In early 2026, roughly $1 trillion in software market capitalisation evaporated. Anthropic launched Claude Cowork as a research preview on January 12, and when it released a broader enterprise version with deep integrations on February 24, the cumulative damage became unmistakable: $285 billion in enterprise software value wiped out, the IGV software ETF down 22% (its worst stretch since October 2008), Salesforce off 40% from its highs, Monday.com down 22% in a single session. Most commentary has treated this as a correction, a sentiment-driven repricing that will eventually reverse. I think something more interesting is happening, and understanding it requires thinking carefully about what capital value actually represents and where it comes from.
A business is worth the present value of its expected future profits, and profits require a gap between what customers will pay and what it costs to produce. That gap only persists where replication is difficult. If a competitor can assemble the same capability at a similar cost, competition will grind economic profits (the returns above the cost of capital) toward zero, leaving only the baseline return needed to keep capital in the business. Premium valuations exist because investors believe economic profits will persist, and economic profits persist because something about the business is genuinely hard to replicate.
Scarcity is always relative, though, which is where things get interesting. In any economy short of true post-scarcity abundance, every resource is scarce to some degree, simultaneously, and the price mechanism is constantly balancing all of them against each other. The useful question is never “is there scarcity?” but rather “what are the relative scarcities, and how are they shifting?” When one resource becomes relatively more abundant, its price falls relative to others, and economic profits migrate toward whichever resources have become relatively more scarce. GDP can grow, technology can advance, living standards can rise, yet scarcity never disappears: the ratios shift, and capital values follow.
For the past two to three decades, the relative scarcity of cognitive work has been increasing. Globalisation, industrial automation, and digital manufacturing made physical production progressively cheaper, which meant that designing a product, writing code, analysing data, and managing complexity all became relatively more expensive compared to actually manufacturing and distributing things.
Software companies turned out to be the perfect vehicle for capturing the economic profits this shift created. The economics of a SaaS business are strikingly elegant: a thousand engineers build a product once, embedding their cognitive work into code, then sell access to it as a subscription to millions of customers. The marginal cost of serving one additional customer rounds to zero, yet the subscription price reflects the cost each customer would otherwise bear to replicate that cognitive work in-house. SaaS was an arbitrage on the relative scarcity of cognitive output. You were renting access to integrated thinking that was too expensive to replicate yourself, which is a wonderful business to be in, right up until the moment it becomes cheap to replicate.
Per-seat pricing made this arbitrage beautifully explicit: each human who needed access to packaged intelligence represented a unit of demand for the scarce resource. At their 2021 peak, public SaaS companies reached median valuations of 18.6x revenue, a multiple that only makes sense if you believe the cognitive output embedded in the product will remain costly to replicate at the incumbent’s scale and quality for a very long time.
Anyone could write code before AI, of course. The moat was the prohibitive cost of replicating the integrated whole: thousands of engineers, years of iteration through millions of customer interactions, deep integration ecosystems, established distribution and brand. Each component was individually available, yet assembling all of them into a credible competitor took years and hundreds of millions of dollars. That is what kept economic profits elevated and what justified those multiples.
Artificial intelligence is making cognitive output relatively more abundant, which is excellent news for civilisation and rather less excellent news for businesses whose economic profits depend on cognitive output remaining scarce. The effects arrive through two mechanisms worth distinguishing.
The first is on the demand side. AI is increasingly capable of performing the workflows that humans previously needed software to support. When an AI agent can navigate a CRM, run a sales analysis, and generate the output directly, fewer humans need to sit in front of the software to accomplish the same work. Enterprise customers are already reporting what the industry calls “seat compression” at contract renewals: organisations that previously licensed 500 seats are renewing for 50, because ten people plus an AI agent can now do the work that previously required a hundred people plus a SaaS product. The seats are disappearing because the cognitive work they represented is being done by something cheaper.
The second mechanism is on the supply side, and it goes straight at the replication cost that sustained the moat. AI is lowering the cost of building software itself. When the cognitive work required to create a software product becomes relatively more abundant, the total cost of assembling a credible competing product falls with it. Something that previously required a thousand engineers and ten years might now require fifty engineers and two. The integrated whole that once took hundreds of millions to replicate becomes replicable for a fraction, and as more competitors enter, margins compress regardless of how sticky the incumbent’s existing customer relationships might be.
The market has noticed. The sector’s forward price-to-earnings ratio has collapsed from roughly 35x at the end of 2025 to around 20x, a level last seen in 2014. Median public SaaS valuations have fallen to 6-7x revenue from nearly 19x at the peak. You can read those numbers as a repricing of the relative scarcity of cognitive output, and with it, the durability of the economic profits that software businesses have historically earned.
There is a useful distinction between first-order and second-order effects of a new technology. The first-order effects are visible and priced quickly: revenues at Nvidia surge, Microsoft integrates copilots, investors bid up anything adjacent to AI. The second-order effects take longer to arrive, and they tend to surprise people because they work in the opposite direction. They show up when the technology diffuses broadly enough that it begins to erode the very scarcity it initially seemed to enhance.
