A POST-SCARCITY MANIFESTO ON SCARCITY

THESEUS CAPITAL is an investor and operator in businesses that drive civilizational progress for humanity. Our mission is to steer capitalism’s acceleration of technological progress towards what is unchanging and valuable with respect to humanity’s prosperity and longevity.

HUMANITY is akin to a ship sailing across a vast, uncharted ocean. To reach new worlds, we must first understand the currents quietly setting our course, then be able to steer the ship with strength and confidence.

To accelerate the journey, we may upgrade the ship; replace thin sails with robust engines, navigate with advanced GPS instead of faulty compasses, until at some point, the ship could run on its own without any hands needed at the helm.

When technology replaces plank after plank of human life, will we still be the same ship? What will our purpose be then? Where will we wash ashore?

CAPITALISM is a fundamental force that has set humanity’s course for centuries and will continue to shape it for many more.

At its core, it can be understood as a self-organizing and self-expanding process. Imagine a firm that produces and sells goods for a profit. Through an accumulation of know-how, it improves its operational efficiency and lowers its cost of production. The firm, facing market competition, reinvests its surplus in R&D to further reduce costs and maintain price competitiveness. Finally, since the most reliable way to lower costs is to improve the means of production, capitalism selects for technical advance as a matter of structure.

The firm’s newly achieved efficiency then expands the market rather than conserves, as Jevons observed in 18651 when better engines raised British coal consumption instead of reducing it, because cheaper power made economical a whole range of uses that had previously been out of reach.

The history of capitalism, then, is this loop closing on itself and widening. For instance:

• Mechanized spinning in Lancashire cut the labor required to produce a yard of cloth2, and falling prices opened markets to people who had worn homespun. This provided funding for further mechanization that demanded better steam engines, which then transformed transport and mining.
• Railways and trains collapsed the cost of moving goods overland and fused regional markets into a single continental one, which raised the scale at which any producer could profitably operate and rewarded whoever mechanized fastest to serve it. Their appetite for iron, coal, and steel expanded the industries that supplied them, and their capital requirements ran so far beyond private fortunes that financing them produced the modern corporation and the modern capital market3.
• The invention of containerization collapsed the cost of ocean freight by roughly 90-95% after 19564, which severed the link between where a good is made and where it is sold, since manufacturers no longer had to locate near their customers. Thus, production migrated to the cheapest available labor along the Chinese coast5, and factories of the North Atlantic closed.
• In semiconductor manufacturing, the work of designing a chip with hundreds of millions of transistors exceeds what any engineer can lay out by hand, so the layout is performed by software, and that software runs on chips produced by the previous generation of the same process.

Time and time again, we bear witness to capitalism developing technologies that produce the instruments of their own successors. With each wave of innovation, from textiles to railways to containers to chips, the time between these cycles has tended to grow shorter as the underlying technologies become more powerful.

Historical attempts to forestall capitalism’s seemingly anti-human, mechanistic cycle has failed stunningly, resulting in opposite intended effects. The most serious attempt by the Soviet Union rejected the free market because they saw it producing immiseration alongside abundance, and because they took its crises and its class divisions to be permanent features rather than growing pains. Their solution was to place investment under a single plan, so that what got built would answer to human decision rather than to profit. During the late 1920s and 30s, their plan seemingly worked: an agrarian empire was transformed into an industrial power in two decades. However, the Soviets did so by hiring American engineers to build the factories, modeling the steel city of Magnitogorsk on the United States Steel works in Gary, Indiana6, and contracting with Ford to build its automobile plant. It had taken capitalism's technology and rejected only its pricing. What it lost in the bargain was the information that prices carry7, which is why the planners could build rockets on command and never learn to make a decent refrigerator8. In the end, the Soviets’ system collapsed in 1991, and China had already reversed course in 1978 when Deng Xiaoping’s reforms marked a shift away from Map-era central planning to wards a more market-oriented “socialism with Chinese characteristics.” Over the following decades, sweeping changes in agriculture, special economic zones, and export-led manufacturing helped drive sustained double-digit growth, lifting hundreds of millions out of poverty and propelling China to industrialize faster, and on a larger scale, than any country in history.

Fast forward to today, we believe that capitalism’s self-reinforcing loop will run faster than ever with artificial intelligence, until we reach a point of technological singularity where any good or service can be produced abundantly with widespread standardization and automation (at near zero marginal costs). But in a world where everything is cheap and abundant, do goods and services become valueless? No. Even if technology solves the scarcity of physical goods and basic services, economic theory dictates that human desires are infinite. Therefore, if technology makes all our current needs incredibly cheap, our desires will simply shift upward, creating new categories of “expensive” things.

