Are we in a data centre bubble?

That dreaded question—are we in a data centre bubble?—has been gaining traction after TD Cowen’s ill-fated report revealed that Microsoft was cancelling several major data centre projects.

Speculative investment bubbles are nothing new. There was the great Beanie Baby boom of the late ’90s, the Tulip Mania in 17th-century Holland, and more recently, the NFT market, which sank faster than the Titanic.

But when it comes to the data centre space, is it really just a case of overbuilding? Or is Tsai right: Has the hype around AI, the main force behind the sector’s expansion, inflated expectations beyond reason? Maybe the real issue isn’t overcapacity but overvaluation.

The argument in favour of a bubble in the data centre market was made clear in late February when news broke that the tech giant had walked away from leases totalling several hundred megawatts.

“DeepSeek of the week” was how Pim Rothweiler from Natixis described it as the news broke at the same time Metro Connect 2025 was taking place.

Some dismissed the analyst report, arguing it was just Microsoft simply getting out of agreements on projects suffering from construction and power connection issues. It was straight out of the playbook from rival Meta, who just a few years ago employed similar tactics to get out of several US data centre leases following its lacklustre attempt to make the metaverse a reality.

But then a second TD Cowen report surfaced, saying Microsoft had scaled back yet more data centre projects, this time in Europe as well as the US. Fears of a potential oversupply position were starting to appear.

Then another, with projects in Indonesia and Australia joining the earlier cancellations, signalling the hyperscaler was possibly rethinking its AI strategy.

In total, Microsoft had planned to invest $80 billion in AI data centres in 2025, with half of that earmarked for the US.

Ed Barrow, CEO and founder of cloud cost management vendor Cloud Capital, told Capacity that “clear signs of froth” were beginning to appear on the hyperscale and infra-investment side.

“The AI boom has driven demand signals that were off the charts, but now the rubber is meeting the road – actual usage, ROI on training runs, and cost-to-serve are starting to matter again. It’s no surprise we’re seeing hyperscalers pause or renegotiate leases as they recalibrate forecasts.”

The Microsoft news alone was enough to spook investors. But it wasn’t just analysts raising eyebrows.

In late March, Alibaba chair Joe Tsai weighed in, warning attendees at the HSBC Global Investment Summit that he was “astounded” by the sums being “thrown around” for AI-focused data centre expansion.

“I start to see the beginning of some kind of bubble,” he said. “I start to get worried when people are building data centres on spec.”

Tsai’s concern wasn’t just about scale but timing. “People are investing ahead of the demand that they’re seeing today, but they are projecting much bigger demand,” he noted, suggesting that enthusiasm around generative AI may be clouding more pragmatic infrastructure forecasts.

His comments came just as Oracle pledged $5 billion to expand its UK cloud footprint, and Nvidia, SoftBank, and Elon Musk’s xAI joined Microsoft in a $100 billion AI infrastructure fund.

Meanwhile, France announced a €109 billion initiative aimed at rivalling the $500 billion Stargate Project from OpenAI, SoftBank and Oracle.

Then there’s Meta, which is reportedly mulling whether to expand its $65 billion data centre plans to $200 billion, and Damac’s Hussain Sajwani’s $20 billion pledge to invest in US data centres.

These are just some of the multitude of investment pledges that have continuously cropped up throughout the last six months. The sheer pace and volume of announcements seem to back Tsai’s core fear: that investors and operators are racing to deploy capital before the market itself is ready.

If Microsoft’s retrenchment signalled a strategic recalibration, the market reaction to CoreWeave’s IPO offered a more visceral litmus test of sentiment—and it wasn’t exactly a confidence booster.

CoreWeave, one of the two AI cloud darlings alongside Nebius, was billed as the first major IPO of the generative AI boom. But despite a 42% bounce on day three of trading, the debut was wobbly, to say the least.

Shares dipped below the $40 IPO price on day two, and the company had already cut its original pricing range ($47–$55) and scaled back its offering from 49 million to 37.5 million shares to drum up enough buying interest.

The Nvidia-aligned infrastructure provider has since seen things level out somewhat, but the initial volatility wasn’t just market jitters; it reflected deeper questions about the company’s fundamentals.

