Lessons from the CLEC Bust: Why we prefer Hyperscalers over Neoclouds

Recent market movements were interesting. On 12 October, Broadcom's CEO said that capital will be the biggest constraint facing the ​AI buildout as development continues. Not demand. Those comments added to selling pressure to the sector. In particular, neoclouds – represented by companies such as CoreWeave, IREN and, Nebius – were some of the largest detractors.

This was in stark contrast to the period from 2025 to early this year, when the market was very optimistic about Neoclouds. The reason is simple: the demand for AI computing power is stronger than ever before, and the supply of computing power will face shortages over the next few years. Recent developments continue to support the case for strong demand: CoreWeave raised prices by about 25% across all its GPU types in July 2026, and Nebius raised its on-demand H100 rate from $3.85 to $4.50 per GPU-hour from 1 October 2026, an increase of about 17%.

So, why do neoclouds continue to face so much selling pressure despite such strong demand? As we highlighted in both of our previous articles – Digital Economy (Internet) 4Q26 outlook: Sailing through the tides; Maintain 3.5 Stars Cloud Computing Update: Increasingly diverging  strong demand alone does not establish the investment case.

Figure 1: Neocloud vs Big tech performance since early September

Demand is not the only scoreboard

Neoclouds are unlikely to fail for lack of AI demand. If they fail, it will be because the capital markets stop funding them. That was the lesson of the competitive local exchange carriers (CLECs) of the late 1990s, and in October 2026 the bond market began applying it.

The bull case for CoreWeave, Nebius and IREN is simple: demand for AI compute exceeds supply and will for years. That is probably true. But over the past 12 months the three spent a combined US$34.7bn on capex, and interest-bearing debt now equals about 42% of their combined assets. For businesses built this way, the health metric is not demand alone, but whether contracted revenues generate sufficient cash returns to service debt and fund future investment. The cost and availability of the next dollar are therefore critical.

Figure 2: Neocloud backlog

Figure 3: Net Debt/ EBITDA

How the CLECs failed

The CLECs were the neoclouds of the dot-com era. After the 1996 Telecom Act opened US local phone markets, they borrowed heavily to build local networks. They pledged the equipment as collateral, raised more money, built more, and signed more customers. Equipment makers such as Cisco, Lucent and Nortel lent to them so they could buy their gear.

The long-term demand trend was real, but it did not prevent excess capacity and weak economics during the buildout. Growing internet usage ultimately benefited consumers without ensuring that the companies financing the infrastructure survived. Global internet traffic has grown more than 1,000-fold since 2000, and almost every household is now connected. Yet 14 of the 15 major CLECs filed for bankruptcy between 2000 and 2003.

The trigger was funding. The Fed raised rates six times between June 1999 and May 2000, from 4.75% to 6.5%. Once capital markets tightened, the most highly leveraged CLECs were the first to go bankrupt, which made investors increasingly anxious about the CLEC business and unwilling to continue lending them money. Although the Federal Reserve subsequently cut interest rates 11 times in 2001, lowering the rate from 6.5% to 1.75%, it was still to no avail. One bankrupt CLEC listed the near-total closure of telecom capital markets as the first cause of its failure.

The case of Neocloud; Nvidia as the New Cisco?

Neoclouds run the same loop with different hardware. Neoclouds' business model consists of purchasing GPUs from Nvidia, placing them into data centers, and then renting out the computing power to AI model companies, large cloud providers, or other Neoclouds. Unlike cloud giants such as Microsoft, Amazon, and Google, whose established businesses generate substantial cash flow, Neoclouds rely heavily on external capital to finance GPU purchases and expansion.

Their capital intensity and credit risk can make unsecured borrowing expensive. Contract-backed financing can reduce that cost by giving lenders greater visibility over future cash flows. Therefore, Neoclouds' strategy is to sign contracts with clients first, and then use these contracts as collateral to secure financing. For example, the interest rate for a contract-backed loan secured by CoreWeave in March of this year was SOFR plus 2.25%; meanwhile, the yield on its unsecured bonds had already reached around 15% by October.

To explain the Neoclouds' business model simply: sign contracts with clients, use the contracts as collateral for financing, use the financing to build more data centers, secure more contracts, and repeat.

This resembles the financing loop used by CLEC business model during the internet era. Back then, tech giant Cisco lent money to its CLEC clients so they could purchase Cisco network equipment. After building the network infrastructure, the CLECs used that equipment as collateral to borrow money, continuing the cycle to build out further. Today, Nvidia is doing the exact same thing.

•         It agreed to buy CoreWeave's unsold capacity through 2032, under a deal initially worth US$6.3bn.

•         It invested US$2bn of equity each in CoreWeave and Nebius.

•         In August 2026 it provided residual-value guarantees up to US$105bn of leases for OpenAI-linked data centres development, with payment obligations contingent on specified tenant default events. The site is intended to host Nvidia infrastructure exclusively.

Vendor financing on this scale is a signal. It does not cause a bust, but it makes it harder to tell how much demand would exist without the supplier's balance sheet behind it.

Figure 4: AI financing chain

Source: IDC. Data as of June 2026.

