The AI Boom Isn't a Bubble.

It's Something Weirder.

We’re living through one of the largest corporate capex waves in history and most of us have no idea it’s happening. In 2025, the big US hyperscalers are on track to invest ~$300 billion in AI-centric infrastructure, with analyst roadmaps pointing toward ~$490 billion in 2026. To put that in perspective, that’s roughly equivalent to building the Apollo programme about once a year. Microsoft is around $80B (FY25); Alphabet ~$75B (CY25); Meta $66–72B (2025); Amazon ~$100B (2025, with the majority to AWS). These aren’t speculative bets by starry-eyed venture capitalists. These are the most profitable corporations on Earth, spending cash they already have.

And yet, commercial gen-AI revenue is still only in the tens of billions; meanwhile total AI spend in 2025 is ~$300B, with gen-AI spend ~$69B. The direction is clear; the gap is real.

This has cleaved observers into two camps. On one side, sceptics like journalist Derek Thompson warn we’re witnessing a bubble that will make the dot-com crash look quaint. On the other, technology leaders from Jeff Bezos to Sam Altman argue we’re in a necessary, if chaotic, infrastructure boom that will define the century. Both sides marshal compelling evidence. Both could be right. Or both could be catastrophically wrong.

Because here’s what I’ve come to believe after examining the data: the AI investment cycle is neither a pure bubble nor a sustainable boom. It’s a fragile hybrid, a solid core wrapped in a speculative shell, and its ultimate fate depends on a race against time that’s already underway.

The spending-to-revenue gap is just the beginning of the alarming bits. According to Bain & Company, for today’s AI investments to become profitable, the sector needs to generate $2 trillion in new annual revenue by 2030 - even after assumed savings -leaving a very large shortfall if monetisation lags. The entire market capitalisation of these companies rests on the assumption that demand for AI will grow by orders of magnitude within five years.

And then there are the financial mechanisms that should make anyone who lived through 2008 deeply uncomfortable. Nvidia and OpenAI have a letter-of-intent partnership to deploy ~10 GW of Nvidia systems, with Nvidia intending to invest up to $100B as deployments progress. Microsoft’s OpenAI tie-up includes very large Azure commit/usage, so OpenAI’s Azure consumption contributes to Microsoft’s reported revenue. Meanwhile, GPU-rich clouds are raising multi-billion debt with GPUs as collateral.

Perhaps most unsettling: depending on the measure, ~40–60% of US real GDP growth in H1-2025 is explained by IT/AI investment. If Big Tech pulled back spending to 2022 levels, sell-side estimates suggest it would erase ~30% of the revenue growth currently expected for the S&P 500. The American economy has effectively outsourced its growth engine to a handful of corporate boards betting on an unproven technology.

History suggests we should be careful about crying bubble, though. Historical work finds most major innovations come with localised speculative bubbles, one landmark sample shows ~73% of big innovations since the 1800s had bubble-like price patterns. The railway mania of the 1840s bankrupted thousands of investors and destroyed hundreds of companies. It also left Britain with a national rail network that powered a century of industrial dominance. The fibre-optic boom of the late 1990s wiped out about $5 trillion in market value across the broader dot-com crash. It also wired the world for the internet age.

The pattern is consistent: massive over-investment, financial carnage, then decades of productivity gains built on the infrastructure left behind. As Jeff Bezos recently noted, calling the AI boom a bubble misses the point. “It’s an industrial bubble,” he said, “and that’s different.”

The quality of capital matters here. The dot-com bubble was fuelled by speculative debt, billions loaned to companies with no revenue and no path to profitability. Today’s AI build-out is primarily financed by companies generating hundreds of billions in annual free cash flow. Microsoft, Alphabet, Amazon, and Meta aren’t gambling with borrowed money. They’re spending profits from businesses that print cash.

And adoption is happening faster than it looks. AI companies on Stripe reached $1 million in annualised revenue in a median of 11.5 months, faster than any previous technology cohort. The St. Louis Federal Reserve estimates that generative AI is being adopted faster than PCs or the internet. A Gallup–Walton survey found that ~60% of teachers now use AI for lesson preparation, with weekly users saving ~6 hours per week (~6 weeks per year). These aren’t hypothetical benefits. They’re measurable productivity gains happening right now, just not always showing up in traditional ROI calculations.

At the core is a legitimate, fundamentals-driven boom. AI is almost certainly a general-purpose technology on the scale of electricity or the internet. The infrastructure being built will likely underpin decades of economic activity. The investment is coming from strong balance sheets, not fragile debt. This is real.

