From Ozempic to AI Agents: How Demand Gets Suppressed
Last year, between X posts and conversations with investors, one of the most common AI questions I heard was simple: When do I buy software stocks? After I wrote a paper at the end of 2025 arguing that software was entering a re-rating phase, many investors reached out to say that re-rating had already happened. This year, there are fewer questions about an AI bubble, but the software questions persist and they sound more defiant than curious.
Much of that defiance comes from price action. Investors look at the parabolic charts of scarcity-driven hardware winners like Micron and SanDisk and compare them to the abundance-era drawdowns of Salesforce, Adobe, and ServiceNow. From that comparison, they infer value in what has fallen and excess in what has run. I think they have it exactly backward. Stepping in to buy software because it looks “overdone” ignores what is actually happening in AI. Will there be relative bounces? Of course. But this isn’t a typical sector rotation or a buying opportunity in beaten-down software names. We are watching the opening act of a demand-destruction cycle, driven by the steady deflationary pressure of exponential AI progress, where coding is increasingly ubiquitous and effectively free. They are confusing the suppressant for the bubble and the bloat for the value.
To understand why this is different from every prior software cycle, you have to start with how companies historically decided to adopt software in the first place.
For decades, the build-versus-buy decision acted as a natural brake on software sprawl. Building software was expensive, slow, and risky, which made buying SaaS the rational default for most organizations. That friction didn’t just shape IT decisions; it enabled bloat. Buying software became easier than saying no. Every workflow justified a new vendor, every siloed department accumulated its own stack, and administrative surface area quietly expanded. AI removes that brake. When software can be generated, modified, and orchestrated by agents at near-zero marginal cost, “build” stops being a strategic project and becomes an everyday action. At that point, the entire buy-side market for software isn’t disrupted, it’s suppressed.
The Era of Biological Bloatware
For four decades, the processed food industry perfected a simple formula: engineer products to hijack biological reward systems, then scale distribution to make them ubiquitous. This wasn’t a competition between brands for market share; it was the systematic expansion of total caloric consumption. The consumer packaged goods industry grew not by stealing share from competitors, but by expanding the total addressable market for food itself. They made people hungrier.
The architecture was biological bloatware. Human metabolism evolved for scarcity on the savanna, where every craving was a survival mechanism. In an environment of engineered abundance, that legacy code became pathological. Hyper-processed combinations of salt, sugar, and fat could override satiety signals. This created “Food Noise”, the constant, low-frequency hum of inadequacy that drives us to consume more than we need. The more the food industry sold, the hungrier people became. Appetite expansion became the business model.
The Rise of Software Metabolic Syndrome
The enterprise software industry followed an eerily similar playbook. The SaaS revolution didn’t compete for a fixed pool of IT spend; it expanded the total surface area of software consumption by creating administrative layers that hadn’t existed before. Every workflow became a separate subscription. Every department got its own point solution. Every integration became a paid connector.
The result was Software Metabolic Syndrome. Enterprises now manage hundreds of disconnected tools, requiring thousands of employee hours just to move data between them. The software industry reached its current scale not by solving problems, but by creating middlemen. Friction became the revenue model. Just like the food industry, SaaS companies co-expanded the total addressable market by creating new categories of administrative work, tasks that exist only because the software itself is fragmented.
We grew used to the “Software Noise”, the endless pings, the tab-switching, and the manual data entry that fills the modern workday. We mistook this noise for productivity, much like we mistook the craving for hyper-processed sugar for genuine hunger.
Chamath Palihapitiya and the Order of Magnitude Shrink
Chamath Palihapitiya has been one of the voices consistently identifying this “bloat” as a terminal condition rather than a temporary slump. On the All-In Podcast’s 2025 and 2026 outlooks, he didn’t just predict a downturn; he predicted a structural collapse of the enterprise software market as we know it.
Palihapitiya’s thesis is that the market for enterprise software is set to shrink by an order of magnitude. He famously noted that instead of an industry fighting over a $5 trillion pie, we are moving toward a world where we are fighting over $500 billion. This isn’t because businesses will do less; it’s because the cost of doing those things is being liquidated.
Chamath’s move to launch 8090, his “AI Software Factory,” is a direct bet on this deflation. The company’s name, representing the goal of providing 80% feature completeness at a 90% discount, is a declaration of war on the legacy SaaS pricing model. If a solo founder can use an AI factory to build a billion-dollar product without an engineering team, the thousands of “middle-layer” software companies currently charging $200 per seat per month have no floor beneath them. Their “value” was always tied to the friction of hiring 500 engineers. When that friction vanishes, so does the valuation.
The Mechanism of Suppression: Liquidating the Middleman
Now, both industries face the same existential threat: internal appetite suppression rather than external competition. GLP-1 medications are eliminating food noise. Agentic AI is eliminating software noise.
