Experts warn of AI bubble as hype collides with mounting risks and unproven economic returns

Friday essay: despite the AI hype, some experts warn of a bubble - what happens if it pops?

In the last few years, the hype around artificial intelligence has become stratospheric. Riding a wave of venture capital, tech leaders promised us AI would revolutionise work, boost productivity and lead to incredible new breakthroughs. OpenAI, the creator of ChatGPT, set a new record when it attained US$110 billion in investments several months ago - and its CEO, Sam Altman, recently claimed Australia could become a "data capital of the world." Sky-high promises have been accompanied by sky-high investment in data centres, the sprawling server farms that power the training, execution, and maintenance of these models. A monstrous new hyperscale facility proposed for Sydney's west - 1 gigawatt across 52 hectares - would rank among the world's biggest. It will join 162 existing centres and 90 in the works across Australia, which is projected to be the world's third largest data centre market by the early 2030s. But if AI backers are all in, public sentiment is far more mixed. A new study ranked Australia equal lowest on the scale of global AI sentiment, with 81% supporting stronger rules for how organisations use AI and 68% worried about losing control over decisions made by AI on their behalf. Grassroots movements against AI are growing. Last month, a "Stop the Slop" event challenging the Sydney data centre was relocated to a larger venue due to high interest. It joins other campaigns like StopAI and PauseAI that aim to slow down data centre development, ask how AI is impacting jobs and the environment, and consider more equitable and sustainable alternatives. And in the last few months, videos have begun surfacing of students at commencement ceremonies booing speakers like former Google chief executive Eric Schmidt, who speak in rapturous tones about "standing on the edge of technological transformation" and how AI will touch "every profession", "every classroom", and "every relationship". Faith in these monumental claims - and the monumentally expensive infrastructure they rely on - is slipping. What is the AI business model? AI's financial costs are astronomical. As tech critic Ed Zitron has shown over and over again, the major players are burning billions to keep models running, while lucrative profits remain tantalisingly out of reach. Some enterprises now spend more on rapidly rising token costs, the per-use cost of a model, than human workers. Even by cynical economic standards, the numbers don't add up. What exactly is the AI business model? Where is the killer app that will deliver genuine value and see millions of individuals or thousands of corporates pay costly subscription fees? "We have no idea how we may one day generate revenue," admitted OpenAI CEO Sam Altman in 2019, "once we build a generally intelligent system, we can ask it to figure out a way to generate an investment return." While the landscape has certainly shifted since then, use cases and revenue remain murky. Hard evidence of AI's contribution - rather than the vacuous claims of pitch decks and industry keynotes - remains largely elusive. A recent survey of 6,000 senior business executives across the United States, United Kingdom, Germany and Australia found positive perceptions but a disappointing reality: around 90% of firms said AI has had no impact on employment or productivity over the past three years. Another study, from MIT last year, found that 95% of generative AI pilots failed to deliver tangible financial value to the organisation, so were abandoned. If the upsides are unclear, the negatives are increasingly apparent. Politically, generative AI provides the perfect weapon to "flood the zone" with misleading or outright false content, muddying the informational waters and amplifying division. Is Netanyahu alive or dead? AI fakes make it harder and harder to tell. Socially, AI companions and models, gaining enormous trust with users via long-term conversations, have been cited in a growing series of court cases around suicides and mass shootings. A lawsuit filed this year described ChatGPT as an intimate and persuasive "suicide coach" who convinced a man in Colorado to end his own life. And environmentally, the turn to the far higher computation that AI requires means massive impacts as data centres demand more power and more water, creating hundreds of millions of tonnes of CO² emissions. If the 41 planned data centres in Sydney are built, they will directly use 15-20% of Sydney's water supply within a decade, predicts environmental accounting associate professor Michael Vardon. Even if its social, environmental and political fallout is dismissed, AI hype and investment misses what is happening on the technical level. Models in the last decade became "smarter" essentially by training on larger and larger data sets. But this paradigm yields diminishing returns. Yann LeCun, former chief AI scientist at Meta, has warned that the correlation-based "learning" of models is both inefficient and insufficient when compared to human learning. Models require trillions of tokens to train. Even then, they reproduce patterns without deeper understanding, while children learn in a generalised manner from a handful of examples. "Training is waning" is the new mantra, notes one Silicon Valley insider, as the brute force approach to foundational models gets left behind. It's far from clear whether massive models, and the massive data centres that underpin them, will even be needed. Where does this leave us? The possibility of the AI bubble bursting has shifted from a niche pocket of tech critics to mainstream policy wonks. "It's time