We are in the business of investing in this post-scarcity society and its technological transition in the interim.

IN THE EARLY 1880s, electricity arrived as a luxury good that had to be custom built for each buyer. There were no monthly bill and no wire running in from the street. A customer who wanted electric light had to purchase a generator, hire an engineer to run it, and install a small power plant somewhere in the building. J.P. Morgan did exactly this at his New York mansion, where Thomas Edison's people installed a private plant to light the rooms. Electricity in that period was local, costly, and understood as the answer to a single problem, which was replacing gas lamps.

The first constraint on growth was physical. Edison's direct current could carry usable power only about a mile from its source before the voltage sagged, which meant that every neighborhood needed its own generating station. A system built that way could never behave like a commodity, because each new district of customers required another plant and another crew to operate it. Then came George Westinghouse and Nikola Tesla, who commercialized alternating current (AC). AC could be stepped up to very high voltage for transmission across long distances and then stepped back down to a safe level at the point of use9, rendering scaling physically and mathematically possible in roughly the way the Transformer architecture later did for AI.

The economics were solved by Samuel Insull10, who was Edison's secretary and became the true architect of electricity as a mass market product. He understood the industry's central financial problem, which was that generating plants cost enormous sums to build while households used them for only a few hours each evening, leaving expensive machinery idle for most of the day. His answer was to aggregate demand until the plants could run near capacity around the clock. He bought up small, decentralized neighborhood grids and wired them together into large, centralized networks. He introduced time of use pricing, charging different rates at different hours so that factories were pulled toward daytime power while homes filled the evenings. The result created a flywheel effect: large and steady demand justified building giant, highly efficient turbines, those turbines drove the cost of electricity down, and cheap power invited more consumption, which justified the next round of capacity expansion.

The last requirement to commoditization was interchangeability, since a commodity must be the same everywhere. Early cities ran on a confusing variety of voltages and frequencies, so a device built for one town might be useless in the next. Over time the industry converged on common standards, and in the United States that settled at 120 volts and 60 hertz. The two-prong wall outlet became the universal interface of the twentieth century, the equivalent of a public API. Once the plug was fixed, an inventor anywhere could design a product with reasonable confidence that it would work in any building in the country.

The most striking thing about the history of electricity is that almost no one anticipated it would become a commodity. The public treated electricity as a better candle and assumed that illumination was the whole of it. The idea that cheap and ubiquitous power would produce entirely new categories of sat outside the ordinary imagination of the period. Refrigerators, washing machines, electric assembly lines, and eventually computers all followed from a resource that had become so inexpensive and so reliable that people stopped thinking about it. The commodity was the precondition, and the applications arrived afterward, built by people who never had to consider where the power came from.

We are long AI’s commodification and long the downstream innovations that would continue to benefit humanity.

COMMODITIZATION is usually described as something that happens to a product, as if the good itself gradually loses its distinctiveness through the ordinary passage of time. The more accurate account is that buyers do it, and they do it deliberately once a product becomes indispensable to their operations but hurts the bottom line. The pressure runs from the demand side back toward the supplier, and it operates through a few reinforcing mechanisms.

• Interchangeability. When several producers can supply roughly the same thing, buyers stop caring about the identity of the seller and begin caring only about whether quality and reliability clear a threshold. Above that threshold, the purchase decision collapses into price. A supplier who tries to hold a premium discovers that switching costs the customer very little, and price competition turns brutal.
• Supply balance. As capacity expands, whether that means more power plants, chips, or data centers, the balance of power shifts further toward the buyer. When capacity exceeds demand, producers underbid one another to keep utilization high and avoid the visible waste of idle assets.
• Capital intensity. Rigs, pipelines, fabs, and GPU clusters are expensive to build and largely sunk once built. After the capital is committed, the operator has a powerful incentive to produce and sell at almost any price above short-run variable cost, since a poorly priced asset at least contributes something toward fixed costs. That logic is individually rational for every operator and collectively ruinous for the industry, and it keeps both prices and returns low over long stretches.
• Buyer base composition. Bargaining power tends to concentrate in large and sophisticated buyers who can multi-source, negotiate with real information about supplier costs, and move volume between vendors as leverage. These buyers capture the surplus in the form of lower prices while the producers compete away their own profits.

