On a recent episode of the Uncanny Valley podcast, guest Brian Merchant laid out a historical framework he used to judge whether the current rush into artificial intelligence meets the classic signals of an economic bubble — and what that could mean for investors, workers, and the broader economy.
Hosts Michael Calore and Lauren Goode opened the conversation by noting how corporate forecasts and quarterly calls have shifted to heavy talk about spending plans. At the center of that talk was capital investment in compute and facilities. Lauren quipped, "CapEx?" and Michael replied with the plain form: "Capital expenditures." The exchange set the tone for a discussion about massive budgets, big hardware plays, and the question on many minds: is the current surge in AI spending sustainable?
The timing of the episode matters. Since a high-profile comment this past summer by Sam Altman that mentioned the possibility of an AI bubble, the topic has been debated in boardrooms and on trading floors. Bill Gates told CNBC that yes, we are in a bubble, and added that "Something profound," will come out of it. That tension — between the sense of excess and the hope for lasting innovation — was the focus of the hour-long talk.
Merchant, who writes Blood in the Machine and contributes reporting on technology, told the hosts he wanted to move beyond mood-based takes. He described being drawn to a 2019 book by Brent Goldfarb and David Kirsch, Bubbles and Crashes: The Boom and Bust of Technological Innovation, which offers a method for comparing investment frenzies across eras. Goldfarb and Kirsch studied multiple historical booms and identified recurring patterns that can help identify a bubble before the fall.
The authors isolate four signals that tend to accompany tech bubbles. The first is uncertainty about how the new technology will actually translate into profitable business models. Early demonstrations can be dazzling, yet the markets for those technologies and the revenue mechanics that will sustain firms often take decades to crystallize. Goldfarb and Kirsch then point to the number of "pure play" firms — companies whose entire business depends on a single technology or service. The greater the number of firms with a single, narrow exposure, the more fragile that sector becomes if customer demand fails to materialize. The third signal is the presence of many novice investors, especially retail participants who can pour money into new public offerings and secondary markets without deep expertise. The fourth is a widespread, coordinated belief that the technology will be dominant; that narrative alignment among investors and corporate buyers can raise valuations far beyond what fundamentals support.
Merchant walked through historical examples to make those signals concrete. He pointed to the first decades of electrification, when electric light and power appeared miraculous yet it was unclear which commercial uses would drive mass adoption. Would households buy bulbs? Would cities invest in tower lighting? The right commercial model took decades to form. Radio, too, captivated the public and investors before a stable business model emerged. Aviation, accelerated by Charles Lindbergh’s transatlantic flight, became a magnet for speculative capital once a vivid demonstration suggested the technology could open big new markets. Those episodes show how a powerful demo can align belief among investors before the economic plumbing catches up.
That dynamic is visible in the current AI rush, Merchant argued. The arrival of widely used chatbots and public-facing models created a rapid, visible proof that people might want to use these systems. But demos and viral adoption do not automatically translate into durable revenue streams. The MIT study that has circulated in recent months — which reports that about 95 percent of firms that have deployed generative AI have yet to see meaningful returns — emerged during the same window when companies were announcing gigantic infrastructure plans. That mismatch between investment and payoff is exactly the kind of signal Goldfarb and Kirsch would flag.
The "pure play" signal in AI is a little different than in past cycles, Merchant said. In previous booms, a wave of new public companies provided retail investors with direct ways to place bets on a single technology. In the current cycle, capital is flowing into a much narrower set of public names and into private deals that concentrate exposure. Nvidia, for example, has become the center of many AI bets. Once primarily known for graphics chips for gamers, the firm now supplies the processing hardware that undergirds many AI systems. Merchant used the old metaphor about selling shovels during a gold rush: chips and data-center capacity are the shovel equivalent. That makes Nvidia an obvious target for investment, because demand for compute may continue even if some AI applications disappoint.
That concentration carries risks. Merchant noted that Nvidia now represents a very large slice of market capitalization in the public equity markets — he put it around 8 percent. A severe decline in Nvidia's value would ripple through portfolios and funds that hold that exposure, including retirement plans and institutional investors who may not realize how heavily they are invested in the hardware side of the AI story. The public company presence is not the only route for retail and institutional exposure. Firms like CoreWeave and a growing number of data-center and real-estate plays have carved out businesses that are effectively pure-play services for AI compute. Some private funding rounds have linked chip suppliers, cloud renters, and model operators in complex financial arrangements. Merchant pointed to deals that tie chipmakers and AI firms together in nontraditional ways; he described an arrangement in which OpenAI and chip partners have produced equity cross-holdings that change how risk would propagate if demand falls short.
