Bubbles, Productive & Unproductive; Builders; & Bots: Why the AI Boom Isn’t One Story, But Rather the Vector Driving the AI Economy Is at Least 12-Dimensional

Six dimensions are entrepreneurial-technological-industrial aspects of bubble dynamics that are at least somewhat familiar from history; six dimensions are wild and new. The AI boom isn’t a single...

Six dimensions are entrepreneurial-technological-industrial aspects of bubble dynamics that are at least somewhat familiar from history; six dimensions are wild and new. The AI boom isn’t a single narrative; it’s a tangle of grifters, overbuilders, positive externalities, and coordination plays colliding with platform power and techno‑millenarianism as natural‑language access to data becomes a general‑purpose upgrade like literacy, but with sustainable business models that look mo re like commoditized plumbing…

---

The AI surge looks to me half like a familiar “productive bubble” and half like something much more complicated and new and strange. The “productive bubble
terrain of grifters, wasteful overbuilding, socially valuable but privately unprofitable infrastructure construction, coordination cycles, a few rock‑solid business models, and financial-crisis risk is at least somewhat familiar. But then we also have

• Platform near‑monopolists investing defensively at staggering scale;

• Millenarian enthusiasts with their religious-cult agendas;

• Natural‑language interfaces promise massive user surplus while commoditizing producers, as modes of human interaction with the infosphere are transformed utterly;

• These transformations do not just produce new technologies of nature-manipulation and human coöperation, they also rewire the brain and restructure human thought in unpredictable ways;

• Newer and stronger forms of attention extraction looming as the default monetization path.

• The downstream consequences of what will be a revolution in the modes of human collective cognition

• most durable value likely sits in small, task‑specific models tied to trusted data, and in moats built on workflow, reliability, and proprietary information. Even if many investors lose money, the infrastructure and capabilities will persist.

As an optimist, I see the likely equilibrium is user surplus rising fast—cheap, ubiquitous natural‑language access to data—while margins migrate to trusted data, integration, and uptime rather than model scarcity. I see policy choices around competition, energy, and data governance determining whether we get a broad productivity growth acceleration, or another round of attention enclosure.

But the future is one I cannot see.

The best cut at trying to set this out that I have seen this fall comes today, November 7, 2025, from Bill Janeway:

Bill Janeway: In Search of the AI Bubble’s Economic Fundamentals <https://www.project-syndicate.org/onpoint/will-ai-bubble-burst-trigger-financial-crisis-by-william-h-janeway-2025-11>: ‘A surge of investment in data centers and in the vast energy infrastructure… rising investment volumes fuel soaring valuations… new multibillion-dollar AI infrastructure projects. At the same time, a growing body of reports indicates that AI’s business applications are delivering disappointing returns, indicating that the hype may be running well ahead of reality…..

The history of modern capitalism has been defined by a succession of… “productive bubbles”… mobiliz[ing] vast quantities of capital to fund potentially transformational technologies whose returns could not be known in advance…. The companies that built the foundational infrastructure went bust. Speculative funding had enabled them to build years before trial-and-error experimentation yielded economically productive applications. Yet no one tore up the railroad tracks…. The infrastructure remained… to support the… “new economy’… after… delay and… with new players….

Brian Cantwell Smith…. “It’s not good that [ChatGPT] says things that are wrong, but what is really, irremediably bad is that it has no idea that there is a world about which it is mistaken.”… In business settings, tolerance for error is already low and approaches zero when the stakes are high…. What is the value-creating potential of LLMs? Their insatiable appetite for computing power and electricity, together with their dependence on costly oversight and error correction, makes profitability uncertain….

There are two distinct alternatives…. One lies in developing small language models—systems trained on carefully curated datasets for specific, well-defined tasks…. The other… is the consumer market, where AI providers compete for attention and advertising revenue… [and] value is often measured in entertainment and engagement, [so] anything goes…. given that Google’s and Apple’s browsers are free and already integrate AI assistants, it is unclear whether OpenAI can sustain a viable subscription or pay-per-token revenue model that justifies its massive investments…. [In] the Gartner Hype Cycle… a “trough of disillusionment” precedes the “plateau of productivity”…

I think that this is very good as it goes. But I think that it is greatly oversimplified. I see at least twelve different balls being juggled in the air here, only six of which are found in typical “productive bubbles”. And thus the situation is so complex that I find myself largely at sea.