CoreWeave lost $863 million last year, and over 60% of its revenue depends on one customer: Microsoft.

The company is also carrying a multibillion-dollar debt load, with aggressive expansion plans requiring tens of billions more in capex. And while CoreWeave pitches itself as a hyperscale alternative, its S-1 filings reveal razor-thin margins, limited diversification, and significant operational risks.

A potential lifeline could come from another hyperscaler, however, with The Information reporting that Google was in talks to rent AI chips from the firm.

“With CoreWeave’s IPO, it’s tough to draw conclusions given the wild market swings of late,” Jack Backes, principal strategist at Texas-based Provident Data Centers, said. “It was up 40-50% pre-tariffs and still holding 10% above open.

“The efficiencies in chips are going to keep outpacing their costs, and the real winners will be companies with the infrastructure (and power supply!) to serve that demand.”

Backes’ view on CoreWeave’s timing was echoed by Bill Major, the CEO of FiberLight, who said: “If their IPO was simply a few weeks ago, we would have a different perspective on this matter."

Nvidia has long sat at the centre of the AI data centre gold rush, its GPUs forming the backbone of virtually every major hyperscaler and AI startup’s compute stack. But that dominance may not be so eternal.

Recent developments, including DeepSeek’s purported $6 million model training using older, cut-down Nvidia H800s and Microsoft’s new Majorana 1 quantum chip, have shown the sector may be entering into the beginning of hardware diversification.

As Dr Seth Dobrin, IBM’s former Chief AI Officer, pointed out at the recent Metro Connect 2025, we’re approaching a turning point where technologies like photonics, plasmonics, and neuromorphic chips could erode the reliance on GPUs altogether. These new architectures promise faster processing, lower power requirements, and, crucially, freedom from the Nvidia ecosystem’s closed loop of chip supply, software (CUDA), and hardware lock-in.

Startups are already attracting serious investment. Companies like Lightmatter, Conscium, and others are building alternatives that don’t require expensive, bleeding-edge fabs to produce chips, potentially accelerating their path to commercial deployment.

Even before any of these new technologies reach scale, operators are already looking to diversify. AMD’s EPYC CPUs are being used to run AI workloads in place of GPUs, and companies like DataBiome are training models on CPU clusters to avoid bottlenecks and cut costs.

It’s not the end of Nvidia, not by a long shot, but it might be the beginning of a more pluralistic hardware future. And if that happens faster than expected, today’s massive bets on GPU-specific data centre infrastructure could look decidedly shortsighted.

“We haven’t even seen what the next gen foundation models trained on Blackwell+ chips will be capable of,” Backes said. “It will surprise everyone.”

While headlines about cancelled leases and cautious IPOs feed speculation, many insiders argue that what we’re seeing is not a bubble but a moment of strategic correction in a market that’s still fundamentally strong.

“Microsoft pulling back on some leases feels more tactical and strategic, part of their broader AI narrative, rather than a sign of oversupply,” Backes told Capacity. “This checks out with what I’m hearing from other well-informed people.”

Generative AI may have kickstarted the boom, but the next wave, agentic AI, multimodal synthesis, and foundation models trained on Blackwell-class hardware, is still in its infancy. As FiberLight’s Major put it: “As AI continues to evolve from generative to agentic, the demand for larger compute will be realised.”

That shift underpins continued expansion and goes beyond AI hype, with Cloud Capital’s Barrow arguing there was a more nuanced approach toward compute demand.

“Enterprises are moving from ‘experiment at any cost’ to ‘scale responsibly’ which affects how they contract for infra,” Barrow said. “That shift plays right into what we do at Cloud Capital: helping companies model usage, manage commitment risk, and only pay for what they need.”

While plans and projects for new data centres are emerging thick and fast, actually building them sadly isn’t as simple as flipping a switch. Delays due to power, cooling, and skilled labour bottlenecks are already slowing the pace of delivery.

“Skilled labour shortages, grid constraints, and zoning regulations mean that even in high-demand markets, rollout timelines are long,” said Daniele Viappiani, portfolio manager at GC1 Ventures. “That’s not a bubble dynamic—it’s a pacing one.”

Modular data centres, retrofits, and expansion into emerging markets are helping ease these constraints, and the sector is adapting, not collapsing.