The novice-investor factor in the Goldfarb and Kirsch framework is familiar from past booms: when easy investment channels give everyday buyers access to hype, prices can detach from fundamentals. Merchant said the current retail base looks a lot like the last time the market chased a story en masse. Apps that make trading frictionless, broad publicity for a small group of winners, and an appetite for anything labeled "AI" have recreated the conditions under which nonexpert buyers can flood companies with capital. Goldfarb and Kirsch argue that retail participation matters because it extends the reach of the speculative wave beyond a small circle of professional allocators.
Merchant and the hosts spent time on the fourth factor, the alignment of belief. That alignment is what creates the common story investors tell themselves about a technology's future. In the AI case, the story has been wide: AI will automate jobs at scale, improve medical research, help combat climate change, and in some telling scenarios deliver artificial general intelligence. Michael spelled out the broad claims that circulate in the investor class: "AI is going to do everything. It's going to automate jobs, it's going to cure cancer, it's going to babysit your kids, it's going to fight climate change." Those sweeping promises are a potent sales pitch for capital allocators who fear missing out on the next dominant platform.
The Goldfarb and Kirsch schema includes a numeric scale to rate booms. Merchant reported that the two scholars gave AI the highest possible mark on their scale: an eight. "A big fat eight, a buyer beware," Michael quoted when reflecting the authors' assessment. That top score signals that all four structural pieces are present and operating at high intensity.
The conversation steered toward what might happen if the bubble popped. Merchant emphasized that bubbles do not mean the underlying technology vanishes. The internet did not go away after the dot-com collapse; parts of the telecom build-out persisted and enabled later value creation. He said AI will likely leave behind useful artifacts: models that serve certain tasks, improved tools for content production, automation layers inside enterprise workflows. But the scale of the current capital concentration could produce economic pain that far exceeds past corrections. The dot-com wobble shuttered many firms and inflicted losses; a much larger contraction around a tech as broadly funded as AI could cascade into wider financial stress.
That possibility raises questions about exposure and systemic risk. Merchant and the hosts discussed the ways institutional investors, public pensions, and ordinary savers can be more linked to AI-centric investment than they realize. Private equity placements, corporate ventures, and tangled cross-holdings can make apparent diversification illusory. He warned that bundles of real estate tied to data centers, or mutual funds with outsized stakes in the hardware winners, could transmit shocks to otherwise diversified portfolios.
Lauren asked whether core gains from AI could prove durable even if speculative valuations come down. She described a conversation with AMD CEO Lisa Su and recalled agreeing that AI might yield concrete benefits in fields like medicine; she noted their shared experience of having relatives in the ICU and how that made potential benefits feel real. Brian agreed that practical value is likely to persist. He also cautioned that some valuable changes could exacerbate inequalities: winners in the AI boom may capture outsized gains while many workers face wage pressure as automation reduces demand for certain kinds of labor.
The hosts returned to the question of how AI's long-term profile might compare to past waves of innovation. Lauren asked whether generative AI might be more of a content-focused development — akin to broadcast radio or social networks — rather than a foundational layer like electricity or fiber optics. Merchant said that resonates. Many generative systems today produce content that is useful for certain tasks and cost-efficient for companies trying to reduce labor expenses, but quality often trails human output in ways that matter for complex work. That suggests a split: AI will likely remain an active part of the software stack in many consumer and professional settings, but it may not become a universal infrastructure in the same mold as wired broadband or electrical supply.
Michael pushed on the hardware angle. He noted that telecom infrastructure built during the dot-com era — fiber and switching — offered long-lived value after the bubble cooled. Chips present a different case: they improve rapidly and can become obsolete as the next generation arrives. Merchant agreed, adding that the transience of hardware advances creates a distinct risk profile for chip-heavy investments. Demand for particular chip architectures may spike and then diminish as models and deployment patterns change.
The group also touched on the role of government. Merchant noted recent public-sector moves that alter the investment landscape. He said the U.S. government has taken stakes in domestic chip production in ways that would have been rare a decade ago, such as a significant ownership position in a major chipmaker. That kind of state involvement changes the options available if markets deteriorate: governments could intervene, buy stakes, or provide stability to firms seen as strategically important. Merchant described that scenario as another new factor in how a potential bust could play out.
