, govern, and maintain it correctly.
That is why I think the application's complexity matters much more than the article allows.
Modern SaaS is already highly configurable. If a company is willing to standardize workflows and business processes around leading practices, many SaaS products are relatively easy to maintain with a much smaller internal team than was required in the past. In those cases, the value is not just that the vendor wrote the code. The value is that the vendor has already packaged years of process assumptions, operational patterns, controls, upgrades, and support into something that can be run at scale.
Where I agree strongly with the post is that AI is compressing the cost of building and modifying software, and that this will pressure parts of the SaaS market. But I do not think the conclusion should be that SaaS broadly gets dismantled. I think the correct conclusion is that some categories are highly exposed, while others remain durable for much longer.
ERP is the clearest example. In ERP, you are not just building screens and workflows. You are combining finance, procurement, inventory, manufacturing, reporting, controls, permissions, auditability, and, often, regulatory requirements into a single operating system for the company. Even if AI makes it much easier to generate code, that does not mean it makes it easy to own and maintain institutional correctness across that system. For mid-sized and large companies, I think broad in-house replacement of ERP cores is still a long way off.
By contrast, for other categories—CRM, ServiceNow-like workflows, internal case management, departmental apps, lightweight operational systems, and various custom extensions—the economics are already shifting. In those areas, in-house or semi-custom solutions may soon have an edge, if they do not already.
The other missing variable in the post is that replicability is not the same as advisability. Yes, most SaaS software can be replicated. That does not mean most companies should do it. If you want to maintain a large internal team to continuously upgrade the system, handle security, satisfy audit and compliance requirements, manage integrations, and preserve business continuity, you can. But I would not recommend that path for most companies. Business and domain knowledge are critical, and most IT departments lack an end-to-end understanding of the solution from both the business and technical sides to own that responsibility safely.
So I think the strongest version of the thesis is not “SaaS is going away.” It is something more like this:
AI will unevenly compress the moat of software categories whose main value came from packaging relatively standard workflows into configurable products. The more a system depends on cross-functional integration, process discipline, controls, compliance, and deep domain knowledge, the more durable the incumbent platforms will remain.
Put differently, here is how I would segment it:
- Simple departmental tools — High vulnerability
- Easier workflows
- Limited control burden
- Easier to replicate or replace with AI-assisted custom solutions
- CRM / case management / ticketing — Medium-high to high vulnerability
- Often configurable and reproducible
- More exposed where differentiation comes from workflow fit rather than deep platform complexity
- Internal workflow apps / portals / dashboards — High vulnerability
- Strong candidates for AI-generated or bespoke solutions
- Especially where the process is specific to the company
- Service management platforms — Medium vulnerability
- Replaceable in parts
- Still valuable where governance, ecosystem, and enterprise controls matter
- ERP core systems — Low vulnerability in the near term
- Deeply coupled processes
- Significant audit, security, compliance, and business continuity requirements
- Hard to replace wholesale, even if pieces around them become easier to build
- Regulated industry systems — Low to medium vulnerability
- Validation, traceability, and risk management matter as much as functionality
- Replacement cost is not just technical; it is organizational and regulatory
- Large cross-enterprise platforms — Medium vulnerability
- Difficult to replace end-to-end
- More likely to be surrounded, extended, or selectively displaced than fully rebuilt
So yes, I agree with the article that AI changes the economics of software. I also agree that many current assumptions about value capture are too broad and too bullish. But I think the real story is not that all SaaS is threatened equally. It is that application complexity, process standardization, control requirements, and domain knowledge determine where AI breaks the model first.
Part 2: I also think the model discussion is being framed too much as “cloud vs local,” when the more important shift may be toward a more heterogeneous stack: frontier models in the cloud, smaller specialized models in private or edge environments, and orchestration layers that route tasks based on cost, latency, privacy, and reliability requirements. Some workloads will run locally, some privately, and some in the cloud. The key distinction is not just the hosting location. The question is whether the workflow actually requires a large, centralized, general model or whether a narrower model, a tuned system, or a hybrid approach is good enough.