In a “standard” “productive bubble”, there are, typically, five aspects to worry about: grifters, wasteful overbuilders, privately-unprofitable overbuilding made societally useful via positive externlities, coördinated implementation cycles à la Andrei Shleifer, and productivity increases driving the emergence of new rock-solid business models:

Start with the grifter con artists looking for suckers, many of them moving over from crypto. The persistent and recurring ecologies of opportunists who thrive wherever information is asymmetric and accountability is weak: from penny‑stock promoters and crypto “influencers” touting pump‑and‑dump schemes to “miracle cure” supplement sellers and political-consultant astroturfers. The playbook is depressingly consistent—manufacture urgency, evoke insider access, and weaponize social proof. Historically, waves of fraud have accompanied technological or financial shifts: the bucket shops of the Gilded Age, boiler‑room cold calls in the 1980s, mortgage‑origination kickbacks pre‑2008, and token issuances during the crypto boom. Today’s digital infrastructure accelerates the cycle. Oversight lags innovation and distribution platforms privatize attention without commensurate liability. Thus grifters proliferates. And that is, to a substantial degree, where we are.

There is also, in a typical “productive bubble”, wasteful overbuilding by enthusiast true believers. It springs from a potent mix of optimism bias, status competition, and misread economies of scale. Falling unit costs and modularity make it tempting to buy the “next tier up”. The result is high upfront capital outlays, underutilized capacity, and deferred maintenance burdens, with negative spillovers when best‑practice guides generalize niche enthusiast builds into mainstream advice. A better norm is staged investment: instrument first, measure peak and variability, then add capacity only where bottlenecks truly appear. But in many cases those arguing for such are overwhelmed by those who insist that they have to be first, and that a fast better-informed second bite at the apple will be too late.

This category overlaps with a third: overbuilding justified for the sociery as a whole by positive externalities, but unprofitable for investors. To investors and entreprenuers it looks like and is the second: wasteful and money-losing overbuilding. From the standpoint of society as a whole, however, it is socially justified when spillover benefits firms do not capture in their profits boost utility. Think of fiber-to-the-home rollouts that boost firm productivity and enable new services across a city, or clean-energy transmission that lowers wholesale prices and reduces emissions across regions: both expand the opportunity set for everyone, but the builder earns only the tariff or subscription revenue, not the broader productivity gains. Historically, the U.S. interstate highway system and mid‑century electrification overbuilt capacity by commercial standards, yet they catalyzed growth that no single investor could monetize. In digital markets, cloud capacity and AI compute often look like overinvestment until downstream innovations (e.g., new applications, research breakthroughs, complementary business models) arrive; the social return materializes in learning-by-doing, network effects, and option value that capital markets discount. This wedge between private and social returns—amplified by financing frictions, coordination failures, and the tendency to undervalue long-run spillovers—explains why public policy has often stepped in with subsidies, guarantees, or regulation to align incentives, and why what looks like “too much” from an investor spreadsheet may be exactly enough for society’s dynamic increasing-returns path.

Fourth comes the societally-beneficial and privately-profitable value of signals—even signals of ludicrous overenthusiasm—in coördinating the undertaking of mutually-reinforcing investment projects: “big pushes” and “implementation cycles'‘ à la Robert Allen and Andrei Shleifer. Coördination risk is a big risk. Bubble enthusiasm mitigates it. Complementary projects move together: upstream capacity, downstream adoption, and enabling infrastructure arrive in tandem, generating spillovers that no single firm could justify alone. Think of electrification waves where appliance makers, utilities, and factory retrofits reinforced one another; postwar reconstruction programs in Western Europe that bundled capital inflows with institutional reform and trade liberalization; or the U.S. semiconductor–EV–battery push, where subsidies, procurement, and expectations coordinate fabs, supply chains, and charging networks. Even bubbles with “ludicrous” sentiment leave productive residue. In Robert Allen’s and Andrei Shleifer’s terms, “implementation cycles” arise in equilibrium and government commmitment-triggered “big pushes” are optimal whenever confidence cues and commitment devices align timing, reduce threshold fears, and let complementarities be harvested—turning scattered investment into a self-amplifying development path rather than a set of stranded bets.