Midway through the episode, the tone shifted to the immediate question of whether the current mania would continue to inflate or begin a major correction. Goldfarb and Kirsch's high score suggests substantial bubble risk, Merchant said, and that worries him because the scale of capital flowing into AI seems to exceed near-term revenue prospects. Yet he stopped short of asserting a particular timeline for a collapse. Bubbles can deflate quickly or linger while valuations shift; the presence of free cash and institutional appetites makes precise timing hard to forecast.
The hosts and guest traced how the market has concentrated winners. Nvidia's valuation has ballooned rapidly; it crossed multiple trillion-dollar thresholds in a compressed time window. Merchant reminded listeners that it was not long ago that a single $1 trillion company was rare. Rapid capital concentration and accelerating accumulation in a handful of companies can magnify market effects. When a few firms capture vast investor attention, moves by those stocks can sway index flows and pension fund returns.
Merchant returned to the book that shaped his analysis. Goldfarb and Kirsch examined many historical examples with an eye to what differentiates sustainable booms from bubbles. Their framework is diagnostic rather than predictive in a narrow timing sense: it shows whether structural conditions match patterns that have led to crashes in the past. Merchant stressed that the framework flagged AI at the highest risk setting, but it does not produce an exact forecast of the scale of damage if markets correct.
That observed risk produces two countervailing possibilities. One path sees a painful contraction that exposes losses across private and public balance sheets, perhaps forcing portfolio managers to mark down assets tied to AI capacity. Another path has a milder correction in which much of the compute capacity and model development continues to deliver incremental benefits and a smaller group of firms captures the long-term upside. Merchant said both outcomes are plausible. He warned listeners not to confuse the existence of useful technology with the absence of speculative excess.
Discussion turned to specific public statements by AI executives and founders. The hosts and Merchant reviewed Sam Altman's early quip about business planning at OpenAI: "We're going to build AGI and we're going to ask it." The line has been quoted widely as evidence of the uncertainty that has hung over some strategies. Merchant used the remark to illustrate how companies sometimes framed their roadmaps as open-ended bets on future breakthroughs rather than grounded product plans. That kind of rhetoric can be compelling to investors craving a big, transformative narrative, yet risky if profitability remains distant.
The conversation also revisited Bill Gates's CNBC comment that a bubble exists but something major would emerge from the turmoil. Merchant acknowledged that sentiment is credible; even after losses, innovations and infrastructure often remain. The question is scale and the distribution of gains and losses. In some scenarios, the economic distress from a sharp correction could be deep enough to leave lasting scars on employment, capital allocation, and innovation patterns.
The second half of the episode examined how AI is likely to be used in practice and where it may fall short of investor expectations. Merchant argued that generative systems are strongest as tools for producing or augmenting content. That means marketing, creative production, and certain kinds of routine technical output may become less costly and faster to produce. But in domains requiring deep understanding, human judgment, or long-term reasoning, current models still make errors and can be unreliable. The tension between promotional claims and operational reality is one reason many firms have yet to see returns on their AI investments.
Merchant referenced an MIT study that found 95 percent of firms deploying generative AI had not yet realized returns from those efforts. That statistic reverberated through the discussion as evidence that the delivery mechanisms for value are not yet in place at scale. Corporations may be deploying AI to experiment and to keep pace with competitors, which inflates investment even when profit generation lags.
Two broad use cases drew attention. One is consumer-facing assistants: chatbots and conversational agents that will compete for user attention in ways similar to social platforms. The second is workplace automation: tools that augment or replace parts of professional workflows, such as drafting copy, synthesizing research, or streamlining routine transactions. Merchant said both tracks are likely to persist, yet each will affect labor and markets differently. Automation of creative and knowledge tasks could reduce demand for certain roles, exerting downward pressure on wages in affected sectors. At the same time, new services and efficiencies may arise that generate different forms of employment and revenue.
The hosts framed a practical question about what endures. Radio and broadcast content became a staple in American life after initial speculation. The internet outlasted the dot-com crash because it found stable, essential uses. Merchant placed generative AI closer to the content-production axis: powerful and pervasive in some contexts, but not necessarily the foundational physical infrastructure that undergirds entire industries. He added that certain parts of the AI stack — data centers, specialized chips, and large-scale model training capabilities — may have more transient utility than the long-lived assets created in prior cycles.