For most of 2024 and 2025, the market priced AI as additive to technology company valuations: a new feature, a new revenue line, a reason to pay more for software businesses. February 2026 was the month the second-order effects became visible. Anthropic’s product launches demonstrated that AI could increasingly substitute for the human cognitive work that software companies exist to support, putting downward pressure on the relative scarcity those companies were monetising. Wall Street, it turns out, had been paying a premium for the thing that was about to become cheap.
The underlying trends had been building quietly for some time. Public SaaS revenue growth has declined every quarter since Q2 2021, with the median falling to 12.2% by Q4 2025 from 20-25% in prior years. Price increases had been masking the slowdown in organic growth, a strategy that works only until customers realise they need fewer seats at any price. The February selloff was the catalyst rather than the cause: the structural erosion of cognitive scarcity had been underway well before the market acknowledged it. Perhaps most amusingly, 70% of software providers now report that the cost of delivering AI features is eating into their margins, which means these companies are spending money to build the very technology that compresses their own economic profits. It is difficult to think of a purer example of Schumpeterian creative destruction.
If cognitive output is becoming relatively more abundant, the price mechanism tells us that other resources must be becoming relatively more scarce, and their prices should be rising accordingly. The most obvious candidates are physical: energy, materials, land, and infrastructure. AI’s demand for these is large and growing, while their supply is constrained by factors that cleverness alone cannot quickly resolve.
Energy is the clearest case. AI requires enormous quantities of electricity, the demand curve is steepening, and new generation capacity takes years to permit and build. Hyperscalers have announced over $470 billion in AI infrastructure spending for 2026, the vast majority of which will flow into physical assets: data centres, power generation, cooling systems, and transmission capacity. Each of these requires land, materials, construction labour, and regulatory approval, none of which become cheaper because cognitive output is more abundant. If anything, the surge in demand makes them relatively more expensive, which is exactly what the framework predicts.
Semiconductor manufacturing tells a similar story. A leading-edge TSMC fab costs upwards of $20 billion and takes years to construct. The extreme ultraviolet lithography equipment required to run it is produced by a single company on earth (ASML, for those keeping score at home). High-bandwidth memory, critical to AI workloads, is manufactured by a handful of firms whose decades of accumulated process knowledge is inseparable from the physical plant itself. You can throw all the intelligence you like at optimising these processes at the margin, yet the fundamental constraints are physical: the cleanroom, the equipment, the yield curves earned through years of production, the supply chains for exotic materials.
Critical minerals face geological constraints and multi-year development timelines. Land remains fixed in supply, as it has been since Ricardo first observed it two centuries ago. Regulatory and permitting bottlenecks, far from yielding to AI, may represent one of the more durable forms of scarcity in a modern economy: it turns out that the ability to get permission to build something in the physical world is extraordinarily hard to replicate or circumvent, which makes it exactly the sort of resource that earns economic profits.
The pattern is consistent across all of these examples. The more successful AI becomes, the more physical resources it demands. The technology that makes cognitive output relatively more abundant simultaneously increases the relative scarcity of physical resources, and capital follows the scarcity.
Software valuations are compressing while capital flows toward AI infrastructure, energy assets, semiconductor manufacturing, and physical supply chains. The direction of flow is already visible in the data, and the framework suggests it has further to run, because the forces driving it are accelerating rather than fading. AI capability is improving rapidly, which means cognitive output will become relatively more abundant, not less, and each increment of AI progress puts further pressure on software margins while increasing demand for physical resources.
Two caveats are worth stating plainly. Owning “atoms” is a necessary condition for benefiting from this rotation, not a sufficient one. The same competitive dynamics that erode software margins apply within the physical economy. Mining companies are often commodity price-takers with poor returns on capital. A power plant in a deregulated market may not capture much of the scarcity rent. The framework identifies the direction of the rotation, not which specific businesses will earn durable economic profits within it; that still requires finding physical assets with genuine bottlenecks that competition cannot easily replicate on the timescales that matter.
The physical resources that are relatively scarce today will not necessarily remain so, either. Technology has historically resolved physical scarcities: solar costs have fallen 99% in four decades, and the cost of computing itself has declined by orders of magnitude. The rotation toward atoms could eventually be followed by another shift as new technologies expand physical supply. The claim here is not that physical scarcity is permanent, only that the current trajectory of AI development is shifting the ratios toward physical resources and that this shift has further to run before it is fully priced.
Every major technological revolution has shifted the relative scarcities across resources, and capital values have followed with a lag. The loom made cloth relatively more abundant and the machinery to produce it relatively more scarce. The automobile made transport relatively more abundant and oil, roads, and suburban land relatively more scarce. The internet made information distribution relatively more abundant and the cognitive work required to create, analyse, and act on that information relatively more scarce.
In each case, the technology itself was enormously beneficial to society: living standards rose and total economic output grew. Scarcity never disappeared, because it never does so long as human desires outrun the resources available to satisfy them. The ratios shifted, and the investors who understood which resources were becoming relatively more scarce, rather than which were becoming more abundant, were the ones who preserved and grew their capital.
The cognitive scarcity era was a historically specific period in which cognitive work was the relatively scarce factor, commanding premium prices and generating outsized returns for the businesses that controlled it. That era is ending. What comes next is a shift in the ratios, perhaps a great rotation, from bits back to atoms.