When Microsoft CEO Satya Nadella warned at Davos that artificial intelligence “risks becoming a speculative bubble,” the statement carried particular weight. This was the chief executive of a company that has invested $14 billion in OpenAI. For AI not to be a bubble “by definition,“ Nadella argued, “it requires that the benefits of this are much more evenly spread” beyond tech firms. A telltale sign of bubble territory? “If all we’re talking about are the tech firms.”

When even the most invested parties start hedging their language, the big question shifts. It’s no longer whether we’re in a bubble – we’re already there, even if Azeem Azhar’s tracking dashboard hasn’t officially declared it yet. The question becomes: what infrastructure survives the inevitable correction of the AI bubble?

Ben Thompson’s November 2025 piece The Benefits of Bubbles offered the most sophisticated framework for answering exactly this question. Drawing on Carlota Perez and the recent work of Byrne Hobart and Tobias Huber in Boom: Bubbles and the End of Stagnation, Thompson distinguishes between bubbles that leave useful physical capacity and those that don’t. The dot-com crash was devastating, but it left behind dark fibre networks that enabled everything from Netflix to remote work. The question isn’t whether the AI bubble will pop – Thompson concedes it almost certainly will – but whether the speculative investments being made today will prove as transformative.

His most compelling case centres on power generation. Unlike GPUs that depreciate over five years, power infrastructure lasts decades. And unlike dark fibre that served a single industry, electricity serves every possible use. As Thompson writes:

It’s hard to think of a more useful and productive example of a Perez-style infrastructure buildout than power. It’s sobering to think about how many things have never been invented because power has never been considered a negligible input from a cost perspective.

The argument gains force when Thompson identifies what already constrains AI progress. Microsoft’s CFO Amy Hood noted on the company’s earnings call:

As you know, we’ve spent the past few years not actually being short GPUs and CPUs per se, we were short the space or the power, is the language we use, to put them in.

Amazon’s CEO Andy Jassy made a similar observation: power is currently the bottleneck, not semiconductors. If AI demand forces massive new power generation capacity, and if that infrastructure survives the inevitable market correction via bankruptcy or stranded assets, the post-bubble society benefits from essentially free electricity for decades of future innovation.

There’s just one problem with the discussion Thompson sparked: everyone has fixated on the wrong technology.

The nuclear distraction

The conversation around AI and power generation has become dominated by small modular reactors (SMRs). Tech giants are signing memoranda of understanding with nuclear startups. OpenAI’s Sam Altman backs Oklo. Google has deals with Kairos Power. Microsoft has committed to buying power from Three Mile Island. The National Governors Association publishes dispatches on “powering a new era of innovation” through this technology.

It’s easy to see why nuclear captures attention. The technology provides 24/7 baseload power independent of weather, exactly what AI workloads supposedly require. SMRs promise factory fabrication, shorter build times, and incremental capacity additions. The symbolism resonates too – serious infrastructure for a serious technology revolution, not the intermittent generation of renewables.

But examine what’s actually being built, and nuclear barely features.

According to the International Energy Agency’s analysis, SMRs won’t enter the mix until after 2030. The first commercial deployments target the late 2020s and early 2030s, with Ontario Power Generation’s BWRX-300 fleet expected to start operating in 2029 at the earliest. Google’s Kairos deal stipulates bringing their first SMR online in 2030. Most projections show nuclear becoming “increasingly important” towards the end of this decade “and beyond” – industry euphemism for “not yet relevant.”

Even data centre infrastructure specialists such as Schneider Electric acknowledge this reality:

SMRs remain an emerging power source facing regulatory hurdles, public concerns over nuclear fuel, and high upfront costs currently hindering their adoption.

The economics explain the timeline. Lux Research estimates that first-of-a-kind SMRs could cost nearly three times as much as natural gas – $331/MWh versus $124/MWh – and sees cost-competitiveness delayed until 2035. Idaho National Laboratory suggests costs could decline through modular construction and standardised deployment, but only after enough units are built to move down the learning curve. The first real-world test of whether this theory holds arrives with Ontario’s fleet, years from now.

Meanwhile, between 2024 and 2030, what’s actually meeting the surge in AI data centre electricity demand? IEA data predicts that renewables – primarily wind, solar and hydro – will meet nearly 50% of additional demand, growing at 22% annually. Natural gas and coal together meet over 40%. Nuclear’s contribution to incremental demand growth remains minimal until the 2030s.

Tech companies are the world’s largest corporate buyers of clean energy, not through aspirational SMR deals but through massive renewable power purchase agreements (PPAs) being signed and built right now. Meta has contracted over 12 gigawatts of clean energy capacity. Amazon, Microsoft and Google dominate the corporate renewable PPA market. Google just acquired Intersect Power specifically for its expertise in co-located solar, battery storage and flexible backup power.