That matters because it changes where value accrues. If the future is increasingly hybrid and specialized rather than dominated by a small number of centralized general models, then distribution, integration, workflow fit, governance, and operational deployment may matter more than raw hosting position alone.
Apple has every incentive to push AI toward edge computing and on-device inference, because that aligns directly with its strengths: hardware integration, control of the platform, privacy positioning, and monetization through device sales and ecosystem lock-in. Hyperscalers have the opposite incentive. They want AI to remain cloud-centric because their economic model depends on centralized compute, recurring usage, and infrastructure rents. But incentives do not fully determine outcomes. Edge constraints are real, and the cloud will remain the better architecture for many workloads. Another thing, less as one architect pushing another and more as different players pushing the stack in the direction that best fits their business model.
Part 3: Before I jump to the last part, I wanted to say thank you for writing this post, and it made me think deeply about this topic more than ever before.
Regarding the bubble question, I think my response below overlaps with several of the points you mentioned in your post. I would distinguish between a genuine technological wave and an overheated investment cycle. The two are not mutually exclusive. As you also said, AI can be real, important, and ultimately transformative while still generating pockets of excess in which capital, expectations, and narrative run ahead of durable value capture. My reading of past bubbles, especially the railway bubble of the 1800s, is that this is often how major technology cycles unfold. The central paradox is that the technology is already useful, but making it broadly useful beyond a relatively narrow set of high-value applications may require enormous spending on compute, power, data centers, integration, security, workflow redesign, governance, and human oversight before the returns become clear. In that sense, the technology may be real even if parts of the investment case are overheated.
The first area of possible bubble behavior is hyperscaler and frontier-model spending. That spending may eventually prove justified, but once capital deployment moves materially ahead of demonstrated demand and monetization, the conditions for overbuild are in place. The risk is not only that demand disappoints; it is also that usage grows while economics remain weaker than current spending implies because pricing compresses, competition intensifies, or customers capture more of the value than suppliers do. There is also a reflexive element to the cycle. Hyperscalers and frontier labs may feel compelled to keep investing because progress sustains the narrative, and the narrative sustains the capital, strategic relevance, and customer confidence needed to keep progress going. Part of the capex cycle may therefore be self-reinforcing rather than purely demand-led.
A second area is the startup ecosystem built on top of foundation models. Some companies will build real businesses through distribution, workflow integration, proprietary data, or execution. Still, many may prove to be thin wrappers with limited differentiation, weak pricing power, and deep dependence on upstream model providers they do not control. Closely related is the broader belief that because coding has become easier, building durable software businesses has become easy. It has not. AI-assisted coding can radically accelerate prototyping, but a prototype is not a production system. Architecture, security, integration, testing, governance, maintenance, and product judgment still matter, especially inside large organizations.
A third area is the widening infrastructure complex forming around anticipated AI demand: semiconductors, networking, cooling, power, construction, real estate, and data center supply chains. Whenever capacity is built primarily against expected future demand rather than realized utilization and durable returns, the risk of overinvestment rises. Bubble behavior often spreads outward in precisely this way: investors stop buying only the core theme and begin buying everything adjacent to its buyers.
There is also a softer but important bubble risk in enterprise behavior. Many companies are funding pilots, copilots, demos, and AI initiatives because they feel they must show momentum, not because they already have a clear path to scaled deployment and measurable return. That can produce a kind of pilot theater: visible activity with little impact on production. The same is true, to some extent, of the consulting layer around AI transformation. Some of that work will be valuable, but some of it will monetize urgency before the operating model is mature.
Another area of overconfidence may lie in the technical thesis around LLMs themselves. The question is not whether they are powerful; clearly, they are. The question is whether LLMs alone can carry us all the way to robust, production-grade intelligence in critical applications. I am skeptical. Hallucination, traceability, and robustness remain serious constraints where correctness matters more than fluency. At the same time, I am not persuaded that “world models” alone provide a clean answer. The more plausible path seems to be hybrid systems that combine LLMs with retrieval, structured data, symbolic methods, domain-specific models, tool use, verification layers, and tighter workflow constraints. If that is right, then part of the market may be overpricing a simpler and cleaner technical story than reality will support.