Listeners heard concerns about how quickly chips can become outdated. Whereas fiber and telecom equipment provided lasting transport capacity, AI chips face rapid cycles of iteration and replacement. That dynamic can leave firms with stranded investments if demand patterns shift. Merchant said that risk complicates the narrative for investors seeking lasting returns from hardware bets.
State action and industrial policy entered the discussion as wildcards. Merchant noted that the U.S. government's willingness to underwrite or acquire stakes in semiconductor production alters historical dynamics. A state willing to own a chunk of a critical supplier changes the private-sector-only model that has governed many past bubble episodes. If prices drop and firms suffer, governments might step in for strategic reasons, affecting who bears losses and who gets protected.
After returning from a short break, the hosts shifted toward lighter segments of the show that mark trends and pet peeves. They launched a short round in which each named a thing they think is losing relevance and a thing they think deserves attention. The hosts framed the segment as a way to close a dense discussion with practical takeaways about behavior and technology.
Lauren picked meetings as her top "tired" item. She argued that productive gatherings fall into two useful categories: meetings with a clear agenda that move quickly through bullet points, and unconstrained brainstorms meant specifically to generate fresh ideas. Her complaint targeted the middle ground — sessions that claim structure but drift into meandering conversation. "The no bad ideas meeting," she said, works when it is explicit about its intent, she added, and filler meetings that neither set an agenda nor invite creative play should be retired.
For something to highlight as "wired," she recommended a simple social skill: "Read people's faces better." If your phone is not dominating your attention and you look at someone while they speak, you often understand tone and nuance more clearly. Lauren said that more presence yields better understanding and makes conversations more memorable.
Brian's choices cut against a recent trend he finds troubling. He said he is "tired" of AI companions: forms of personalized bots or avatars marketed as emotional confidants or virtual partners. He described these products as amplifying loneliness and feeding unhealthy patterns, particularly for users in vulnerable states. "No shade on anyone who is using an AI companion," he added, "but I think it's time to wind it down." His "wired" pick was a grass-roots reaction: groups and chapters that encourage less screen time and more in-person interaction. He described those efforts as a counterweight to the pattern of replacing human contact with digital substitutes.
Michael offered a more mundane pair. His "tired" item was wool base layers for outdoor activity, which he finds heavy and itchy in many situations. His "wired" pick was capilene base layers, a synthetic fabric he has tested and praised for light weight and adequate warmth. He framed the item as a practical swap for anyone planning outdoor exercise in variable weather.
The conversation wound down with a final exchange about whether the current investment pattern will end in a sharp correction. Merchant repeated his cautionary view: the structural signs Goldfarb and Kirsch identify are strongly present and merit attention from investors and policymakers. Yet he did not argue that AI will vanish; he said parts of the industry will persist and produce value. The real risk is the scale of concentrated capital and the potential economic stress a correction could inflict.
Before the episode concluded, the hosts and guest agreed that the next few quarters would be decisive for the market and for public perceptions. Merchant suggested that observers watch where actual revenues and profits start to accrue and how funds and public pensions are exposed to hardware and service bets tied to AI compute. He also mentioned that policy responses could change the damage profile if markets retreat.
The hour produced a mix of alarm and nuance. Merchant used a historical lens to show how similar patterns have played out in the past; the Goldfarb and Kirsch framework offered a structured way to see the overlap between past bubbles and the present AI boom. The conversation never lost sight of concrete details such as Nvidia's role, CoreWeave's business model, the MIT return-on-investment finding, and Sam Altman's memorable line: "We're going to build AGI and we're going to ask it."
That material was paired with moments of practical guidance and cultural observation during the show's closing features. The hosts urged listeners to pay attention to how companies convert hype into profits, to scrutiny over exposure in pooled investments, and to the social effects of automation. Merchant cautioned that no one should treat a vivid demo alone as a reliable predictor of business outcomes; he said that the mechanics of turning demonstrations into sustained revenue are often messy and take time.
The episode left a clear impression: the present moment contains both genuine technological breakthroughs and structural signs that have preceded painful market corrections in the past. Investors, corporate managers, and policymakers who want to make measured choices will need to weigh demo-driven mania against hard metrics on revenue, adoption, and durable cost savings. The historical patterns Goldfarb and Kirsch identified provide a lens for that assessment, and their highest score for AI should prompt more than casual attention from people who manage capital or steward public funds.