But surrounding that core is a speculative periphery exhibiting every classic bubble symptom. Index-level valuation is stretched: the S&P 500 price-to-sales is ~3.4 versus a long-run median ~1.6 (as of 6 Oct 2025, AEST). Venture capital has become a monoculture: for the first time since the dot-com peak, more than half of global VC dollars in H1-2025 flowed into a single sector—AI. Start-ups are raising billions at ten-figure valuations with no products and no clear plans.

The system’s stability hinges on a race: can enterprise adoption and monetisation scale rapidly enough to validate the investment before macroeconomic headwinds, rising capital costs, or a crisis of confidence in the speculative periphery triggers a correction?

And here’s the uncomfortable bit: the race isn’t going well.

Behind closed doors, corporate technology leaders are increasingly worried. A survey by Solvd found that 71% of CIOs and CTOs believe their executive leadership holds unrealistic expectations about AI’s return on investment. They’re being pressured to deploy AI whilst simultaneously struggling with poor data quality, fragmented systems and integration nightmares that make successful implementation nearly impossible.

The costs are spiralling beyond projections. Gartner warns that organisations failing to understand AI’s scaling dynamics can miscalculate costs by 500% to 1,000%. The proof-of-concept phase alone costs large enterprises an average of $2.9 million. And Gartner predicts that most enterprise software will increase in price by at least 40% by 2027 due to embedded AI features, whether customers want them or not.

This is creating a dangerous feedback loop. The initial wave of enterprise AI investment was driven by fear of missing out. That phase is ending. Recent market reads show the percentage of finance leaders planning to increase gen-AI budgets has fallen to ~27%, down from ~53% a year ago. The market is bifurcating: the minority seeing strong returns will invest more, but the majority experiencing negligible ROI are pulling back.

This matters because hyperscale cloud providers, the ones spending hundreds of billions on data centres, have based their entire investment thesis on exponential growth in enterprise demand. If that demand flattens because enterprises hit an “ROI wall,” the hyperscalers will revise forecasts and cut capital expenditure. That, in turn, would squeeze Nvidia’s order book and pressure the private-credit structures now financing portions of the periphery, including GPU-collateralised loans, potentially triggering the very correction that sceptics predict.

There’s a crucial difference between the AI boom and previous infrastructure build-outs that both optimists and pessimists tend to gloss over: asset durability. When the railway bubble burst in the 1840s, the steel tracks remained. When the fibre-optic bubble burst in 2001, the “dark fibre” buried in the ground was still there, ready to carry traffic for decades. These crashes were painful, but they left behind durable infrastructure that society could repurpose.

AI infrastructure isn’t like that. The core assets are GPUs with short economic half-lives: in practice, they’re depreciated over ~3–5 years, and architectures are turning over faster (Hopper to Blackwell in roughly two years). Data centres filled with current-generation chips aren’t valuable, salvageable infrastructure when the bubble bursts. They’re warehouses full of rapidly depreciating silicon. This means two things. First, investors can’t wait decades for returns because the assets won’t last that long. Second, when the correction comes, and some form of correction is inevitable, there may not be a cushion of residual value to soften the landing. The capital destruction could be more total than in previous industrial bubbles.

There’s another factor that doesn’t fit neatly into either the bubble or boom narrative: geopolitics. Governments increasingly view AI infrastructure as critical national infrastructure, on par with highways or the electrical grid. The race between the United States and China isn’t just economic. It’s seen as existential by policymakers on both sides. This means that even if private investment falters, state intervention could prop up the build-out.

The UK now treats compute as national infrastructure in its Compute Roadmap. Across Europe, sovereign-cloud and “strategic autonomy” agendas are accelerating to reduce dependence on US hyperscalers. In the Gulf, Saudi Arabia and the UAE are pouring sovereign wealth into AI data centres. When investment becomes a matter of national security rather than pure economics, normal market disciplines break down. This could prevent a crash or make one more catastrophic. State involvement can sustain investment through periods when private capital would rationally pull back. But it can also delay necessary corrections, allowing imbalances to grow larger and more dangerous.

The dashboard of warning signals is decidedly mixed. Enterprise spending intentions are weakening but not collapsing. Hyperscaler capital-expenditure guidance remains strong, though growth is slowing. Nvidia’s order book is still robust. Market breadth is narrowing, a worrying sign, but hasn’t reached crisis levels.

We’re somewhere between the peak and the hard-work phase, the point where initial hype gives way to the grind of implementation. This is normal. Every general-purpose technology goes through this. The question is whether we emerge from the grind into real productivity gains fast enough.