And fifth and last, of course, there are the productivity increase enabling rock-solid business models that are profitable without either coördination with other producers, or finding a greater fool to unload the position onto. They generate cash flows from true value creation. Think of containerization and modern logistics platforms that cut frictions and turnaround times, large‑scale cloud infrastructure that amortizes compute and storage over millions of users, or enterprise resource planning systems that standardize workflows and data—each reduces unit costs and raises output per worker without relying on speculative exit dynamics. Today, digital payments that shrink transaction costs, AI‑assisted coding that compresses development cycles, and industrial robotics that improve precision and throughput may become cases of such. But at the moment those of this fifth category that are well-proven are limited to two kinds:

• One, there is what Janeway calls “small language models—systems trained on carefully curated datasets for specific, well-defined tasks”. Natural-language interfaces to trusted and scrubbed structured and unstructured databases can be very valuable. Moreover, they do not require neural networks of absurd size. You do not need that big an LLM by our standards to achieve good-enough natural-language fluency. And if you have it pointed to and governed by a scrubbed-and-trusted database, you really do not want it to be complicated enough to start trying to think outside its box. Plus these ones that will soon be runnable on device. The poster children for this use case are today’s programming copilots.

• Second, there is what I think of as the sphere of articulate entertainment-creation and entertainment-management tools, and what I think of as articulate software pets of various kinds. These uses will spur all kinds of industries—but the actual foundation-model providers will be in the business of selling a commodity product as they compete against each other, and megafortunes will not be made here.

All of this is more-or-less standard for a “productive bubble”. And in addition there are the standard risks of debt- and vendor-finance, overleverage, financial crisis, and recession.

It will be interesting to watch how this particular repertory theatrical run of this now-old entrepreneurial-technological capitalism story evolves.

---

This time, however. Things are different. This episode brings with it a bunch of additional considerations, each of which is a separate ball being juggled through the air as well:

There are: millennarian religious enthusiasts forseeing the Rapture of the Nerds—either a good or a bad rapture, and deploying money, other resources, and programmer time to, in various ways, make themselves worthy. These sects and cults are richer than they have any right to be, and are spending more money and using up more resources than they have any right to do.

Imagine a technologically driven end-times in which salvation or damnation arrives via AGI or brain–computer interfaces rather than angels and trumpets. In the optimistic telling, a benevolent superintelligence cures disease, abolishes scarcity through near‑zero marginal‑cost production, and uploads minds into digital paradises—Kurzweil’s singularity meets Olaf Stapledon. In the pessimistic mirror, the same acceleration yields a paperclip apocalypse, a panopticon of algorithmic surveillance, or a caste society where the enhanced rule and the unaugmented serve. If there are any humans who qualify as enhanced enough to be among the ruling caste. Or, indeed, if the ruling caste sees humans as worth their concern at all—or as vermin to be simply tolerated until they become too annoying.

These secular eschatologies borrow heavily from older millenarian forms—Joachimite timetables, Great Awakenings’ purifying crises—transposed onto silicon: prophets become founders, relics become GPUs, and sanctification becomes alignment. Contemporary examples abound: effective altruists debating “x‑risk” as a form o f theodicy; venture manifestos promising “fully automated luxury” futures; policy briefs warning of irreversible lock‑in once systems surpass human oversight. What’s striking is not the novelty but the continuity: a familiar human hunger for meaning and redemption now articulated in engineering diagrams, with the same bifurcation between utopia and catastrophe, and the same politics of authority over who interprets the signs. And these secular eschatologies do and will have impact. Consider that at the start of the 1100s power- and military-political relationships in the triangle between Constantinople, Aswan, and Baghdad were upset and transformed by the eruption of the First Crusade, most of the participants in which had no knowledge of and little concern with the arrangements of power, control, and wealth they were upsetting.

