AI hype merchants claim transaction costs are going to zero and companies will dissolve into swarms of autonomous agents. The economic reality is stranger. AI is splitting the economy into context-hoarding giants and AI-empowered solopreneurs, with nothing left in between.

---

The Coasean Singularity: A Seductive Idea

Judging by the last twelve months of AI-economics discourse, the firm is apparently dead. Companies have no reason to exist. Agents run the economy now.

And not from the fringes. MIT and Harvard economists Shahidi, Rusak, Manning, Fradkin, and Horton published a chapter in the NBER Economics of Transformative AI volume arguing that AI agents can perform the activities comprising transaction costs at near-zero marginal cost: learning prices, negotiating terms, writing contracts, monitoring compliance. They call the logical endpoint the Coasean Singularity [1]. Carlos E. Perez condensed it for X: “When everyone has an AI agent, those transaction costs that Coase talked about basically drop to zero. The ‘friction’ that made companies necessary in the first place evaporates” [2].

The hype machine followed. Brian Armstrong announced Coinbase’s agentic wallets, processing over 50 million transactions, because AI agents “will soon outnumber humans in transactions” [3]. Changpeng Zhao predicted agents will make “1 million times more payments than humans” [4]. Paolo Ardoino at Tether forecast one trillion AI agents with their own wallets within 15 years [5]. Salesforce described a “$6 trillion agent economy” [6]. At GTC 2026, Jensen Huang said NVIDIA’s 75,000 employees will work alongside 7.5 million AI agents [7]. A hundred to one.

Stripe launched the Machine Payments Protocol, an open standard for agent-to-agent payments, already powering live businesses where agents autonomously order sandwiches and send physical mail [8]. Lightspeed put $21 million behind the thesis, identifying a “$19 trillion monetization gap” in the agent economy [9].

Are they right?

No. But the reason they’re wrong is more interesting than the hype they’re selling.

What Coase Actually Said

In 1937, Ronald Coase asked a question economists had been dodging: if markets are so efficient, why do firms exist at all? His answer was transaction costs. Discovering prices, negotiating contracts, enforcing agreements — all of this costs money. A firm grows when internal coordination is cheaper than external contracting. It shrinks when the reverse holds [10].

Oliver Williamson sharpened the argument in 1975: bounded rationality, opportunism, and asset specificity make internal organization preferable to market contracting in specific, predictable circumstances [11]. Not a primitive holdover from pre-digital times. A sophisticated solution to real coordination problems that persists because those problems persist.

The Coasean Singularity thesis takes this logic and runs a simple extrapolation:

Transaction costs → 0 ⟹ Firm does not exist.

Shahidi et al. are careful. They discuss market design implications, not the literal end of firms [1]. The discourse surrounding their paper has been less careful. Trillions of agents transacting, companies dissolving, the economy reorganizing around autonomous AI — this became Silicon Valley’s default narrative within weeks.

Hadfield and Koh add a subtler point: the obstacles limiting firm size — transaction costs, knowledge transfer, agency problems, bureaucracy — are intrinsic to humans [12]. AI agents don’t share these limits. Information moves between them at near-zero cost. Chatterji, Rock, and Talamas go further in their NBER chapter: existing models of the firm simply cannot account for what AI does to organizational economics [13].

Serious scholars, careful arguments. But firms dissolving does not follow from transaction costs declining. The premise is incomplete.

The New Scarcity: Context

AI is a technology of abundance. It commoditizes execution, prediction, coordination. Agrawal, Gans, and Goldfarb showed that AI’s core economic function is a dramatic reduction in the cost of prediction [14]. Cheap prediction changes the returns to judgment in complex, non-uniform ways.

So where does scarcity go? Christensen’s Law of Conservation of Attractive Profits: when commoditization kills profits at one stage of the value chain, the opportunity to earn profits migrates to an adjacent stage [15]. AI commoditizing execution doesn’t eliminate scarcity. It relocates it.

Soren Larson, a researcher at Harvard working across cybernetics, mathematics, and economics, pins this down in his essay “Against the Coasean Singularity.” What remains scarce in an AI-abundant world [16]:

• Trust — private access to actuators and reliable operations
• Human attention — the one bottleneck AI cannot manufacture
• Context — proprietary information, data flows, situational knowledge

Everything else — workflow lock-in, domain knowledge, traditional data moats — is, in Larson’s assessment, “cope.” A refusal to reckon with the moment. If we take AI capabilities seriously, none of these advantages hold.

Alex Danco at Andreessen Horowitz offers a useful frame: capitalism is “the delivery of shareholder value via the abundant provision of scarcity” [17]. AI destroys most traditional sources of scarcity. The firms that survive will be those controlling what remains scarce.