We’re probably 12 to 24 months away from a moment of truth. That’s roughly when the current wave of enterprise AI pilots will have run their course and companies will make go-or-no-go decisions on scaling. The first generation of AI-focused data centres will have been operational long enough to show whether utilisation meets projections. The rapid depreciation of early GPU investments will force a reckoning about replacement cycles and total cost of ownership. Macroeconomic conditions - interest rates, inflation, potential recession, will either provide tailwinds or headwinds. That window runs through mid-2027 for this capex cohort.

If enterprise adoption accelerates and companies start seeing genuine productivity gains that translate to willingness to pay, we’ll look back on 2025 as the moment when the sceptics lost the argument. The boom will have validated itself. If adoption stalls or fails to generate ROI at scale, we’ll see a pullback that could range from a mild correction to a systemic crisis, depending on how much hidden leverage exists in the private-credit markets financing the periphery.

For those of us not running technology companies or managing billion-dollar investment portfolios, this might all seem abstract. It isn’t. If the AI boom continues, we’re likely entering a period of rapid technological change that will reshape work, education, healthcare and creative industries. The productivity gains could be transformative, but they’ll be unevenly distributed. The same Gallup survey that found teachers saving six weeks a year also revealed that AI adoption is highest among management and information services - white-collar knowledge work, whilst remaining minimal in manufacturing, construction, and service industries. The gap between those who benefit and those who don’t could widen dramatically.

If the boom becomes a bust, the consequences extend beyond tech stocks. An economy that’s grown dependent on AI capital expenditure for growth would face a sharp contraction. The Federal Reserve would face pressure to cut interest rates aggressively, potentially reigniting inflation. Venture capital would retreat from technology entirely, potentially starving the next generation of genuine innovations. And public trust in technology companies, already strained, could collapse entirely.

Either way, the energy implications are massive. Baselines suggest triple-digit gigawatts of new capacity are needed this decade: one McKinsey case puts ~156 GW by 2030, and Goldman Sachs sees up to +165% global data-centre power demand by 2030 versus 2023. Unless matched by clean-energy investment, this represents a significant climate challenge.

We need honest accounting. The financial engineering obscuring the true cost and risk of AI infrastructure helps no one in the long run. Special-purpose vehicles, circular vendor financing, and creative revenue recognition might smooth quarterly earnings reports, but they prevent the price discovery and risk assessment that markets require to function. We need realistic expectations about timelines and returns. AI is genuinely transformative, but transformation takes time. The internet wasn’t built in three years. Neither will AI’s economic impact materialise on a venture-capital timetable. Companies and investors who can maintain perspective, who can tolerate the messy, non-linear process of turning infrastructure into value, will do better than those chasing quarterly metrics.

We need to focus on the unglamorous fundamentals that make AI actually work: data quality, integration architecture, change management, training. The shiniest model is useless if your data is rubbish or your employees don’t trust the system. The enterprises seeing genuine ROI are the ones that did the boring, expensive groundwork. And we need broader distribution of benefits. If AI’s productivity gains flow primarily to a handful of companies and their shareholders whilst communities bear the costs - higher electricity prices, environmental impacts, labour disruption, the social licence for continued development will erode. That’s how you get a real bust: not from financial mechanics, but from political backlash.

Here’s what I keep coming back to: both the optimists and pessimists are probably right about their core claims. The sceptics are right that the current spending-to-revenue ratio is unsustainable. They’re right that speculative excess is real. They’re right that much of this capital will be destroyed. The optimists are right that AI is transformative. They’re right that infrastructure investment at this scale is necessary for general-purpose technologies. They’re right that focusing on near-term profitability misses the point.

The truth is that we can have a genuine boom and a painful bust. We can build incredibly valuable infrastructure and see a financial crisis. We can be at the beginning of a technological revolution and experience a severe market correction. The question isn’t bubble versus boom. It’s how much damage we sustain during the inevitable turbulence of a boom, and whether the value created justifies the carnage along the way.

Right now, we’re in the fragile middle, past the point of easy retreat, not yet at the point of vindication. The decisions made over the next 12 to 24 months, by executives, investors, policymakers, and enterprises, will determine which version of this story we tell in 2030. My money is on messy. A correction is coming, something between a mild reset and a genuine crisis, but the infrastructure and capabilities being built are real enough that we’ll emerge from it with something genuinely valuable, even if the path from here to there destroys a lot of wealth and a fair few reputations along the way.

That’s not the definitive answer anyone wants. But it’s the honest one. The AI story is still being written. Neither the triumphalists nor the doomsayers have won yet. And that, perhaps more than anything, is what makes this moment so fascinating and so dangerous.