The economics will also depend heavily on model specialization and deployment architecture. We still do not know how valuable AI will be, or how much will require massive centralized models, and how much can be handled by smaller, domain-specific systems running locally, on-premises, or at the edge. If useful workloads migrate meaningfully toward specialized models and edge deployment, inference economics could change materially, dependence on frontier providers could weaken, and more value could shift toward devices, chips, integration, and workflow ownership. If that transition comes faster than the current capital spending assumes, parts of the present investment narrative could deflate. If it comes more slowly, the case for centralized scale and continued infrastructure concentration becomes stronger.
Two broader uncertainties could reshape the entire landscape. One is open source. If open models continue to close the gap, they could compress pricing and reduce the strategic leverage of frontier providers, shifting value toward deployment, integration, and customer ownership. If they do not, concentration around a small number of frontier firms may prove more durable than many expect. The other is geopolitics. Cyber risk, export controls, defense demand, model misuse, and national-security concerns could all influence where models are built, how they are deployed, how open ecosystems remain, and whether parts of the spending cycle are driven as much by strategic logic as by ordinary commercial return.
More broadly, the risk here is not always a classic valuation bubble. In many places, it may be an expectations bubble, a monetization bubble, or a time-horizon mismatch: long-term demand may turn out to be real, but adoption, pricing power, enterprise readiness, and operational maturity may arrive much more slowly than current capital commitments assume. Many firms believe they are AI-ready because they possess data, when in practice, data quality, permissions, governance, and process maturity remain serious bottlenecks. In regulated or high-liability domains, compliance, auditability, and accountability may slow deployment far more than current enthusiasm suggests.
So yes, I think bubble-like behavior is possible, but I wouldn't describe it as a single AI bubble. It is more likely a set of overlapping pockets of speculation: hyperscaler capex, frontier-model funding dynamics, wrapper startups, second-order infrastructure plays, enterprise pilot theater, consulting layers, overconfidence in AI-generated software, and an overly simple belief that scaling LLMs alone will deliver dependable intelligence. The common pattern is not fake technology. It is real capability accompanied by capital, expectations, and extrapolation that may be moving faster than sustainable monetization, operational reality, and the actual system architectures needed to make AI reliable in the real world. And that is why one of my favorite lines feels like the right way to end: “Reality always wins. Your job is to get in touch with it.” The AI industry will meet that reality, too. When it will happen is unknowable, but many of today’s companies will likely not survive the reckoning. As with the dot-com bubble, the hype will fade; the durable businesses will remain.
You just wrote a note about capitalism. This is another of its signatures. Periodic economic crises caused by overproduction, which is generated by universal competition among private firms with a license to amass gigantic wealth by owning the society’s means of production. It’s fancier this time because AI and other sophisticated technology is involved, but it’s the same old collapse.
And capitalism has always been, ultimately, doomed to consume itself and fail — just like fascism.
Unregulated capitalism will consume itself and tend towards monarchy. I believe this to be true. But I don't think the mere existence of a capitalist economy need necessarily be unsustainable. It is a proper function of government to regulate markets towards providing positive social utility. To allow a frontier of creativity through competition. But also to ensure an economic safety net. So that everyone has a house, and healthcare and healthy food to eat and a chance to do their best and contribute to society. We have enough wealth in this country to do all this, and its time for a new deal for the American people.
Capitalism can be productive, and it can become extractive when its untethered from the pursuit of the common good. A lot of businesses contribute to the common good. Some businesses grow to a certain size, and they then capture entire markets, and the politicians and then the government and we get what we have today. I think we can have a world where people are free to start a business, grow a business and succeed. And that doesn't have to be incompatible with making sure everyone is housed, educated, and given opportunities to grow and lead a meaningful life.
I have no argument with this but I’d revise the standard discourse around the idea of restraining capitalist prerogatives through regulation, because such framing endows a capitalist economy with an aura of being the default, which only afterwards is to be rationally controlled, but to which society must defer in the first instance as if it were a law of nature. It is not a law of nature, and is not entitled to be assumed. To do so is to preemptively disenfranchise the commons and subordinate it in stature to the pursuit of private gain - the great right-wing “achievement” of the last 45-50 years.