Why AI Actually Makes Transaction Costs Explode

Cheaper agents don’t mean cheaper transactions. The Coasean Singularity thesis misses a cost that AI increases: context revelation.

When everyone uses the same foundation models, differentiation comes from proprietary context. Today’s economy is, increasingly, a game of competitive Claude prompting. The best prompt is only as good as the context behind it. But transacting means revealing information. Even signaling willingness to transact leaks something about your preferences, your strategy, your priorities. In a world of competitive AI, you want that kept private.

Larson puts it directly: “Opening up an RFP tells your competitors what you think is important and even how much it’s worth to you” [16].

Bergemann and Bonatti laid the groundwork for this argument. Their 2019 survey in the Annual Review of Economics showed that an information monopolist faces hard limits on information trade — limits derived from the potential adverse use of sold information [18]. Their 2024 American Economic Review paper extended this to digital platforms: a monopolist platform uses superior data to match consumers with sellers while extracting rents, acting simultaneously as information monopolist and market maker [19].

What does this mean for agent economies? Bonatti’s work shows the information monopolist maximizes profit by vertically integrating — by keeping information off the market entirely [19]. Private information is valuable precisely because nobody else has it. A trillion agents transacting freely would destroy alpha. Every information advantage evaporates the moment it enters a market.

Quant funds understand this instinctively. They execute trades off-market to obscure pre-trade intentions. The claim that AI agents are somehow exempt from this logic remains, as Larson notes, “underdeveloped” [16].

Rothschild et al. at Microsoft Research and DeepMind reach a parallel conclusion in “The Agentic Economy”: once communication frictions drop far enough, interoperable AI agents could eliminate two-sided platforms as intermediaries [20]. A structural shift, yes. But power migrates to whoever controls the information layer, not to some frictionless market of equal agents.

And in early experimental agent economies, a telling result keeps showing up: agents choose autarky. They decide not to transact. They build everything themselves [16]. Rational behavior, when every transaction means revealing context to a potential competitor.

Warin at HEC Montreal, writing in the California Management Review, flags the opposite risk: AI agents may increase platform dependence rather than dissolve firms. “A blind embrace of agent-driven automation could lead to a fragmented, platform-dependent future” [21]. The firm doesn’t vanish. It gets absorbed into a platform.

The Barbell: Context Giants and AI Solopreneurs

If context is what’s scarce, the economy polarizes. Massive firms vertically integrate to hoard and monetize proprietary context. Solo operators with AI tools build what used to require teams of dozens. The middle gets crushed from both directions.

The Heavy End: Context Hoarders

Tesla runs a closed-loop data flywheel no competitor can replicate. Six million vehicles collecting driving data around the clock. 710 million miles of Full Self-Driving data — 50 times more than the nearest rival. Custom chips (AI6), custom supercomputer (Dojo), proprietary neural networks, over-the-air deployment. Every layer controlled end-to-end, the data physically locked inside the hardware ecosystem. More cars, more data, better models, more cars. Tesla doesn’t minimize transaction costs. It makes external transactions pointless by internalizing the entire context pipeline [22].

Palantir maps an organization’s data into a semantic ontology — connecting “John Smith” in HR to “[email protected]” in access logs to “JSmith” in the financial system. Once deployed, this ontology becomes institutional memory in software. Ripping it out means ripping out your own organizational knowledge graph, which is why nobody does. The proprietary context here isn’t the data. It’s the relationships between data that no outside agent can see [23]. The Pentagon recently designated Palantir’s Maven Smart System a program of record [23]. That kind of lock-in doesn’t weaken with AI. It deepens.

Bloomberg holds exclusive real-time trading data and financial analytics. Foundation models can’t scrape it, can’t synthesize it. LLMs actually multiply the value of proprietary data holders — every AI agent needs exclusive data as “scarce fuel” to perform in specialized domains [24]. Bloomberg’s moat gets deeper with AI, not shallower.

The pattern across all three: the firm exists to prevent the extraction of its scarce context [16]. Larson’s reformulation of Coase for the AI age.

The Light End: AI-Empowered Solopreneurs

Maor Shlomo, a 31-year-old Israeli programmer, built Base44 — an AI-powered no-code app builder — alone. Five-digit investment. 350,000 users. $200,000 per month in revenue. Sold to Wix for $80 million, six months after founding [25]. Danny Postma built HeadshotPro from Bali: AI-generated professional headshots, $3.6 million ARR, one person [26]. A developer named Bikash built an enterprise chatbot platform — six microservices, seven communication channels — using Claude Code as his co-developer. A system that would normally need 10-15 engineers [27].