This is what the bond market is pricing in

Without external funding, this business model is unsustainable. Neoclouds cannot support expansion through their own profits. None of the three companies—CoreWeave, Nebius, and IREN—generates enough operating profit to cover their interest expenses. Over the past 12 months, CoreWeave's operating cash flow was $6.9 billion, while its capital expenditures were $20.6 billion, leaving a US$13.7bn gap before other investing and financing flows. Funding that gap requires external capital or existing liquidity.

But issuing debt creates even more problems, starting with financing costs. By 13th October, the yield on CoreWeave's unsecured debt had risen from 9.5% at issuance to 15%, which is 10 percentage points higher than the 5% risk-free yield of US Treasuries. As interest expenses continue to mount, profits will be further squeezed, necessitating even more borrowing to sustain massive capital expenditures. Its 2031 notes briefly topped 16%, and five-year CDS implied a default probability above 60%.

Bear in mind, higher secondary-market yields do not immediately increase the coupon on existing fixed-rate debt, but they indicate a higher cost of issuing new debt or refinancing outstanding obligations.

CoreWeave now has a two-tier cost of capital. Contract-backed project debt is still cheap: a Meta-backed tranche priced at 5.9% in July, against roughly 10% on its unsecured bonds at the time. That is why management says lenders are underwriting Microsoft or Meta, not CoreWeave.

Project-level financing does not remove the need to fund development costs, overruns and equity contributions. Higher unsecured borrowing costs can therefore constrain expansion even when contract-backed debt remains available. The relevant test is whether project cash returns, after operating costs and equipment replacement, exceed the blended cost of financing.

At 15% yield, earning a positive spread on GPU rental is very hard. Our take is that these Neoclouds could increasingly use customised chips from the hyperscalers, such as TPUs, to lower costs, as they do not have the capability to build their own chips; however, Nvidia’s financing may limit adoption.

More importantly, external funding is not infinite.

Figure 5: Yields

Figure 6: CDS for Neocloud are above the roofs

 No pricing power at either end

Recent news, such as CoreWeave raising prices by about 25% across all its GPU types in July 2026, and Nebius raising its on-demand H100 rate from $3.85 to $4.50 per GPU-hour from 1 October 2026, an increase of about 17%, may create the impression that these companies have pricing power. Our take is that this impression is mistaken.

They are all selling GPU computing capacity built largely on Nvidia chips. While service quality, software and deployment capabilities can differ, it remains uncertain whether those differences are sufficient to sustain premium pricing once supply becomes more abundant. Because they lack pricing power, the profit margins for Neoclouds are actually dictated by upstream suppliers and downstream clients. Today’s pricing power appears to depend heavily on scarcity, rather than on advantages that competitors cannot readily replicate.

Nvidia sets their costs: IREN's annual report warns that suppliers may raise GPU prices even on orders already placed, and that it cannot always pass the increase on. Hyperscalers set their revenue: Nebius warns that larger rivals with broader product lines may compete by cutting prices or bundling. Those rivals are also the neoclouds' biggest customers.

The CLECs were squeezed the same way, between Cisco's equipment pricing and the Bell companies' hold on the customer. When sentiment turned in 2000, both ends tightened at once, which is why CLECs were among the first casualties of the bust.

Coming back to the title, today’s pricing power comes from scarcity, not from anything the neoclouds own that others cannot copy.

Figure 7: Neocloud’s margin

 Source: Bloomberg Finance L.P., iFAST compilations. Data as of 27 October 2026.

But, to be fair, there are also some differences between CLECs and today’s Neoclouds

Despite the similarities between Neoclouds and CLECs, one must admit that the customer quality for Neoclouds is far superior to that of the CLECs back then. CLECs sold to small businesses and dot-coms, many of which failed alongside them. Neoclouds sell to Microsoft, Meta and the frontier labs on multi-year contracts. IREN has a roughly US$9.7bn contract with Microsoft, and Nebius has five-year deals with Microsoft and Meta. That revenue is far safer than any CLEC's. But a strong customer only guarantees that the contract gets paid; it does not make the business profitable, and it also means that customers have more bargaining power over them.

The CLEC collapse was not only about funding. Many depended on leasing the Bell companies' networks at regulated rates, and court rulings in 2004 removed that access. Neoclouds have no comparable dependence on a regulator.

Key Takeaway

Demand is necessary for neoclouds to survive, but it is not sufficient. The CLECs were right about the internet and were still wiped out by their funding structure.

Citadel Securities founder Ken Griffin airline analogy highlights the risk: once capacity exceeds demand, operators may accept prices that cover incremental operating costs but fail to recover their original investment.

Once these massive clients finish building more of their own data center capacity, their reliance on Neoclouds could diminish, and they may use their scale and product offerings to compete more aggressively.

Hyperscalers also face rising funding needs as AI investment absorbs cash flow, but their diversified earnings and stronger access to capital give them greater flexibility to manage those commitments. That is why we prefer hyperscalers over neoclouds. AI demand can remain structural while the returns earned from supplying compute remain cyclical. 

Our preference on Hyperscalers is not a sell call on Neoclouds. Over the near term, Neoclouds can still benefit from compute scarcity. However, a more constructive view would require evidence of:

•         Narrower credit spreads and sustained access to financing on terms that support viable project returns. Lower Treasury yields alone would not resolve company-specific credit concerns

•         Resilient renewal pricing and utilisation, including secondary rental rates for older GPUs.

•         Improving cash generation relative to CAPEX and debt-servicing requirements.