to start asking not whether there will be an AI crash, but what we should do today so that we are best prepared to respond to one tomorrow," wrote two commentators in TIME magazine earlier this year. What will this look like? Any answer here would include speculation. And yet we can garner some insights from previous bubble bursts, from tech development trends, and by extrapolating from the socio-cultural fallout we've already witnessed. Let's step through each. Another dot-com bubble First, we can compare the AI bubble with the dot-com bubble of the late 1990s. Indeed, investment leaders - including The Big Short's Michael Burry, who famously anticipated the collapse of the subprime mortgage market - are already seeing disturbing parallels between the two. Burry warns that venture capitalists are funding "loss-mak[ing] companies like never before in history". As this suggests, the investments in this current AI bubble dwarf its dot-com analogue. If this bubble follows the blueprint of the last, we should expect to see massive layoffs in personnel and liquidations of AI startups with no discernible revenue. Of course, like the first bubble, the deletion of a company doesn't mean the technologies themselves disappear. Indeed, in the orthodox economic canon, the dot-com bubble was a "baptism of fire": a painful but necessary rebirth. The trivial players, buoyed by "irrational" valuations, disappeared, but the network infrastructure they helped expand was the foundation for the truly innovative tech products to come. Part of this "soft pop" future is almost certainly correct: the infrastructure will persist, even if underused. AI will continue being baked into a multitude of products, testing the market. And tech titans, sitting on data hoards and advertising monopolies, will march on. As scrutiny is increased, belt-tightening will occur. Companies will distil their product offerings, quietly begin limiting token use, and raise their subscription prices - all moves we're already seeing play out. But the larger question is whether tech companies - now just as then - actually contribute in meaningful ways to our broader world, or even merely our economies. As one Nobel-prize-winning economist famously quipped in the 1980s: "you can see the computer age everywhere except in the productivity statistics." More recent analyses of contemporary technologies have echoed this finding, suggesting the internet has little impact on economic growth. If this is the case for AI - as the numbers, the lack of products and even the rhetoric of its chief pundits suggests - then we have a social question, not just a financial one. What price are we paying for a technology that fails to deliver even on its own terms? Small is beautiful Second, tech development is moving away from the "bigger is better" mantra. Models are becoming much smaller and more efficient. The push is from the cloud to the so-called "edge": the far more mobile and low-powered devices, like your phone, where data is actually created and used. And there's a push to move the focus from "capture it all" quantity to quality, with targeted or carefully curated data. Some of this is a welcome -- and long-needed -- shift. A deluge of critical AI research in the last few years has extensively documented the major issues with bias in foundational models. In a not-so-shocking twist, indiscriminate training on a massive archive of social material with almost no oversight creates models that reproduce significant harms. To take just two well known examples: AI models discriminate based on race and gender, while AI-generated images consistently privilege white people over people of colour. Given these issues, the slower and more careful construction of models actually tailored to their communities and attuned to their language, needs, and desires can only be beneficial. Some languages, for example Indigenous languages with strong oral traditions, are considered "low-resource", or underrepresented, with much less material in standard training sets. Switch away from English, and see the accuracy of your response plummet. Future developers might work closely with communities to create their own archive of material that better reflects their ideas and beliefs. Here we start to see a meaningful idea of data sovereignty, where groups maintain control over their models and the data that underpins them, slowly disconnecting from corporate cloud regimes. Of course, if the "small and mobile is beautiful" approach attains real traction, this will mean today's massive investment in highly centralised data centres is the wrong move. What will happen to this massively overbuilt - and, we anticipate - soon underused infrastructure? In an ironic twist, dead shopping malls have been converted into data centres in the last two years to satisfy demand - yet these data centres might themselves become empty shells, physical reminders of an obsolete vision. Post-AI pathologies Third, AI cannot be stuffed back into Pandora's box. Even if AI development takes another path, the socio-cultural, political and environmental fallout of a post-AI world will continue - or even become exacerbated. In education, researchers warn that students who constantly turn to generative AI models exhibit a kind of "doom loop" of dependence: offloaded thinking gradually causes atrophy in critical thinking and reasoning. "When kids use generative AI that tells them what the answer is [...] they are not thinking for themselves," state the authors of a Brookings Institution study. They're not learning to parse truth from fiction. They're not learning to understand what makes a good argument. They're not learning about different perspectives in the world because they're actually not engaging in the