A few historic, buyer-driven commoditization examples reinforce these patterns:

Buyer-driven pressures

In the early mainframe era, buyers were entirely locked into IBM’s expensive, proprietary hardware and metered time-sharing contracts, which bottlenecked corporate innovation. Frustrated by exorbitant costs and long queues for computing time, businesses flocked to cheaper minicomputers and, eventually, personal computers. This mass buyer defection forced the industry to shift from bespoke, leased behemoths to mass-produced, standardized, and interoperable hardware (like the x86 architecture), turning computing power into an accessible commodity.

For decades, consumers and enterprises were squeezed by telecom monopolies (like AT&T) that charged massive per-minute premiums for long-distance routing. Corporate buyers actively lobbied for deregulation and eagerly funded upstart competitors (like MCI and Sprint) to drive prices down. Ultimately, businesses bypassed legacy telecom infrastructure entirely by adopting the standardized, open-source Internet Protocol (IP), which turned voice and data transit from a metered luxury into a cheap, flat-rate, interchangeable pipe.

Throughout the 90s, companies like Oracle and SAP forced buyers to pay millions in upfront capital expenditures for rigid, perpetual licenses and punitive yearly maintenance fees. Fed up with predatory software audits and paying for "shelfware" (unused licenses), CIOs aggressively shifted their budgets to early SaaS disruptors like Salesforce. By demanding pay-as-you-go, standard browser-based access, buyers forced legacy software giants to abandon their bespoke installations and commoditize their offerings into cheaper, standardized monthly subscriptions.

In the early days of the cloud, AWS, Azure, and Google Cloud successfully charged premium margins by locking buyers into proprietary managed services and exorbitant data egress fees. As cloud bills ballooned into massive operating expenses, corporate buyers and startups realized they had traded software lock-in for infrastructure lock-in. To regain some negotiating power, they increasingly adopted open‑source, cloud‑agnostic orchestration tools such as Kubernetes and Terraform. These tools made workloads more portable in principle and strengthened buyers’ ability to push back on pricing, especially for basic compute and storage.

Explosive enterprise AI consumption has led to massive cost blowouts. For instance, Uber exhausted its entire annual AI budget by April following a December rollout11. To combat these soaring inference costs, corporate buyers are actively working to commoditize large language models by bypassing expensive, frontier-model lock-in. Coinbase CEO Brian Armstrong recently highlighted this buyer-driven push in an X post, noting that his firm cut its AI spend nearly in half even as its token usage continued to grow12. They achieved this by deploying internal AI gateways that default to cost-effective open-weight models, utilize aggressive caching, and automatically route each prompt to the cheapest capable model based on the specific task's difficulty.

We believe that the commoditization of AI has already begun. Despite a lower global adoption rate, the collective actions and demands of its main buyers---enterprises---will determine both the degree of closed-sourced models’ commoditization and frontier labs’ pricing power.

SUPPLIER-DRIVE PRESSURES could accelerate model commoditization just as much as buyer-driven pressures. The simple logic here is that, since the suppliers of AI infrastructure’s revenues depend on sustained capital expenditure, their incentives are not to build until there is enough capacity; rather, they would like to build as much as possible while demand for AI endures and the market sentiment runs high. Therefore, actual demand may be grossly overestimated from the top-down by these manufacturers to continuously incentivize investors and frontier labs to pour more money into infrastructure build outs. We see this scenario play out in a couple of prominent historical examples when new technologies similarly arrived and promised widespread usage, yet actual demand was far below capacity build out.

Supplier-driven pressures

RAILROADS, 1800s

The railroad boom unfolded as a sequence of manias that followed a similar arc, first in Britain and then, on a larger scale, in the United States. In the British Railway Mania of the 1840s, petitions to Parliament for new companies exploded, investment briefly reached wartime levels as a share of GDP13, and middle-class savers piled in via partially paid shares—until higher interest rates slammed the funding window shut and share prices collapsed, leaving many lines unbuilt and others consolidated by stronger survivors.

Across the Atlantic, the American railroad boom repeated this pattern of speculative overbuilding, financial collapse, and durable physical legacy, but at far greater scale. It arrived in waves through the 1870s and 1880s, and it is in this American version that the underlying incentive problems are clearest.