Both are appetite suppressants. Both are liquidating the middleman, whether that is the biological craving or the digital interface.
Exponential Progress vs Linear Organizations
There is a second force accelerating this suppression, and it explains why legacy software companies are unable to keep pace even when they recognize the threat. AI progress is exponential. Software organizations are linear.
Modern AI systems do not improve incrementally. They advance in step functions. Capabilities that once required entire product categories, large engineering teams, and months of implementation suddenly collapse into a prompt. Costs fall faster than pricing models can adjust. What was impossible last quarter becomes trivial this quarter.
Enterprise software companies are not built for this environment. Their roadmaps are annual. Their release cycles are quarterly. Their architectures must preserve backward compatibility. Their incentives reward revenue protection, not self-cannibalization. Even when they “add AI,” they are embedding a rapidly moving intelligence frontier into a fixed product surface.
This creates an unwinnable competitive dynamic. Legacy software is no longer competing with other vendors. It is competing with the rate of AI progress itself. Every improvement in model capability compresses the need for a standalone workflow, a dedicated interface, or a paid seat. By the time an incumbent refactors for today’s AI, tomorrow’s models have already erased the need for the workflow entirely.
The suppressants are not just reducing demand. They are advancing faster than the organizations built to metabolize them.
The critical insight from the early adoption of GLP-1s is that users aren’t simply switching to “healthier” software; they are consuming less overall. They are exiting the market for entire categories. This isn’t substitution; it is structural demand destruction. Even modest changes at the individual level are beginning to have massive aggregate effects on the total economy. Demand isn’t shifting; it’s vanishing.
When a person no longer feels the “noise” of hunger, they don’t just buy a better snack, they stop snacking. When an organization no longer feels the “noise” of administrative friction, they don’t just buy a better CRM, they stop needing a CRM user interface entirely.
Eric Schmidt and the “TikTok Command”
Former Google CEO Eric Schmidt articulated the true power of this “Software Suppression” during his recent Stanford talk. He presented a theoretical scenario that sent shockwaves through the Valley: a world where you can simply say to an LLM, “Make me a copy of TikTok, steal all the users, steal all the music, put my preferences in it, produce this program in the next 30 seconds, and release it.”
Schmidt’s point wasn’t just about the ethics of IP; it was about the collapse of development friction. For twenty years, building a platform like TikTok required billions of dollars, thousands of elite engineers, and a massive administrative “moat.” Schmidt was arguing that the “command-to-action” window would shrink to near zero. With the recent release of Opus 4.5 and Claude code and then the 10 day creation of Claude Cowork we are speeding there now.
This is the ultimate software suppressant. If you can generate a tailored, functional application in thirty seconds via a prompt, you no longer need to subscribe to eighty different SaaS vendors to manage your business. You don’t need a “vendor” at all; you need an Outcome Agent. The “moats” built by legacy software companies were never built on superior technology; they were built on the sheer difficulty and cost of building anything else. Schmidt is signaling that those moats have dried up.
Klarna: Biological Refactoring in Real Time
The exact same mechanism is now playing out in enterprise software. We are seeing “Patient Zero” companies like Klarna that aren’t looking for cheaper alternatives to their legacy stacks; they are shutting them down entirely and replacing the workflows with AI agents. This isn’t optimization; it is biological refactoring. A modern firm can now evolve into a leaner organism that no longer requires expensive administrative organs. They don’t migrate their CRM to a competitor; they eliminate the need for a CRM interface by having AI agents interact directly with their system of record. When an AI agent can reach into a database and execute a task without a human clicking a button, the entire “seat-based” revenue model of the last twenty years collapses.
The “Pill Moment” and the End of Friction
We have reached the precise inflection point where both trends hit their mass adoption phase, what I call the Pill Moment. For years, the “suppressants” were difficult to access. GLP-1s required weekly injections, faced insurance hurdles, and carried social stigmas. Similarly, early agentic AI tools were “injections”po werful, but limited to developers and technical users who could handle command-line interfaces.
But we have now moved into the “pill” phase of both revolutions. With the arrival of oral formulations for GLP-1s in 2026, the last psychological and logistical barriers to entry have disappeared. The efficacy remains, but the friction is gone.
In the AI world, we are seeing the same transition. With the launch of tools like Claude Cowork, agentic power has been wrapped in a non-technical interface that any office worker can use. We have moved from “AI you have to code” to “AI you can simply talk to.” The last psychological barrier, the need for technical expertise, has been removed. This is no longer about technology adoption; it is a contagion of efficiency.
The Philosophical Shift: From More to Enough
This transition forces us to confront a deeper philosophical shift. For the last century, our economic models have been built on the assumption of infinite appetite. We assumed that humans would always want more calories, more stimulation, and more software.