These aren’t speculative commitments for future technology. Meta’s recent 650 MW solar deal with AES for Texas and Kansas expects projects to be operational in 2-3 years. The company’s partnership with Silicon Ranch exceeds 1,500 MW of solar installations. Amazon’s renewable agreements span utility-scale wind and solar farms across continents. By the end of 2025, Meta alone expected to help add 9.8 GW of renewable energy to US power grids – equivalent to powering over 2 million homes.

The scale is staggering, the timelines immediate, and everyone is focused on nuclear.

Why renewables actually work

Thompson’s infrastructure thesis requires three characteristics:

• Durability,
• Fungibility,
• The right bankruptcy timing.

Solar and wind deliver all three better than SMRs.

Consider lead times. Renewable projects reach operation in 2-3 years, not 5-10. They’re being built now, during the bubble’s inflation phase, not scheduled for the period after it pops. When Thompson identifies power as already the constraint rather than chip availability, he’s describing a problem renewables can actually address on relevant timescales.

Location constraints favour renewables dramatically. You can build solar and wind wherever resources are good, then connect to grids. An SMR requires specific site characteristics, extensive regulatory approval, community acceptance and integration with existing nuclear infrastructure or new purpose-built facilities. A cancelled data centre in Iowa doesn’t strand a West Texas solar farm. It just feeds different demand.

The fungibility advantage is even starker. SMRs sized for data centre loads become awkward assets if AI demand disappoints – too large for some applications, too small for others, in locations optimised for vanished customers. But renewable generation capacity is the most fungible infrastructure imaginable. Every kilowatt-hour goes somewhere useful. Solar panels don’t care whether they’re powering AI training runs, residential air conditioning, or industrial processes.

This matters enormously for Thompson’s bankruptcy scenario. His optimistic case requires companies to build infrastructure, then go bankrupt before extracting full value, leaving society with essentially free capacity. But Microsoft, Google and Amazon have deep enough pockets to absorb losses for years. By the time they admit defeat on specific projects, they’ll likely have paid debt service on those assets, eliminating the “free infrastructure” benefit. Or they’ll cancel mid-construction, leaving nothing useful – plausible for partially-built nuclear reactors, less so for wind turbines and solar installations that start generating returns module by module.

Renewable PPAs create a different dynamic. These are 15- to 25-year commitments that provide guaranteed revenue to developers, regardless of whether specific AI projects succeed. If Microsoft abandons a data centre, the solar farm it contracted keeps generating power for someone else. The financial obligation remains, but so does the productive asset. The bubble provides capital and demand justification, but the infrastructure becomes grid-scale capacity serving everyone.

The numbers bear this out. Goldman Sachs research shows that data centre power demand is rising by 175% between 2023 and 2030, with renewables accounting for nearly half of the incremental growth. That’s roughly 270 TWh of new renewable generation between 2024 and 2030, driven substantially by AI infrastructure deals. These aren’t hypothetical megawatts coming online in 2035. They’re being built right now, during the speculative AI bubble mania, exactly when Perez’s framework suggests infrastructure gets installed.

The invisible revolution validated

This reframes our November piece not as optimistic speculation but as documentation of dynamics already underway. We argued then that AI’s voracious energy appetite, bubble or not, was driving exactly the infrastructure investments that make distributed renewable grids work, through market forces rather than policy mandates. The mechanism was economic self-interest wrapped in smarter technology, not sacrifice driven by climate guilt.

The data validates this almost embarrassingly well. Tech companies aren’t building renewables from environmental conviction – though they’ll certainly claim credit for sustainability leadership. They’re building renewables because power is the binding constraint, and solar-plus-storage delivers faster than alternatives. As data centre executive Jeff Tench of Vantage Data Centers testified to the US Senate:

It does need to be a reliable, grid dispatchable source, which I believe can be accomplished with the right combination of energy source for generation and energy storage.

The market has spoken with remarkable clarity. According to IEA data, natural gas currently supplies over 40% of US data centre electricity, while renewables supply 24%. But for additional capacity being built, the economics increasingly favour renewables. Long-term PPAs provide the financial certainty developers need. Co-location with battery storage addresses intermittency. Build times beat alternatives. And crucially, the approach scales – you can sign 10 gigawatts worth of renewable PPAs far more easily than licensing even a single gigawatt of nuclear capacity.

This creates precisely the coordinating mechanism that Hobart and Huber identify as bubbles’ cognitive capacity benefit:

When one group makes investments predicated on a particular vision of the future, it reduces risk for others seeking to build parts of that vision.

Renewable developers can finance projects knowing tech giants will honour multi-year commitments. Equipment manufacturers can scale production. Supply chains emerge. Standards coalesce. Innovation happens in parallel rather than sequentially.

The irony is that this happens almost invisibly while attention focuses on nuclear. Corporate renewable PPAs don’t inspire TED talks or generate venture capital buzz. Yet that invisible infrastructure layer enables everything that follows – not just AI workloads, but also the broader electrification required for EVs, heat pumps, industrial processes, and grid decarbonisation.

Where scepticism belongs

None of this makes the AI bubble any wiser, or its valuations any more justified. Thompson’s framework asks what infrastructure survives the crash, not whether the crash can be avoided. And enormous grounds for scepticism remain.