We should start from the proposition that certain public goods are not for sale - end of story - and then capitalism can have the remainder so long as it does not infringe on those public goods.
What those public goods are is debatable. I think it includes basically everything we need to live decently: housing, food, education, and health care right off the bat, but also libraries, parks, art, access to nature, and to the democratic process. One can think of others, but it must include limits on the quantity of wealth any one person can own or control, not just because it’s just and protects the commons, but because it is ludicrous to assert that any one person deserves more than that limit no matter what innovation they might have contributed. Private wealth beyond that number just means that other people’s contributions are not being acknowledged, or are being stolen. I think that limit should be set much lower than most: $20,000,000 (that number should be uncontroversial - alas, it’s not).
The real problem with Silicon Valley and its technologists is that all this discussion is framed from their perspective. The Consumer is so left out, the Consumer has become the problem. There is no evaluation of what people want, just how do we make people want whatever new things we can invent.
That leads to Yarvinism where “The Masses are Asses” and you look at the Consumer more as potential BIODIESEL.
Capitalism has two ways of working out. One is the Star Trek Replicator/Holideck World where all your physical needs can be met, letting everyone live lives at the Top of the Maslow Pyramid. And the other is a machine that is a giant Skinner Box that deals everyone reward and punishment reflective of how well they keep the machine functioning as judged by the Architects who become our Masters.
Elon Musk wanted X to be the Everything App. Now it’s a place you can post ideas and get personal insults yelled at you by literal Nazis. But you can also make porn of anyone you like/despise, especially underage girls, courtesy of MechaHitler. I mean @Grok.
The idea of Musk and Peter Thiel torturing/ feeding us hamster pellets for their amusement in pursuit of their Vision and calling that Freedom. Instead of gross, old democracy. 🤮!
This is how you get Luigis and guys throwing Molotov cocktails into toilet paper factories. Which means you need to build more PAIN into the Machine.
Thank you for presenting this informed and important perspective in such an accessable way
Agreed. This is not my wheelhouse, but it def helped me have a better understanding, a clearer picture for what I have innately been feeling about this whole expansion.
Essentially doing something ten times faster and cheaper actually costs everything. That's what I just read.
If you've done any research on how the data centers are being built and what it takes to support them, it's very clear this is unsustainable. There's an egregious amount of waste and excess capacity being built. You have multiple players building what is effectively the same thing, it's redundancy at scale.
AI and LLMs should have been a public/private partnership like nuclear power or the space race. The implications for our society were just as important, probably even greater. We knew this back in 2017. Instead, we have allowed a "winner takes all" system to take root like we did with the Internet. There are way too many people chasing far too few resources and no referee.
Physics says you can't dissipate the heat on these new servers effectively without massive amounts of coolant and power. Using water efficiently is a requirement, we can't change the weather patterns of the world, no matter how many billions of dollars we have. Using oil, glycol or other coolants have their own complications. The latest Blackwell chips from NVIDIA will only have 75 seconds to determine if you have a problem that will cause a catastrophic heat cascade that cooks the server to slag. If a heatpump fails (and it will, entropy is a thing), you need to make a very fast decision on what to do. Is it a false positive? Do you interrupt your 99.99% uptime? What's the useful life of the incredibly expensive server? 2 years? Less? Is that built into your cashflow forecast? It's worth noting that Physics also says you can only move light so fast through fiber optic cable, there are limits we can't upend with more.
The Abilene data center in Texas won't have the power it requires until 2027 at the earliest. Other gigawatt sucking data centers need similar levels of power, something we haven't invested in as a society. It takes YEARS to fabricate power turbines for generators, the connection equipment, etc. That power has to come from somewhere. So are consumers going to go without heat, cooling, and light so a few people can make machines that take our jobs? Elon better get cracking on those security robots, because that sounds like a French Revolution level of casus belli for the masses.
The chips are effectively all manufactured in Korea (for RAM) and Taiwan (processors). There's a war that has completely upended those nation's access to energy (LNG and the right kind of oil may not be available at ANY price soon). What happens when they don't have power to do the laser lithography in Taiwan? The debt burden on the data center is still burning even if it's an empty warehouse without any computer hardware generating inbound cashflow. Sounds like a debt restructure in the making. That will have knock-on effects for equity investors who suddenly will look at the actual economics.