There are: The platform near‑monopolists who see their current extraordinary rent streams at risk from a Christensen‑style disruption. They already habitually mobilize cash, regulatory muscle, and ecosystem control to try to smother the “Next New Big Thing.” Microsoft’s strategic throttling of Netscape’s distribution and browser oxygen in the 1990s is the canonical case, but the playbook has since scaled: Apple’s App Store rules and default placements shape which innovations reach users; Google’s leverage over search placement and Android preinstalls tilts discovery and adoption; Meta’s rapid cloning and bundling of emergent social features (Stories, Reels) into its vast network effects blunts upstarts; and Amazon’s marketplace self‑preferencing and data advantages deter third‑party category challengers. The mechanism is straightforward: combine control over chokepoints (defaults, distribution, payments), cross‑subsidize with monopoly cash flows, and wage long campaigns of acquisition, emulation, and exclusion while lobbying to frame these defenses as consumer benefits. But now they think they face a challenge that requires that they invest in GPT LLM capability at astonishing scale, out of the belief that to have an inferior model is to become the prey of hungry upstarts—or of each other. This makes it more unlikely that anyone else is going to make truly significant money off of GPT LLM MAMLMs. But it does make the build-out and the boom much larger. And it makes the potential value in terms of model-utilizer surplus much greater.

Do not underestimate these likely benefits to model-utilizers. Historically, general‑purpose communication breakthroughs—most obviously human speech and then literacy—reshaped production, coordination, and discovery by lowering the cost of translating intent into action. Natural‑language access to data is a similar general‑purpose capability.

Natural‑language interfaces to structured and unstructured databases promise extraordinary gains in user surplus because they collapse search, query formulation, and interpretation into a single conversational step. These interactions unlock latent value as users no longer need SQL fluency, bespoke dashboards, or weeks of analyst time to interrogate their information environment. Yet because such interfaces will rapidly commoditize—open models, standardized orchestration layers, and competitive SaaS packaging will keep prices low—much of the social surplus will accrue to users rather than producers. As with web search, email, and spreadsheets, the diffusion of a general tool that democratizes capability tends to squeeze margins, even while it transforms productivity across healthcare, finance, logistics, research, and public administration.

The strategic imperative for firms, then, is not to bet on scarcity rents from the interface itself, but to build moats around proprietary data, workflow integration, reliability, and trust—where durable value can still be captured while the broad prosperity gains diffuse through the commons.

Moreover, there are also: The unfortunate consequences of this game-changing. Natural‑language interfaces are genuinely transformative, but they risk becoming exquisitely tuned extraction machines for attention rather than tools for human flourishing. The interface’s conversational ease lowers friction for persuasion and nudging, letting recommender systems and engagement stacks steer users toward outrage, compulsion, and commercial conversion—now with the authority and intimacy of a “helpful assistant.” Think ad‑tech married to a tutor’s cadence: “personalized” prompts that upsell, polarize, or path‑depend your choices; chat flows that obscure whether you are being informed, entertained, or profiled; and “guardrails” calibrated less to civic well‑being than to dwell time.

We have lived versions of this before at lower intensity—newsfeeds optimized for clicks, autoplay loops, casino‑grade mobile games. LLMs , however, raise the ceiling by mimicking empathy, adapting in real time, and harvesting richer behavioral telemetry. If firms monetize attention and data, they will deploy natural‑language front ends as stalking‑horses for a new round of enclosure of the digital commons: proprietary walled gardens, opaque ranking, and pay‑to‑play placement dressed up as assistance. It may become a clickbait eyeball-glueing attention-harvesting world. We will just happen to live in it, and not to our benefit.

And I have not even touched on the downstream consequences of what will be a revolution in the modes of human cognitive interaction with the infosphere, and the currently unknowable downstream consequences of that for the shape of human society.

---

I look at all this, and I have a consciousness of being overwhelmed, and not even knowing where I should start to think. Which is the most important analytical thread to pull first, and how do I start to pull it?

If reading this gets you Value Above Replacement, then become a free subscriber to this newsletter. And forward it! And if your VAR from this newsletter is in the three digits or more each year, please become a paid subscriber! I am trying to make you readers—and myself—smarter. Please tell me if I succeed, or how I fail…

---

#bubbles-productive-unproductive-builders-bots
#mamlms
#bill-janeway
#subturing-bradbot