The chips being purchased depreciate rapidly and deliver value only if the models they train prove commercially viable at scale. We’re witnessing hundreds of billions in capital expenditure chasing revenue that currently measures in tens of billions – the difference between Singapore’s GDP and Somalia’s, one measure of the gap between AI infrastructure spending and current revenue. Market concentration means 75-90% of index gains depend on a handful of AI companies. Scaling laws appear to be plateauing, shifting competition from “bigger models” to architectural innovation nobody can predict.

The constraints are appearing globally. In Germany, Europe’s largest data centre market, Frankfurt’s data centres already consume 40% of the city’s electricity, with grid connections unavailable until the mid-2030s. Marina Köhn of Germany’s Federal Environment Agency characterises the expansion:

The expansion of data centers is strongly driven by speculation based on the current AI hype and, unfortunately, not demand-oriented.

Most importantly, the models themselves might not deliver returns justifying current investment. History suggests every transformative technology triggers infrastructure buildouts that massively overshoot near-term demand. Railways did this in the 19th century. Telecoms did it in the 1990s. The technology eventually proves valuable, but not before a painful adjustment between fantasy economics and reality.

Beyond these familiar concerns, the bubble creates a less visible problem: capital diversion on an economy-warping scale. As investor Paul Kedrosky observed, data centre spending accounted for roughly half of US GDP growth in the first half of 2025. This concentration goes well beyond impressive into the market distorting.

Kedrosky draws parallels to the 1990s telecom boom, when massive infrastructure spending created what he calls a “death star” that sucked capital away from manufacturing. Small manufacturers found it harder to raise money as investors chased telecom returns, forcing their cost of capital higher, just as China joined the WTO and tariffs dropped. The capital crunch made American manufacturers less competitive precisely when they faced new global competition.

The same dynamic repeats today. Private equity firms managing hundreds of billions prefer writing thirty $50 billion cheques to managing hundreds of $5 million investments across small manufacturers. The internal logic of capital allocation creates structural bias against exactly the kind of manufacturing reshoring that policy ostensibly supports. Trump’s tariffs attempt to reverse offshoring, whilst AI’s capital appetite makes financing domestic production harder. The bubble doesn’t just misallocate capital to potentially unprofitable AI ventures – it starves other parts of the economy that might generate real returns.

Yet Thompson’s crucial insight remains: painful adjustments don’t negate infrastructure benefits. They may actually enable them. The railway and telecom crashes left behind physical capacity that subsequent generations used at essentially zero marginal cost. The speculative investors who funded construction lost money. Society gained cheap infrastructure.

If the AI bubble produces this renewable generation capacity through corporate PPAs, that infrastructure remains productive regardless of whether GPT-7 delivers the returns OpenAI projects. The solar farms don’t disappear. The wind turbines keep spinning. The battery installations continue storing energy. And unlike dark fibre serving a single industry or SMRs optimised for specific locations and load profiles, renewable generation serves every possible future use of electricity.

This is where the energy infrastructure story diverges from chip purchases or data centre construction. A cancelled data centre is a stranded real estate investment. Deprecated GPUs are e-waste. But renewable capacity becomes part of the grid’s permanent generation mix, available at fully-depreciated cost for whatever electrification demands emerge next.

The pattern holds

Thompson draws on Perez’s insight that speculative mania enables infrastructure investments that rational calculation wouldn’t support. Nobody builds railways for current traffic levels – you build for projected demand a decade hence, accepting that some routes will prove unprofitable. The bubble provides both capital and a coordinating mechanism to build infrastructure ahead of demonstrated demand.

The AI bubble is doing this for renewable energy infrastructure. Tech companies are signing 15- to 25-year PPAs for generation capacity scaled to AI projections that may prove wildly optimistic. They’re providing financial certainty for developers to build utility-scale wind and solar farms that otherwise might not get financed. They’re creating demand for battery storage integration that pushes the technology down cost curves faster than policy could drive.

This timing distinguishes renewable buildout from SMR commitments. Nuclear capacity contracted today arrives after the likely adjustment period. Renewable capacity contracted today reaches operation while the bubble inflates, gets financed through long-term commitments that survive market volatility, and transitions to serving broader demand as the bubble deflates.

The most important infrastructure revolutions remain invisible until complete. We didn’t recognise TCP/IP as revolutionary when networks adopted it. We didn’t celebrate the electrical grid expansion while watching workers string power lines. But both became enabling infrastructure for economic activity impossible to imagine beforehand.

The AI bubble’s lasting contribution might not be the models everyone’s watching but the gigawatts of renewable capacity being built while attention focuses elsewhere. That infrastructure will outlast current market valuations, survive the inevitable correction, and serve demands we can’t yet articulate. It won’t make the bubble’s victims whole or justify today’s equity prices.

But it might prove exactly the kind of infrastructure benefit that makes bubbles worth surviving – just not the one anyone’s actually debating.

Photo by kp yamu Jayanath from Pixabay