Every LLM company loses money at the margins on every prompt. They charge less per token than it costs to create the compute for the token. You cannot brute force the cost down. The entire pricing structure for the usage of LLMs is a trojan horse. They are going to charge more for these, they don't right now because they are subsidized by investors. At some point the investors need a return and the vacuum becomes a blower. For that to happen they need to substantially increase pricing, which begs the question of whether or not it's better to just have a human doing this work.
The tech companies will spend every last dollar to the point of folly because they think the winner here will be like the winner of Search, Social Media, etc. The reality is far different. The only barrier to entry is the cap-ex for the servers and we already have open competition here. LLMs will be a utility.
More importantly, novel thought is not built into the design and cannot be engineered into the machine. You cannot brute force AGI with an LLM. LLMs are great for remembering what we know as a society (most of the time, when they aren't hallucinating). They struggle with the black swans or projecting events that haven't happened before. They might make connections based on their training routine, but that's not the same as a human doing "what if". A human mind is not a matrix math engine.
All of this is to say that LLMs are incredibly useful. They have a point of diminishing returns. We are investing past that point. It's the Railroads all over again.
When the ‘sub-prime mortgage’ market imploded supposed experts testified anywhere and everywhere they could expressing astonishment that the bankers would have kept writing all those loans to all of those people who clearly could not service the debt. They feigned being dumbfounded that the bank execs would have done that.
I said, “I know why they did it: If the annual bonus dangled out in front of me if I just kept the balls in the air for one more year were that big I woulda kept writing NINJA loans until all those shoes dropped too.”
I don't understand the "it will change everything" argument made in this piece because it doesn't address hallucinations. How can LLM code ever be good enough? What I'm hearing from software devs is that they're pulling back, becoming more and more aware that they're producing more/faster but the mistakes are harder to catch and fix and their people will over time, lose their skill/understanding.
I find enormous value in LLMs, don't get me wrong, but not for coding. Immensely useful in answering conceptual questions.
I'm not going to read your post. The first few paragraphs tell me what I've been thinking for awhile. There's no reason I should be right but we agree fundamentally, for me the why doesn't matter.
Well said. I appreciate the detail without the noise commentary.
A fundamental shift is underway in the structure of global power.
For decades, geopolitical influence was defined by:
Military capability
Economic scale
Strategic alliances
Today, a new variable has moved to the center:
Technology.
Artificial intelligence, semiconductor manufacturing, and clean energy systems are no longer supporting sectors of the global economy.
They are becoming the primary determinants of geopolitical power.
The New Equation
The emerging reality can be summarized simply:
Technological power = geopolitical power
Nations that lead in:
AI development and deployment
Advanced chip design and fabrication
Control over battery and energy supply chains
will shape:
Global economic structures
Security architectures
Information ecosystems
Standards and norms for the digital age
This is not a future scenario.
It is already unfolding.
The Silent Acceleration
What makes this shift particularly significant is its resilience.
Despite:
Ongoing conflicts
Economic uncertainty
Geopolitical fragmentation
Investment in technology continues to accelerate globally.
Capital is flowing into:
AI infrastructure
Semiconductor ecosystems
Clean energy innovation
This indicates a strategic consensus:
Technology is the most reliable path to long-term dominance.
The Expanding Frontier
The competition is not limited to current technologies.
Emerging domains are already reshaping the horizon:
Quantum computing, with the potential to break existing encryption systems
Next-generation cloud and AI ecosystems redefining productivity and control
Integrated digital infrastructure linking economies, defense, and governance
These are not isolated innovations.
They are components of a new global operating system.
The Strategic Shift Required
Global leaders must move from:
Viewing technology as an economic sector → to treating it as strategic infrastructure
Supporting innovation → to orchestrating national and allied tech ecosystems
Competing in markets → to competing in system design
Because the nature of competition has changed.
It is no longer about outperforming rivals within the same system.
It is about building the system others must operate within.
sounds like bitcoin mining may yet just survive. all that data centre infrastructure, some of that will pivot (back?) to hashrate.
Very helpful response. Thanks!