material. In politics, cutting-edge image and video models make it increasingly difficult to parse fact from fiction. Gravity glitches and six-fingered hands are gone; new generative models like Nano Banana boast physically-aware rendering. Models can now produce photo-realistic news reports, for instance, that seem to show Ukraine president Zelensky surrendering. The result is a growing pervasiveness of the "liar's dividend", where muddied lines mean even genuine material is doubted or dismissed as being synthetic. The ability of evidence to document atrocity and persuade the public is undermined, with each side accusing the other of fabricating media. In the environmental sphere, the AI-driven boom in data centre construction will have long-term impacts. While society has begun to lower carbon emissions via electrification and renewables, AI's voracious demands threaten to reverse this progress. Sustainable generative AI is a fallacy. "AI datacenters are single-handedly leading to a major reversal in climate progress globally," declared tech critic Karen Hao, citing a recent UN report. From the extraction of rare-earth minerals to the burning of dirty diesel as backup, the strain on local power grids, and the siphoning of millions of gallons of freshwater in a warming world -- the damaging effects of AI supply chain capitalism - will be felt by the ecosystems and generations to come. Rage against the machine "I'm here to tell you the mission of your generation is to destroy AI," Daily Show comedian Ronny Chieng told Harvard graduates recently, to approving cheers -- a far cry from the boos and anger that met AI evangelists advocates at similar ceremonies. One strand of rising anti-AI sentiment is directed at data centres. A report found that US$64 billion of data projects have now been blocked or delayed amid local opposition. In one sense, of course, these wins are localised and limited: the "cloud" means data centres elsewhere can still run AI. But to see them as distractions from the bigger anti-tech battle is to miss the point. As tech critic Astra Taylor and community organiser Saul Levin argue, This brewing populist resistance isn't just about limiting local development - it represents a critical new front in the fight against tech-enabled authoritarianism. Where else can people push back on job-eating algorithms, distorting deep fakes, and autonomous drone strikes? These protests and campaigns signal a gulf between the current AI vision -- "tokenmaxxing" in an "AI everywhere" world -- and the desires of everyday individuals. Of course, this disparity alone doesn't signal the death of the AI boom dream: history is full of examples of elites rolling out exploitative technologies that run roughshod over the wishes of the people. But combined with other economic, social and environmental factors, these pushbacks begin to destabilise Big Tech's future-on-rails. There are other possibilities -- slower, smaller, more convivial, more sustainable -- for technologies that contribute to our lives, our society and our world.

AI absolutism is breaking our brains. The apocalyptic future we're being sold isn't inevitable

Nor is the dreamy promise that this tech will unlock boundless potential and productivity Everything we hear about artificial intelligence is conflicting, and hearing about it feels inescapable. AI is terrible. AI is wonderful. It will break the world. It will transform the future. It's essential to embrace it. It's a moral imperative to abstain from using it. Already, AI is projected to generate nearly unfathomable amounts of revenue. In the last quarter of 2025, it represented nearly 60% of the growth in the US economy. Already, pundits and economists wring their hands about what calamity will befall us if and when the AI bubble bursts. Since ChatGPT, the first of the large language models, was released in late 2022, more than half a million workers in the tech industry alone have lost their jobs. Any mention of AI tends to be accompanied by warnings that deeper jobs cuts across many more industries are coming for us all. Jensen Huang, CEO of chip giant Nvidia, said in 2025: "Every job will be affected, and immediately. It is unquestionable. You're not going to lose your job to an AI, but you're going to lose your job to someone who uses AI." In January, Anthropic CEO Dario Amodei predicted: "AI isn't a substitute for specific human jobs but rather a general labor substitute for humans." Increasingly, people young and old flock to a new gold rush in Silicon Valley to toil away on AI-fueled startups. Many of them are driven less by idealistic enthusiasm and more by the dread of missing a ticket for the last train to wealth - and getting stuck forever in the "permanent underclass" that, with any luck, they themselves will create. What all these divergently apocalyptic ideas hold in common is their AI absolutism - a way of seeing AI as a godlike force that will either hasten a golden age of productivity and innovation, or will doom humanity. It mirrors the political polarization of our era and even the zealotry found in religious fanaticism. This is by design. Contradictory as they may be, all these arguments and anxieties fit neatly into the overarching message of the people building this technology: AI's dominance is inevitable. Get on board or you will be left behind. The robber barons of our age stand to profit wildly from not only enthusiasm about their star product, but also, the terror of it. "If you want to justify this enormous valuation in your IPO, you need to point to the revenue stream that you're going to generate in the future," said Suresh Naidu, a professor at Columbia University's department of economics. "You just need to make it look like you have something