Discussions around the topic of whether we’re in an AI bubble have been going on for more than three years now.
Every time some major news is reported, the same question resurfaces. Is this all just hype?
Recently, a handful of companies have been spending on AI at a pace that looks more like a national infrastructure project than a software rollout.
The money is staggering, the timelines stretch years, and the power requirements match small countries.
But whether this becomes the foundation of a new industrial era or the next great tech hangover depends on how much of that spending turns into real productivity before the bills come due.
The new industrial boom
Earlier this month, AMD announced it would supply OpenAI with six gigawatts of computing power using its next-generation MI450 chips.
The deal is estimated to be worth tens of billions of dollars, and it gives OpenAI the right to buy nearly 10% of AMD at one cent a share if performance milestones are met.
And this is just one of several enormous supply commitments the ChatGPT maker has signed as it scrambles to secure hardware for its expanding fleet of data centres.
OpenAI has already agreed to buy roughly $300 billion of cloud capacity from Oracle beginning in 2027, the largest cloud contract ever reported.
The deal requires power equivalent to more than four million homes.
Both AMD’s and Oracle’s share prices soared over 40% following the announcements, with Oracle’s market value now approaching $1 trillion.
Broadcom joined the list this month, revealing a partnership with OpenAI to co-design custom AI chips optimised for inference.
The two companies plan to deploy racks of these accelerators starting next year, targeting ten gigawatts of capacity.
Industry analysts estimate that one gigawatt of next-generation data centre infrastructure costs about $50 billion to build.
These figures are almost hard to process. Together, the AMD, Oracle and Broadcom deals point to more than a trillion dollars of future AI infrastructure spending through the end of the decade.
The money keeps flowing
Venture capital has followed the same pattern. According to KPMG’s latest Venture Pulse report, global VC investment hit $126 billion in the first quarter of 2025, the highest in ten quarters.
Yet the number of deals fell to a record low, meaning most of the money went into a few mega-rounds.
The largest of all was OpenAI’s $40 billion raise, the biggest private funding round in history.
The concentration is striking. Investors are showing enthusiasm for late-stage AI bets but hesitation elsewhere.
In the United States, $91 billion of the total VC capital landed, while Asia-Pacific activity fell to a ten-year low.
This pattern of massive sums chasing a narrow set of names is typical of investment manias, even when the underlying technology is real.
And now the corporate world is joining in. JPMorgan announced a plan to invest up to $10 billion directly in companies tied to artificial intelligence, energy, and defence as part of a $1.5 trillion national-security financing program.
Its chief executive, Jamie Dimon, said the United States has become too dependent on foreign supply chains for critical materials and technology. The line between industrial policy and private investment is beginning to blur.
Power, chips and the laws of physics
Behind every glowing chatbot is a web of very physical constraints. Artificial intelligence is not an ethereal technology.
It runs on silicon, memory, and electricity.
The servers used to train and run large models consume enormous amounts of energy.
OpenAI alone now operates roughly two gigawatts of computing capacity.
Its combined partnerships will push that toward thirty gigawatts by the end of the decade.
Power is emerging as the real bottleneck in this case.
In many regions, grid connections and transformer lead times already stretch several years.
Electricity prices for consumers are climbing as utilities race to secure new capacity.
Industry estimates suggest the current wave of AI data centres could require more than 2% of US GDP in cumulative spending if current trends continue.
Some investors dismiss the comparison to past bubbles on the grounds that this infrastructure has residual value.
Even if model training slows, the data centres can still host cloud services. That is true to an extent, but utilisation rates matter.
Empty racks and half-used power contracts can turn expensive assets into stranded ones, as telecom firms discovered with their unused fibre networks two decades ago.
Is it an AI bubble or a build-out?
Economists are comparing the current “AI mania” to Britain’s Railway Mania of the 1840s and the telecom bubble of the late 1990s.
In both cases, investors were right about the long-term potential but wrong about the timing and profits.
Railways changed the world, yet railway stocks collapsed by 70%. The same happened to telecom firms after the fibre-optic glut of 2001.
In the US, spending on AI hardware and software has already risen by about half a percent of GDP since the launch of ChatGPT in 2022.
That makes it roughly the same scale, relative to the economy, as the 1990s telecom build-out.
Power companies are planning an additional $1.5 trillion of generation and grid capacity over five years to meet the surge in demand from data centres.
The question is whether this investment wave ends in a smooth plateau or a sharp contraction.
Unlike the dot-com era, the companies leading this race, such as OpenAI, Nvidia, Microsoft, Google, Amazon, Meta, Oracle and the like, are funding it mostly from cash flow.
Their balance sheets can absorb the costs for now.
But the spending is rising so quickly that even these giants are starting to issue more debt.
The logic of network industries also fuels the excess. Each firm believes it must build the biggest and fastest platform to capture monopoly-like returns later.
In that environment, restraint appears to be losing.
Recent reports suggested that OpenAI alone could burn through $115 billion in cash by 2029 and that the industry as a whole may require $2 trillion in annual revenue by 2030 just to fund its computing needs.
Bain & Co. expects an $800 billion shortfall between projected costs and likely revenues, a gap that would normally make lenders nervous.
What history might be whispering
Bubbles rarely look reckless in real time. They are built on stories that feel inevitable.
Railroads would reshape transport, fibre would connect the world, and housing prices would never fall nationwide.
The narrative around AI is similarly powerful. It’s about machines that reason, software that learns, productivity that compounds.
The technology is genuine. The question is whether the financial expectations layered on top of it are.
Even some of AI’s biggest champions quietly admit the market looks frothy. Sam Altman says investors are “overexcited” but insists AI is still the most important development in decades.
Mark Zuckerberg believes a bubble is possible, but argues the bigger mistake would be under-investing.
Jeff Bezos has called it an “industrial bubble,” one that may still lift productivity across every industry.
The boom today is smaller than the railway or housing manias in terms of GDP share, but the economy is also slower-growing.
The United States expanded by around 4% a year in the late 1990s. Growth now averages 2%.
Much of 2025’s output increase has come directly from AI-related investment. If that spending slows, so might the broader economy.
In terms of stock market valuations, there’s still room to grow before we reach 2000 levels of hype.
The risk is not a catastrophic crash but a drawn-out hangover. The telecom bust in 2001 triggered a mild recession, yet employment took four years to recover.
The recovery that followed relied heavily on the housing bubble, which then produced the next crisis.
If the AI cycle cools in a similar way, the economic impact could be deeper because underlying growth is weaker.
For now, the race continues. Chips are sold out years ahead, utilities are redrawing maps to feed data centres, and financial markets are rewarding boldness over caution.
Whether that confidence becomes the foundation of the next industrial age or the prelude to another correction will depend on something that has never been in short supply in Silicon Valley: belief.
The post What keeps the AI bubble from bursting — and what it means for investors appeared first on Invezz