• First, financing was decoupled from operating economics. American railroads were funded by land grants, subsidies, and bonds sold to distant investors who were buying a story about the future of the continent, while the promoters assembling these ventures earned their profits on construction itself. The Crédit Mobilier scandal is the canonical case14: Union Pacific insiders owned the construction company, overbilled the railroad, and extracted returns regardless of expected ROI on capex. When builders are paid for building rather than for operating, overcapacity is a designed outcome.
• Second, competition drove duplication. A railroad is a natural monopoly on its route, so the rational play for a rival financier was to build a parallel line and force either a rate war or a buyout, as Vanderbilt's West Shore fight along the Hudson demonstrated. Each redundant trunk line was individually defensible and collectively they ensured that no owner earned adequate returns.
• Third, high fixed costs met falling prices. Because building a railroad required massive upfront investment—land, grading, track, bridges, stations, and rolling stock—an operator's costs were overwhelmingly fixed. These assets were heavily financed with borrowed money, meaning bond coupons came due every quarter regardless of whether a single ton of freight moved. Once a train was already running, however, the marginal cost of hauling one additional carload was practically nothing, requiring only a bit of coal, minor wear and tear, and a little labor. As a result, roughly 80 to 90 percent of a railroad's costs were fixed15, while its marginal cost hovered near zero. When capacity exceeded demand, this unique cost structure relentlessly forced prices down. Once debt is incurred, it becomes a sunk cost, which is entirely irrelevant to pricing decisions at the margin. For example, a railroad's fully loaded cost to haul a ton of wheat from Chicago to New York—including its share of debt service—might be $1.00, while the pure marginal cost is just $0.10. If a competing parallel line offers to haul that freight for $0.80, the first railroad has no choice but to match it. Any rate above the $0.10 marginal cost contributes something toward those fixed bond obligations, whereas losing the shipment contributes nothing at all. This dynamic sparks a race to the bottom. In a normal market, these chronically unprofitable prices would drive capacity out, allowing rates to naturally recover. Railroads, however, faced two unique barriers that made escape impossible. First, railroad capacity simply did not exit the market. A new owner could buy the bankrupt assets for cents on the dollar, instantly granting them a drastically lower fixed-cost base. This allowed the reorganized railroad to profitably charge even lower rates, perversely making the price wars worse after each bankruptcy. Second, while operators fully understood this trap and repeatedly tried to stabilize the market through pools and cartels, these agreements were doomed to fail. Every cartel member faced the same relentless temptation to secretly shave rates to fill their own trains, and because rate agreements were unenforceable in court, cheating was immediate and rampant. Ultimately, the industry was so cornered by its own economics that the Interstate Commerce Act of 1887 emerged16, in part, from the railroads' own desperate desire for the government to enforce the pricing discipline they could not impose on themselves.

The investment lesson lies in where the value ultimately settled. The track survived the financial destruction of its builders, and the enduring gains flowed to the users of cheap freight: farmers, steel producers, mail-order retailers such as Sears17, and ultimately consumers, who harvested the benefits for the next half century by enjoying cheap goods. The original capital financed a transformation whose returns accrued mainly to its customers and to the second-generation owners who bought the assets after the wipeout.

FIBER AND TELECOM, 1996—2002

The telecommunications industry had experienced significant growth and investment during the 1990s, fueled by the expansion of the internet and the introduction of wireless technology. Companies such as WorldCom, Global Crossing, and Lucent Technologies had achieved enormous market valuations based on expectations of continued growth and profitability. Total US telecom capital expenditure over the period ran to roughly half a trillion dollars, with more than a trillion dollars of debt and equity raised globally to fund the buildout. The demand assumption underpinning all of this capital was around the claim that internet traffic was doubling every 100 days, a figure popularized by WorldCom's UUNET subsidiary and repeated in a 1998 Commerce Department report18. Actual traffic was doubling roughly once a year19, which still represented spectacular growth, but the difference between eightfold annual growth and twofold annual growth compounds catastrophically when it is used to size a network buildout.

The most underappreciated mechanism of the bust was that technology multiplied supply faster than demand could grow. Dense wavelength-division multiplexing (DWDM) improved so rapidly that the carrying capacity of a single fiber pair already in the ground rose by orders of magnitude20 during the buildout itself. Carriers were laying conduit containing dozens of fiber strands while the effective capacity of each strand grew by factors of tens to hundreds. Supply was therefore expanding along two axes at once, through new physical construction and through the escalating productivity of existing assets. By 2002, common estimates held that only 3 to 5 percent of installed fiber was actually lit21. Bandwidth prices on major routes fell more than 90 percent22, and the collapse in the price per bit destroyed every revenue model that had been premised on scarcity. The lesson here is that in any capacity buildout, the supply forecast must account for the productivity curve of the underlying technology, because efficiency gains function as invisible additional capacity.

