Corrected: RPT-BREAKINGVIEWS-Data centre upstarts evoke earlier broadband craze

Nebius NV

11/09/2026 13:00

Corrects IREN CEO's first name in paragraph eight. The author is a Reuters Breakingviews columnist. The opinions expressed are his own.
By Edward Chancellor
LONDON, Sept 11 (Reuters Breakingviews) - The data centres that run AI models can cost tens of billions of dollars to build. Most of this infrastructure is financed by deep-pocketed tech giants such as Microsoft MSFT.O and Amazon.com AMZN.O. Yet independent operators, known as neoclouds, are also playing an important role in the AI build out. Daniel Roberts, the co-CEO of data centre operator IREN IREN.O, claims the surge in capital spending is “fundamentally different” from past investment cycles. In fact, the AI is frenzy is reminiscent of the internet bubble; the neoclouds resemble the alternative telecoms firms that soared and crashed as the millennium dawned. 
By 2002, the telecoms industry in North America had spent nearly $500 billion over the preceding years building the broadband networks that were required in the internet era. Around half of this investment was made by alternative telecoms firms: upstarts taking on incumbent industry giants. Some of these so-called altnets built local telecoms systems to compete with the established “Baby Bells”; others invested in long-haul fibre-optic networks. Around 40 of the altnets were listed on U.S. stock exchanges.
“Build it and they will come” was the industry’s boomtime mantra. The demand for broadband was deemed insatiable. Tech investment guru George Gilder forecast in early 2000 that data traffic would increase a thousandfold within five years. Enthusiasts believed that every dollar invested in new telecoms equipment would generate an incremental dollar of sales. The imminent arrival of new killer applications – one altnet CEO mooted holographic 3D video calls – would utilise the millions of miles of fibreoptic cable being laid.
Between 1995 and 2000, the assets of the U.S. long-haul fibre operators – among them Global Crossing, Level 3 Communications and Williams Communications – expanded more than twenty-fold. Liabilities grew in tandem. Their businesses, however, remained stubbornly unprofitable. Projected demand for broadband turned out to be grossly exaggerated. As a result, the telecoms businesses were saddled with massive excess capacity.
When the capital markets closed to telecom and technology companies in mid-2000, the altnets’ share prices collapsed, and several high-profile firms went bust. Some long-distance operators, including Global Crossing, faced scrutiny for using bogus capacity swaps to inflate revenue. Telecoms equipment firms like Lucent and Cisco Systems CSCO.O that had lent to the altnets were forced to write off their loans. Around 40% of high-yield telecoms bonds defaulted.
Despite their public protestations of confidence, many telecom insiders bailed out long before the crisis arrived. Gary Winnick, the CEO of Global Crossing, raised more from sales of personal stock than his firm’s assets sold for in bankruptcy.
While the altnets were mostly startups, many of today’s neoclouds were established to mine cryptocurrencies. CoreWeave, the largest of the group, started out in 2017 as Atlantic Crypto and entered the data centre business after the bitcoin crash of 2018. After the crypto market nosedived again in 2022, the launch of OpenAI’s ChatGPT chatbot in November of that year offered other crypto miners a lifeline. Their stocks of graphic processing units (GPUs) and contracted power agreements were indispensable for training and running the large language models that were now much in demand.
While so-called hyperscalers such as Alphabet GOOGL.O, Microsoft, Amazon, Meta Platforms META.O and Oracle ORCL.N account for around half of the sales leading AI chips made by Nvidia NVDA.O, neoclouds make up a sizeable share of the rest. The AI boom is “leading to the largest infrastructure buildout in human history,” declares Joshua Payne, the boss of Nscale, a London-based neocloud firm, which is spending tens of billions of dollars building a data centre in West Virginia. IREN’s Daniel Roberts believes that each unit of new AI supply is bringing on multiple amounts of new demand. The “metric that defines success” in his business, says Roberts, is the number of GPUs you control.
Nvidia’s chips are very expensive and large language models that are scaling up in pursuit of artificial general intelligence require them in ever greater quantities. That has left the neoclouds scrambling for cash. Several privately owned firms, including Nscale, Lambda, and Softbank-backed SB Energy, are preparing stock market listings. CoreWeave CRWV.O CEO Michael Intrator told CNBC earlier this year that the debt markets were dictating his company’s expansion strategy. At the end of June, CoreWeave’s total liabilities stood at $72 billion, an increase of 56% since the start of the year.
Whereas the altnets financed themselves with high yield, payment-in-kind and zero-coupon bonds, neoclouds have been issuing bonds backed by GPUs and setting up special purpose vehicles with financial partners. Just as telecom equipment firms provided vendor-financing for their customers in the late 1990s, Nvidia plays a key role in supporting the neoclouds. The AI chip giant has invested directly in CoreWeave, Nscale, Nebius NBIS.O and others. Nvidia has committed to lease data centre capacity and provides a financial backstop for some neoclouds. It also has revenue-sharing agreements with a number of them.
The investor Jim Chanos believes the neocloud business model is no more viable than the altnets of yesteryear. The famed short-seller argues their businesses don’t scale and lack any clear competitive advantages. Neoclouds are capital-intensive businesses that, in his financial model, cannot earn their cost of capital even when the depreciation schedule for their GPU chips is stretched out to 10 years. They should be viewed as equipment leasors that are dependent on a single vendor, Nvidia. Current bottlenecks restricting the construction of new data centres are unlikely to last. Chanos sees the sale of shares by CoreWeave insiders as a red flag.
The alternative telecoms firms did not fail simply because they oversupplied a market whose demand projections were exaggerated. The industry was stricken with excess capacity in part because technological improvements massively increased the carrying capacity of installed fibre-optic networks. Neoclouds are also at the mercy of uncertain technological developments.
Can Nvidia hold onto its position as the dominant maker of AI chips? Will OpenAI and its rival Anthropic – both of which are prime neocloud tenants - continue to dominate their field? Will the arrival of small language models that can run on local computers obviate the large language models and their expensive data centres? It is this profound uncertainty regarding the future of AI technology, rather than the level of capital spending, that makes the current boom “fundamentally different” from anything the world has seen before.
Follow @Breakingviews on X
     

(Editing by Peter Thal Larsen; Production by Streisand Neto)
((For previous columns by the author, Reuters customers can click on CHANCELLO/edward.chancellor.bv@gmail.com))