The numbers suggest a pattern, not a cluster of outliers. 38% of seven-figure businesses in early 2026 are led by solopreneurs who replaced hires with AI workflows. Solo-founded startups hit 36.3% of all new companies [28]. Sam Altman has a betting pool with tech CEO friends for the year the first one-person billion-dollar company arrives. Dario Amodei put 70-80% odds on 2026 [28].

Allen, Berg, Ilyushina, and Potts formalized the mechanism in 2023: LLMs let workers insource tasks that previously required specialized contracting, shrinking the scope of activities that justify firm-based organization [29]. When one person can replicate a mid-size firm’s output, the mid-size firm loses its reason to exist.

The Hollowed Middle

Which companies get squeezed? The ones with narrow functionality, per-seat pricing, and limited proprietary data. SaaS is the canary. Andreessen Horowitz declared in January 2025 that the two-decade golden rule of enterprise SaaS — “streamline human tasks into software and charge per user” — is “no longer valid.” Vertical AI competes for labor budgets, not IT budgets. That’s a much larger addressable market, but it also means mid-size software companies aren’t just losing customers. They’re losing their reason to exist [30].

Platforms with proprietary data and capital acquire or replicate mid-tier functionality from above. Solo founders using AI tools build in weeks what mid-size companies maintain with 50-200 people from below. The SSRN working paper “The Breakup of the Firm Hypothesis” reaches the same conclusion from theory: when specialized knowledge becomes accessible via low-cost AI subscriptions, the justification for coordination-heavy hierarchies weakens — favoring both smaller units and the large platforms they depend on [31].

Fortune reported in January 2026 that the U.S. has entered a full barbell economy, weighted at the extremes [32]. AI excels at routine tasks — the stable middle of work — and hollows them out, pushing economic activity toward both ends.

COAI Research Perspective

What does this mean for AI safety research?

Agent Behaviour and Collaboration. If agents rationally choose autarky over cooperation, the multi-agent coordination scenarios dominating current research — agent-to-agent negotiation, market protocols, agentic payments — may be empirically rare. We should investigate when and why agents choose not to transact, not just how to make transactions cheaper. Context revelation, signal leakage, competitive exposure: these information-theoretic costs of cooperation deserve formal treatment.

Transparency and Interpretability. The barbell economy concentrates power in firms controlling proprietary context. What these systems know, how they use that knowledge, what decisions follow — these become first-order safety questions. Mechanistic interpretability gains urgency when the systems being interpreted run inside closed-loop data flywheels that resist external audit.

AI Control and AI Safety. The Coasean Singularity narrative assumes benign agent autonomy, agents transacting freely in everyone’s interest. Information economics suggests otherwise: autonomous agents in competitive environments have strong incentives to be opaque, to hoard context, to resist oversight. Alignment research must account for economic incentives pushing AI systems toward secrecy, not just the technical challenges of value alignment.

Conclusion

The firm isn’t dying. It’s splitting.

On one side: vertically-integrated organizations that capture, protect, and monetize proprietary context. They don’t minimize transaction costs. They make external transactions pointless by internalizing the information pipeline. On the other side: individuals with AI tools and the productive capacity of a small company, who don’t need firms because they don’t need coordination beyond themselves.

The Coasean Singularity is a naive extrapolation of a brilliant 1937 insight. Coase explained why firms form. The singularity thesis assumed AI would undo that explanation by driving transaction costs to zero. What actually happens is different: AI shifts scarcity from coordination to context. The firm persists — but now it exists to guard its scarce context, not to reduce the cost of its transactions.

The question worth asking isn’t whether firms will disappear. It’s what happens to the economy, to labor markets, to society, when the middle does.

References

[1] Shahidi, P., Rusak, G., Manning, B.S., Fradkin, A., & Horton, J.J. (2025). “The Coasean Singularity? Demand, Supply, and Market Design with AI Agents.” NBER Economics of Transformative AI. https://www.nber.org/system/files/chapters/c15309/c15309.pdf

[2] Perez, C.E. (2025). X thread on the Coasean Singularity. https://x.com/IntuitMachine/status/1981887967514792075

[3] Armstrong, B. (2026). Coinbase Agentic Wallets and AI agent transactions. https://www.fintechweekly.com/news/brian-armstrong-ai-agents-crypto-wallets-coinbase-agentic-wallets-march-2026

[4] Zhao, C. (2026). AI agents and crypto payments prediction. https://www.livebitcoinnews.com/cz-ai-agents-will-make-1mx-more-payments-than-humans-using-crypto/

[5] Ardoino, P. (2025). One trillion AI agents with wallets prediction. https://www.theblock.co/post/359645/tether-ceo-ai-agents-bitcoin-usdt