As real revenues fell short, financial reflexivity and fraud filled the gap. Carriers swapped capacity via indefeasible rights of use, booking the sales as revenue while capitalizing the purchases, thereby manufacturing growth without net economic activity. Global Crossing and Qwest were notable practitioners, and WorldCom went further, capitalizing roughly $11 billion of operating expenses. Vendor financing added another loop: Lucent, Nortel, and Cisco lent carriers the money to buy their equipment, so reported growth partly reflected the vendors’ own balance sheets rather than genuine demand. When the cycle turned, vendors absorbed those losses alongside the carriers, which is why Nortel and Lucent fell as hard as the network operators they supplied.

Between 2000 and 2002, telecom companies lost on the order of trillions of dollars in market value. WorldCom’s failure became the largest US bankruptcy in history at the time; Global Crossing and 360networks collapsed outright; and sector employment fell by hundreds of thousands. Yet once again, the asset outlived its financiers. Dark fiber was bought out of bankruptcy for cents on the dollar and became the substrate for subsequent internet companies. Google quietly accumulated distressed fiber and backbone capacity in the early 2000s, and cheap, overbuilt bandwidth helped make largescale video services like YouTube and Netflix economically viable as prices fell toward marginal cost. Fiber laid by bankrupt carriers in 1999 was still being lit a decade or two later, and the secondgeneration owners, who bought at postbankruptcy prices, earned the returns the original builders had projected for themselves.

MECHANISMS OF COMMODITIZATION

Commoditization as Prosperity

In this scenario, the hyperscalers’ race to build massive GPU clusters results in a massive supply glut. AI becomes the new electricity—cheap, standardized, and universally accessible. As Meta, Microsoft, AWS, and Google overbuild, switching costs plummet. Open-source models and "good enough" proprietary models converge in capability. The cost of inference drops to near zero as hyperscalers treat compute as a loss leader to keep customers in their cloud ecosystems.

In this scenario, traditional enterprises and consumers win. The big gains accrue to non-AI businesses who reap massive productivity gains and margin expansion. The losers are the foundational AI labs and data center operators. Returns on capital for new GPU clusters collapse and selling "intelligence" becomes a low-margin, high-volume utility business.

Shifting Rents

Here, the algorithmic magic of AI commoditizes, but the physical reality of running it does not. The "brain" becomes cheap, while the "calories" to run it become exceptionally expensive. For instance, open-source models and rapid algorithmic diffusion offer near-SOTA intelligence, but the physical infrastructure—TSMC wafer allocations, power grid connections, and advanced networking—remains fiercely constrained. Thus, startups can download incredibly capable open-weight models for free, but they can't afford the cloud instances to run them at scale. The leverage entirely leaves the software layer and moves upstream. AI labs find themselves squeezed between pricing pressure from competitors and unyielding hardware costs from their suppliers.

In this scenario, the winners are the upstream infrastructure monopolies: NVIDIA, TSMC, major utility companies, and firms (real estate and alternative asset managers) that own data centers with guaranteed power contracts. The losers are the model builders and SaaS wrappers, whose margins are continuously eaten by their cloud hosting bills.

Non-Commoditization

In this scenario, compute demands wildly outpace physical reality. Revenue demand for AI could grow 10x, but compute capacity can only scale by a fraction of that due to the death of Moore's Law, slow fab construction, and stringent energy limits. As a result, frontier labs must bid aggressively for massive, high-security compute tranches just to train the next generation of models, which drives server prices far above standard spot rates. Because the baseline cost of compute is so exorbitantly high, buyers only want to use the absolute best, most efficient models to maximize the return on investment of every single computation. Thus, the Alchian-Allen effect takes place23 and leaves no market for a “second-best” closed-source AI. Structurally, these mid-tier AI labs are also forced to pay the same massive compute costs as the top labs but lack the intelligence and/or capabilities to charge high prices, ultimately forcing them into bankruptcy.

Crucially, this dynamic prevents a pure shifting rents scenario where infrastructure providers capture all the industry's value. Because the top models can do the work of a senior software engineer or a corporate lawyer, the model layer does not commoditize into a race to the bottom. Instead, the sheer cost of infrastructure acts as an impenetrable barrier to entry that thins the herd. The top one or two surviving frontier labs command massive pricing power over what is essentially a normal/luxury good, while open-source models face commoditization pressures and become a “good enough” inferior good that powers the infrastructure for non-elite SMEs and consumer usage. The immense economic rents of the AI revolution are therefore shared between the infrastructure oligopoly and the frontier model duopoly, rather than being squeezed entirely down to the hardware layer.