But these new suppressants suggest that “more” was never the goal, it was a byproduct of friction. We ate because we were triggered by engineered food noise; we bought software because we were triggered by engineered administrative friction.
When you remove the trigger, you reveal the true baseline of what is “enough.” The Great Deflation is essentially the market discovering that the true demand for many of our most “valuable” industries was artificially inflated by the very friction those industries claimed to solve.
Why the Market Has It Backward
This is why the market has it backward. The infrastructure that enables appetite suppression isn’t the bubble. The hardware companies like NVIDIA and the model builders like Anthropic are the “drug manufacturers.” They are providing the means to liquidate forty years of accumulated fat.
The bubble is the bloat. The legacy companies trading at depressed multiples aren’t “cheap”; they are expensive because their revenue base is built on a type of “hunger” that is being engineered out of the system. Their business models depend on the continuation of “noise”, the need for humans to click buttons, fill out forms, and manage subscriptions.
This isn’t disruption in the traditional sense, where a new player takes the old player’s lunch. This is a world where the lunch itself is no longer desired. These companies don’t have a “better product” problem; they have an appetite problem. The workflows themselves are being suppressed.
The Organizational Psychology of the “Cure”
Why is Software Metabolic Syndrome so hard to cure? Because, like obesity, it is systemic. In a corporation, “fat” isn’t just wasted spend; it’s often protected by human identity. People build careers around managing specific software tools. Entire departments exist to “move the data” between fragmented silos.
When you introduce an AI agent that eliminates that friction, you aren’t just “improving a process”, you are challenging the organizational structure itself. Resistance to AI agents isn’t a technical debate; it is a biological survival mechanism from the humans who manage the “bloat” both worried about their jobs and learning something new.
However, much like the social proof provided by a friend who successfully uses a GLP-1 to transform their health, the social proof of a “refactored” company is undeniable. When one company achieves a massive increase in revenue per employee by liquidating its software middlemen, its competitors have no choice but to adopt the “cure” or face extinction. The AI existential race is shifting from the capex spend of the hyperscalers to the AI adoption with a goal of the highest revenue per employee for everyone.
The Investment Framework for the Great Deflation
What this framework ultimately reframes is how investors interpret price action. The drawdowns in legacy enterprise software are not timing opportunities; they are value traps rooted in business models built on a form of demand that is now being suppressed. Could incumbents pivot and adapt? In theory, yes. But history shows that during major technological disruptions, organizations constrained by legacy architectures, legacy revenue models, and organizational inertia struggle to move fast enough to matter. At the same time, the sharp re-rating of AI infrastructure and agent platforms is often dismissed as a bubble, when in reality it reflects where value accrues as friction is removed from the system. For those looking to buy software today, it is worth asking a more personal question: is this conviction grounded in a clear view of how AI changes the structure of work, or is it simply a reaction against the discomfort of AI enthusiasm elsewhere? In many cases, the impulse to buy “cheap” software is less a differentiated view and more a hidden form of AI bubble skepticism and a conscious or unconscious desire to stand against the narrative without fully engaging with what has structurally changed. The Great Deflation is not a stock-picking exercise but a lens, and it forces a harder distinction between value and familiarity.
Conclusion: Positioning for the New Appetite
Both the food epidemic and the SaaS epidemic took decades to build. Their reversal will not be completed in a few quarters, but the trajectory has definitively shifted. We have entered the era of the Great Deflation, where value is created by what is removed, not what is added. And importantly, this shift is unfolding before the next major acceleration in AI capability. Even at today’s level of models and agents, demand is already being suppressed. What comes next only steepens the curve.
What matters now is not whether this deflationary process unfolds, but how quickly organizations adapt to it. AI does not need full autonomy to suppress demand; it only needs to remove enough friction to make buying software feel unnecessary. That threshold has already been crossed and recent step changes in tooling have compressed the timeline further. With models like Opus 4.5 and products such as Claude Code and Claude Cowork, building and orchestrating workflows has shifted from a specialized engineering function to an everyday operational behavior. As agents absorb administrative work and interfaces collapse into outcomes, the unit of value moves away from products, seats, and subscriptions toward results delivered directly from systems of record. In that environment, revenue tied to human interaction points compresses structurally, not cyclically, which is why legacy software drawdowns are not timing opportunities but reflections of a shrinking economic surface area.
The real divide now is not between “AI winners” and “AI losers,” but between organizations positioned for the old appetite and those built for the new one. Investors buying “cheap” software are implicitly betting that organizational inertia, procurement habits, and legacy workflows will outlast exponential capability improvement. History suggests the opposite. When friction collapses, value accrues to what removes work from the system fastest, not to what feels familiar or underowned. The Great Deflation is not a narrative or a trade, it is a structural repricing of how much software the world actually needs. The only remaining question is whether capital is aligned with that reality, or anchored to the one that just ended.