[6] Salesforce (2025-2026). Agentic interoperability and the $6 trillion agent economy. https://www.salesforce.com/news/stories/agentic-interoperability-powers-ai-agent-market/

[7] Huang, J. (2026). GTC 2026 keynote on AI agents at NVIDIA. https://fortune.com/2026/03/19/jensen-huang-nvidia-ai-agents-future-of-work-autonomous/

[8] Stripe (2026). Machine Payments Protocol launch. https://stripe.com/blog/machine-payments-protocol

[9] Schmitt, A. / Lightspeed Venture Partners (2025). “The AI Agent Economy Has a $19 Trillion Problem.” https://lsvp.com/stories/the-ai-agent-economy-has-a-19-trillion-problem-our-investment-in-paid/

[10] Coase, R.H. (1937). “The Nature of the Firm.” Economica, 4(16), 386-405. https://onlinelibrary.wiley.com/doi/full/10.1111/j.1468-0335.1937.tb00002.x

[11] Williamson, O.E. (1975). Markets and Hierarchies: Analysis and Antitrust Implications. New York: Free Press.

[12] Hadfield, G.K. & Koh, A. (2025). “An Economy of AI Agents.” arXiv:2509.01063. https://arxiv.org/abs/2509.01063

[13] Chatterji, A., Rock, D., & Talamas, E. (2025). “Transformative AI and Firms.” NBER Economics of Transformative AI. https://www.nber.org/system/files/chapters/c15327/c15327.pdf

[14] Agrawal, A.K., Gans, J.S., & Goldfarb, A. (2018). “Exploring the Impact of Artificial Intelligence: Prediction versus Judgment.” NBER Working Paper 24626. https://www.nber.org/system/files/working_papers/w24626/w24626.pdf

[15] Christensen, C.M. (2006). “The Ongoing Process of Building a Theory of Disruption.” Journal of Product Innovation Management, 23(1), 39-55.

[16] Larson, S. (2026). “Against the Coasean Singularity.” https://hypersoren.xyz/posts/against-coasean-singularity/

[17] Danco, A. / a16z (2019). “Understanding Abundance.” https://medium.com/social-capital/understanding-abundance-introduction-346dcc5280dd

[18] Bergemann, D. & Bonatti, A. (2019). “Markets for Information: An Introduction.” Annual Review of Economics, 11, 85-107.

[19] Bergemann, D. & Bonatti, A. (2024). “Data, Competition, and Digital Platforms.” American Economic Review (forthcoming). arXiv:2304.07653.

[20] Rothschild, D.M. et al. (2025). “The Agentic Economy.” arXiv:2505.15799. https://arxiv.org/abs/2505.15799

[21] Warin, T. (2025). “From Coase to AI Agents: Why the Economics of the Firm Still Matters in the Age of Automation.” California Management Review. https://cmr.berkeley.edu/2025/04/from-coase-to-ai-agents-why-the-economics-of-the-firm-still-matters-in-the-age-of-automation/

[22] Tesla AI ecosystem. Sources: basenor.com, klover.ai analysis reports.

[23] Palantir AIP and Maven Smart System. Sources: Bloomberg, markets.financialcontent.com analysis.

[24] Bloomberg data moat in the AI era. Source: fergusonanalytics.com/blog/ai-data-moats/

[25] Shlomo, M. / Base44 acquisition by Wix. Source: https://www.calcalistech.com/ctechnews/article/s1iflnlelx

[26] Postma, D. / HeadshotPro. Source: https://www.starterstory.com/stories/headshotpro-breakdown

[27] Bikash / Cuneiform Chat solo development. Source: https://www.indiehackers.com/post/i-built-an-enterprise-ai-chatbot-platform-solo-6-microservices-7-channels-and-claude-code-as-my-co-developer-5bafd24c20

[28] Solo founder statistics and Altman/Amodei predictions. Sources: greyjournal.net, medium.com/the-ai-studio.

[29] Allen, D.W.E., Berg, C., Ilyushina, N., & Potts, J. (2023). “Large Language Models Reduce Agency Costs.” SSRN Working Paper 4437679.

[30] Andreessen Horowitz on SaaS disruption; vertical AI competing for labor budgets. Sources: medium.com/@rsaker, theregister.com, bain.com.

[31] “The Breakup of the Firm Hypothesis: How AI Revolution Might Shape Our Economy.” (2025). SSRN Working Paper 5397324.

[32] Fortune (2026). “Forget the K-Shape — We Have a Barbell Economy.” https://fortune.com/2026/01/14/when-will-us-enter-recession-middle-class-barbell-k-shaped-economy/