Another scenario in which non-commoditization plays out is one where frontier labs solidify absolute dominance by shifting their moat from compute to cognition. Once deployed by elite enterprises, these models utilize continual learning to adapt to the specific, proprietary workflows, lexicons, and strategic habits of their buyers. The AI evolves from a generalized tool into a highly personalized asset, absorbing institutional memory with every interaction and creating a powerful intelligence flywheel. As the model becomes deeply integrated into core enterprise infrastructure, an invisible barrier emerges in the form of prohibitive switching costs.

Even if a rival lab or infrastructure provider were to somehow offer a cheaper model, the cost of forgetting becomes insurmountable. Ripping out an entrenched, continually learning model means sacrificing years of tailored, compounded knowledge. Ultimately, the winners of this scenario are the top frontier labs and the elite enterprises that can afford them. By combining the pricing power born from physical scarcity with the unassailable lock-in of continual learning, these labs could transform their models from expensive software into the irreplaceable cognitive nervous systems of modern business.

However, as of August 2026, it looks increasingly the case we have largely moved away from the third scenario towards the second, and as the Hyperscalers continue to add compute capacity, we may move towards the first.

THE SHIP OF THESEUS is ultimately the same ship by the judge of its exterior form and not its material composition. By way of analogy, we believe that technology’s impact in what is scarce and valuable for humanity will be limited, because technological cycles’ nature is transient, whereas human beings’ DNA is coded to be timeless and unchanging. The same human mechanisms of mimetic desire, time and attention, and status and authenticity has persisted throughout our history.

While our core human qualities have endured technological revolutions, many tech products have been designed less to cultivate what is uniquely valuable about us than to exploit our most primal vices. In our past hunter-gatherer society, status meant survival. Today, because technology has made basic goods cheap, we fulfill our primal need for status through hyper-consumption of positional goods. Social media, a massive technological revolution, is essentially a global engine for primal status-signaling and social comparison. In another example, the internet was theoretically supposed to create a unified "global village" by democratizing information. Instead, it accentuated our primal instinct to form tribes. Algorithms cater to our confirmation bias, naturally sorting us into hyper-specific ideological or cultural tribes that defend their boundaries fiercely. Lastly, as automation and AI remove human interaction from routine transactions (like self-checkout or automated customer service), our primal need to connect with other humans becomes a premium commodity, further exploited by apps like Tinder. Counterintuitively, we now place higher value on things that display "authentic" human effort, flaws, and emotional resonance precisely because the world has grown so algorithmicized and mechanical.

THE RED QUEEN

In evolutionary biology, there is a concept called the "Red Queen Hypothesis" that says that species must constantly adapt, evolve, and proliferate in order to survive while pitted against ever-evolving opposing species. Coined in 1973 by evolutionary biologist Leigh Van Valen24, the hypothesis was named after a scene in Lewis Carroll’s Through the Looking-Glass25, where Alice finds herself running frantically alongside the Red Queen, only to realize the scenery around them isn't moving. When Alice points out that in her country, running usually gets you somewhere, the Red Queen replies:

"Now, here, you see, it takes all the running you can do, to keep in the same place."

In technology, the Red Queen effect explains why companies can never stop innovating, even if their profits do not dramatically increase as a result. For instance, when Amazon introduced free two-day shipping in 200526, it was a massive competitive advantage. Today, two-day shipping is the baseline expectation for e-commerce, and Amazon's competitors had to spend billions of dollars upgrading their logistics just to stay in business. Similarly, if Coca-Cola spends $1 billion on advertising, Pepsi must also spend $1 billion just to maintain its current market share. The $2 billion spent between them doesn't necessarily create new soda drinkers; it just maintains the stalemate.

Equally salient is the hypothesis’s stab at the paradox of modern human life. Despite living in an era of unprecedented technological abundance, we often do not feel any happier or more secure than our ancestors. The hedonic treadmill shows27 that as technology makes our lives easier, our brains rapidly adapt to the new comfort level. What was a luxury yesterday (like smartphones, air travel, or air conditioning) becomes an absolute necessity today. We must keep acquiring new experiences and better technologies just to maintain our baseline level of happiness. Because humans are deeply social creatures, our sense of success is often relative, not absolute. If you get a 10% raise at work, you feel great, but if you find out all your peers got a 20% raise28, you suddenly feel poor, even though your absolute wealth increased. In a society where technology makes everyone richer, the markers of status simply move further out of reach.