“While some fear the bubble is about to pop as demand stabilises due to diminishing returns. The industry is adapting through innovation, ensuring continued scalability,” Viappiani added.

The tech sector has seen its share of froth before, with the dot-com bubble serving as the industry’s go-to reality check. But proponents against the data centre bubble point toward the long-term arc of infrastructure investment tends to bend toward long-term value.

“The internet bubble never truly 'burst.' Because it didn’t, we’ve continued to advance technology, including AI,” Major said. “With any new technology, there are always ebbs and flows, and its lifecycle will include periods of boom, bubble, and bust.

“We have been riding the boom of AI, and while this may be a moment in its evolution, I believe any bubble will ultimately rebound into another boom.”

Some see parallels in today’s AI infrastructure cycle. There may be short-term overbuilds, especially around specific locations or use cases. But over time, most expect demand to catch up and even surpass today’s forecasts.

“No bubble here,” Backes said. “If anything, the market is being overly cautious.”

If there’s one thing many investors potentially didn’t account for, it was tariff-megaddon and the resulting global market plunges not seen since the pandemic.

Investors likely saw the return of Donald Trump as President and expected some level of disruption in the pipeline — but could they have predicted reciprocal tariffs between the US and China as high as 125%?

While the levies are on hiatus after Trump paused them, the threat of economic uncertainty is a spectre that'll eventually reappear once the 90 day reprieve ends.

Major took the view on Microsoft's lease cancellations that the hyperscaler was preparing for the impact of the new US tariffs: “The world and, more importantly, the stock market didn’t like the tariffs that were imposed on other countries. Microsoft, with a nearly $3 trillion market cap, has seen its stock fluctuate by 20% from their 52-week high and low.

“Now that the tariffs have been put in place, the markets are reacting, and it forced their management to pivot and be good stewards of their working capital. It doesn’t mean they no longer need the facilities, but simply a pause to see the longer-term impact of the US tariffs."

Some tariff concerns arose at the turn of the year following President Trump’s initial wave of tariffs, a 25% levy on steel imports, which could affect project construction. Though the steel market was already in flux prior, what with the conflict in Ukraine forcing firms to source steel from China and Sweden all while Saudi Arabia is snapping up 20% of the world’s entire supply to build it’s to build its One Line city in the desert.

Beyond steel now, though products that are either manufactured in China or rely on components from targeted countries are expected to face increased pressure. This will likely heavily affect IT equipment such as rack and blade servers, routers, switches, and even passive components like resistors and capacitors, as well as batteries, solar panels, IoT sensors, cement and drywall, and fire suppression systems.

It might not affect everyone, though. Earlier this year, Sebastian Sassi, VP of sales at Atlantic Vision, which supplies FTTH and data centre fibre optics cables, made the bold claim that even if tariffs were as high as 200%, operators would be better off using their services.

Despite tariffs edging closer to that figure, with 125% the number hitting China at the time of writing, Sassi reported a 20% increase in sales inquiries in the first quarter of 2025 than the same time last year.

“The black swan event we’re seeing right now is going to induce some short-term uncertainty, but much like the dot-com bubble bursting wasn’t the end of the internet and e-commerce, a short-term price correction in equities, markets, and capex conservation by Microsoft won’t mean AI goes away,” he said.

There’s no denying that we’re in a high-stakes phase of the AI infrastructure cycle. Capital is flowing fast, demand curves are shifting, and uncertainty, fueled by tariffs, tech transitions, and market fatigue, is rising.

But to call it a bubble implies an imminent burst. Joe Tsai seems certain, but most insiders aren’t so sure.

Instead, what we’re likely to see is a moment of recalibration where AI’s potential remains vast yet untapped, and the rush to build must now contend with economic, political, and technological realities.

If history is any guide, short-term corrections won’t cause the data centre market to go pop. They may just make the next chapter more sustainable. Whether the market is overheated or simply adjusting its stride, one thing is clear: the race to build the future isn’t over. It’s just getting smarter.

Microsoft hits pause on more data centre builds amid AI rethink

How investors are meeting the continued demand for capital in data centres

Trump's tariffs threaten to upend the data centre boom with rising costs

Steel shock: How Trump’s 25% tariffs could delay data centres and tower builds