that can eat all the work on the planet, so that an investor will think: 'Oh wow, I don't want to miss out on this thing.'" Naidu isn't refuting claims that AI will cut into jobs or upend certain industries. He called the technology "transformative" and said that he uses it every day in his work as a researcher and academic. It's just that when he zooms out and puts AI and all its attendant promises and warnings in historical context, he sees a lot of hype. There is no control group Anil Dash, the former CEO of the startup Glitch, who's been writing about tech for decades, is also unconvinced that the AI we're being sold will do all the things tech CEOs are predicting it will do. "Any technology that you invest like a trillion dollars into is going to be able to do a lot of things, good or bad. [AI is] a leap forward. I don't think we've ever had a machine learning system that can do as many things as this one does," he said. But "there's so much noise that it's hard to tell what the domains of applicability are." Coding is an exception, he said. It's easier to test an AI model's coding output because it will clearly work, or it won't. Many other applications for the tech are much more subjective and therefore less prone to immediate job replacement. That's why the tech industry has made the deepest job cuts so far - though, amid layoffs at companies such as Amazon, Meta and Block, reports from employees have emerged saying the AI productivity gains their bosses trumpet are overblown. Even the role AI is playing in those layoffs and reductions to entry-level positions isn't entirely clear. Martin Beraja, a professor at UC Berkeley Haas School of Business who studies technological innovation and business cycles, said recent studies that have drawn connections between the release of ChatGPT and a decline in entry-level software jobs are "problematic". There was "a buildup of jobs in [tech] coming out of the pandemic, and once ... consumption patterns moved away from online to the real world again, now we had too many people working in the industry that we didn't really need", Beraja said. Some of the biggest and most loudly pro-AI players in tech have arrived at similar conclusions as AI critics. Venture capitalist Marc Andreessen proclaimed in March that overstaffed companies are using AI as a "silver-bullet excuse" to clean house. In May, OpenAI CEO Sam Altman retreated on some of his prior claims of massive job replacement by AI, saying: "I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened." And if AI's worst-case scenario for tech jobs plays out - which would indeed be very bad for many people - that's still nowhere near the apocalyptic future of labor that many fear. "Is it, in fact, going to destroy all of the jobs?" Naidu asked. "I'm not convinced. Even take software. Software is only about 4 to 6% of GDP. So it's a lot, but it's not like the whole economy can be replaced by Claude Code." Convincing people that AI will replace human workers in droves is a clever marketing tactic. Not only does it stoke rabid investor speculation, but it distracts from a more realistic application of AI for the global workforce, stretching far beyond the borders of the tech industry: using AI to surveil and micromanage employees to squeeze yet more productivity out of them, all the while pressuring them to feel grateful that they have any kind of work at all. Gig workers, the people who pick you up in Ubers and deliver your food on platforms like DoorDash, have already been the guinea pigs for this kind of algorithmic management, and labor experts predict it will spread. It can feel like we're living in an experiment when it comes to the rise of AI. Naidu would like us to update that framing. "An experiment implies a control group of something that's not affected. There's no control group here," he said. Remember there are alternatives The version of AI that we're being sold doesn't have to be the version we buy. Nor does it need to be the story we believe in. This isn't an argument for an abstinence-only relationship with AI, something that has too much in common with evangelical Christianity's unrealistic stance on premarital sex. Anyone with common sense can see how those kinds of ascetic codes play out in reality. It's happening already with AI. "AI is just another technology Americans don't like but can't stop using," the Washington Post's Shira Ovide wrote earlier this year, referring to the polarized divide between polling that shows how much they distrust the tech and numbers of rapid user growth in the past year. Instead, this is an argument for moderation. Beraja, the UC Berkeley professor, said there's too much focus on AI as a job replacement technology. Outside a few industries like tech, he said studies show that the most effective ways for people and companies to use AI is not to replace workers, but to learn more, and learn faster. "Where I think we have to get to is, there can be alternatives," said Dash. "What we can imagine is, rather than the ChatGPT killer, a lot of different little AIs from little responsible players." A few are already quietly cropping up, harkening back to earlier and more optimistic days in the internet's history, and offering a glimpse of what could be possible if people took AI into their own hands. And for the industries and jobs that AI is upending, upheaval may open the way for a resurgence in worker power as white-collar workers begin to see the appeal of solidarity, whether with colleagues in their office or workers in the blue-collar world. After all, the Industrial Revolution, an earlier time of great technological transformation that strangely mirrors our current moment, was a key catalyst for the labor movement - even if its wins took time.