Such is the nature of human beings: we run a lifelong marathon simply to not fall behind; yet we have no other choice but to participate in it and to feel fulfilled and happy. But at Theseus, we don’t think this has to be the case. We believe technological progress can be steered to make products that genuinely improve the human condition, not exploit our vices.

As artificial intelligence drives the production of commodity goods and routine knowledge work toward zero marginal cost, we believe that economic value will dramatically shift toward the "relational sector". Coined by Alex Imas, Chief AGI Economist at Google DeepMind29, this sector is characterized by human-intensive, provenance-rich professions where consumers specifically desire a human in the loop. We hold conviction in this future economy because, as technology fulfills our basic needs cheaply and abundantly, consumers will move up Maslow’s hierarchy of needs30, placing a much higher value on authentic human effort, emotional resonance, and even human flaws to gain authentic connection and fulfillment.

Structurally speaking, the relational sector will consolidate around two distinct poles: the highly charismatic, authentic, "front-end" services that can command a premium, and the massive AI infrastructure platforms that enable, route, and capture the revenue from these relational businesses in the background. As investors, we are opportunistic on both ends of the technology sector and relational goods sector as we face a bifurcated mandate. To generate outsized returns, capital must be aggressively deployed at the two extreme ends of the barbell: we must own the most valuable, moat-heavy parts of the technological stack that powers this new economy, while simultaneously capturing the growth in scarcity-driven relational goods and services. We believe that the middle ground of standardized, mass-market services will be decimated by automation and margin compression.

I. Why the Middle Fails

Our conviction rests on three bodies of theory in economics, behavioral science, and marketing that converge on a single conclusion. Value does not accrue to layers of activity but to positions the buyer cannot escape and to outcomes the customer can perceive and credit.

The economics of derived demand, appropriability, and added value.

The psychology of evaluability, attribution, and the division of labor.

The marketing discipline of distance to the outcome.

II. Paired Case Studies

Technology, in LangChain versus Cursor. These two AI companies founded within months of each other doing similar, highly sophisticated background work saw their valuations wildly diverge by roughly fifty times over four years. The massive difference wasn't their underlying technology, but rather how they chose to package it.

LangChain launched in 2022 as a free, open-source framework to help developers connect AI models to other systems and databases.

It became incredibly popular and raised massive funding, reaching a $1.25 billion valuation by October 202547.

However, its actual revenue lagged far behind its fame, hitting only about $16 million in 202548.

Why the disconnect? LangChain gave away its core product for free, and competent engineers could replicate its main functions in a matter of weeks. Furthermore, major AI companies started building those same features directly into their own models. Ultimately, LangChain had to make its money on adjacent management and deployment tools because it realized it couldn't charge for the core technology itself.

Cursor, an AI coding tool built by Anysphere, uses similarly complex AI coordination behind the scenes.

Instead of selling the underlying technical tools, Cursor hid the complexity inside an application and sold the highly visible final outcome: working code appearing directly on a developer's screen.

Cursor intentionally priced its product against expensive engineering salaries (framing it as a productivity investment) rather than standard tooling budgets (which companies view as overhead).

Because buyers could easily see and value the immediate results, revenue skyrocketed from roughly $100 million in January 2025 to about $4 billion by mid-202649. This success culminated in a historic $60 billion acquisition by SpaceX50.

But Cursor's strategy isn't flawless. Since they have to buy access to the underlying AI models, their operating costs are very high—costing $0.40 to $0.70 for every dollar of revenue51—resulting in profit margins much lower than traditional software. Additionally, the very companies supplying Cursor with AI models are now moving into their territory by building their own competing coding agents.

However, the takeaway here is to not sell the underlying mechanics; sell the finished product. Customers are willing to pay vastly more for a completed, visible outcome that saves them expensive labor than for the complex tools required to build it themselves.

The relational sector, in Equinox versus LA Fitness. Gyms face a unique problem: they cannot do the actual workout for their customers. Because the client must put in the effort, the physical results are delayed, uncertain, and credited entirely to the customer's own hard work.

Companies like LA Fitness and Planet Fitness accept the natural limits of the gym business and simply sell access to workout equipment. Because equipment is incredibly easy to substitute with rival gyms, running shoes, or free videos, prices are aggressively forced down to roughly $10 to $40 a month52. While profitable at scale, this model lacks pricing power and relies heavily on members who pay their monthly fee but rarely show up. If you only sell a functional tool that requires the customer to do the work, you will be constantly forced to compete on price.

Equinox escapes the standard gym trap by selling an entirely different outcome, which is instantly deliverable social status. The moment a customer joins, carries the branded bag, or mentions their membership, they immediately receive the full value of that exclusive identity. Because Equinox provides this immediate, easy-to-measure feeling of elite belonging, they can comfortably charge $200 to $500 per month53. You can charge massively higher prices by substituting a hard-to-deliver physical outcome with an instantly deliverable identity.

III. Investment Principles

Technology deflates functional goods and cannot deflate positional goods. Freight, bandwidth, and tokens all fell toward marginal cost once capacity caught demand, and the same fate awaits every functional layer of the AI stack. Positional goods are exempt because their scarcity is constitutive, meaning they are valuable precisely because others lack them and abundance would destroy the product. Deflation cannot reach a good whose value is exclusion.

Human provenance is the commitment device that makes scarcity credible. Human involvement proves to a buyer that a product is genuinely scarce. Studies by Imas show people will pay roughly twice as much when they know a product is genuinely exclusive54. His research also shows human-made work increases in value by 44 percent when made exclusive, whereas AI-generated work gains only 21 percent55. Anything made by a machine feels infinitely reproducible, while a human's limited time and judgment concretely prove that the supply is restricted

The willingness-to-pay test separates the two regimes. We can determine a product's category by testing what customers will actually pay. If buyers refuse to pay once they learn a machine could do the job faster and better, you are selling a functional good that ultimately belongs to automation. If customers still pay your price specifically because a human did the work—even if a machine is technically superior—the human element is your core value. The classic example is the mechanical watch56. When highly accurate quartz watches won on pure function, mechanical watches leaned entirely into craftsmanship, making the high-end market more profitable than ever.

The expanding price gap. This test always comes down to price, because the premium people will pay for human involvement has limits. A customer might pay a five-dollar premium for a human barista over a vending machine, but a fifty-dollar premium is heavily debated, and at five hundred dollars, the purchase is no longer about coffee. Automation constantly widens this gap because machine costs drop toward zero while human labor stays anchored to expensive wages. This widening gap pushes out buyers who only want to pay a small premium for human work. Durable, human-driven businesses only survive by targeting customers who are entirely price-insensitive to the human element and whose incomes grow faster than the automation gap.

IV. What We Own

Both successful ends of this strategic barbell share the same basic structure: the firm makes it easy for customers to stay while delivering a clear, upstream result.

At the technology extreme, a firm does this through functional advantages. It builds physical or organizational scarcity—manufacturing bottlenecks, tightly integrated ecosystems, or proprietary data that would take years for a buyer to replicate. Its product gets the customer to their goal with no extra steps, which lets the firm price against the value of that ultimate outcome rather than a standard tooling budget. It also has to sit in a place that larger adjacent tech giants cannot easily turn into a free commodity.

At the relational extreme, the firm does it through social belonging. Here, quality—which demands scarcity---is the product itself, created through memberships, limited allocation, or strict admission. This model depends on a credible human commitment that supply really cannot scale without diluting the quality of the good or service.

Everything in the middle—functional goods without real chokepoints and premium goods without real exclusion—is where automation-driven deflation and easy alternatives eventually compress margins. The discipline of the mandate is to concentrate capital at the two ends and to avoid owning what lives in the middle.

While optimistic, we think that the transition into a relational economy carries profound structural and longevity concerns. If only a small, wealthy segment can consistently afford these high-touch artisanal offerings, the vast majority of displaced workers may be forced to compete to provide them, risking a massive supply glut that creates a "feudalistic economy of relational labor". Furthermore, this hyper-competitive environment risks trapping both providers and consumers in a relentless positional arms race, a dynamic captured by the Red Queen Hypothesis. Providers will be forced to constantly adapt and run on an emotional and technological treadmill just to maintain their current market standing, without necessarily seeing any increase in net profits. Or worse, deep-pocketed tech platforms where these goods and services are advertised could naturally drive consolidation along the value chain, effectively counteracting those distributed gains. Ultimately, despite living in an era of unprecedented technological abundance and consuming bespoke relational goods, the hedonic treadmill suggests that modern humans may not actually feel any happier or more secure than their ancestors.

Precisely because these concerns are about the world we will all one day inhabit, they call us to take responsibility for helping steer the forces shaping humanity’s future as investors and operators. This is why we are committed to: (1) distinguishing real progress from its illusions, (2) evaluating and benchmarking that progress against the broader macro environment, and (3) directing capital and operations so that uniquely valuable human experiences can flourish.