When the Bond Market Spoke First: CoreWeave, AI Debt and the Telecom Bust of 2000–2002

In the week of 12 to 16 October 2026, the chip index fell 11.41% and the companies that borrowed to buy the chips fell twice as hard. The PHLX Semiconductor Index had its worst week since April 2025, closing lower eight sessions in a row. A basket of eight US-listed “neocloud” stocks, the firms that borrow to buy Nvidia GPUs and rent them out, lost 22.29% on average. Nvidia itself, the company selling the chips, fell 6.61% (Portfolio Terminal, which compiles its own neocloud basket).

The share prices were not the first warning. By the time CoreWeave’s stock dropped 15.46% on Wednesday 14 October, its bonds had already been telling anyone who looked that something was wrong. Yields on its 2031 and 2032 bonds were near 15%, and the cost of insuring its debt against default had moved above the level that credit analysts call distress.

Meta, one of the biggest spenders on AI infrastructure, reports results after the close on Wednesday 21 October, with Tesla, IBM, Texas Instruments and Lam Research the same evening. So the question readers are asking this week is a fair one: is this a wobble, or the start of a credit problem? This piece does not answer that, because nobody can. What it does is walk through the last time this exact pattern played out, when the telecom and fibre-optic build-out of 1998 to 2002 ended in the largest wave of bankruptcies in American corporate history, and what traders took from it.

What happened in the week of 12 to 16 October 2026

Three blows landed on the AI trade in four days.

Monday 12 October. Broadcom CEO Hock Tan told a conference in Singapore that many investors had little idea what they were actually investing in, according to Bloomberg’s reporting of the event. Chips fell 2.67% that day. Broadcom itself ended the week down 9.07%.

Wednesday 14 October. Gartner published a forecast that more than $3.5 trillion of AI infrastructure spending could be exposed to impairment by the end of 2030 if providers keep spending at the current pace. In the same release it predicted that cost pressure and a lack of a differentiating business model will force over 80% of neocloud providers to cease independent operations by 2030 (Gartner). CoreWeave fell 15.46% to $65.78 and Oracle 8.83%. US CPI, released the same morning, came in at 3.7% for the 12 months to September, with core at 2.6%, both a tenth above expectations (BLS).

Thursday 15 October. Nvidia CEO Jensen Huang, whose company holds roughly a 10% stake in CoreWeave, told Bloomberg that CoreWeave’s financial strength was improving every day and that he saw no need for Nvidia to backstop the company. CoreWeave bounced 5.94%. On Friday it fell 11.91% and finished the week down 25.21%.

The rest of the market did not panic. The S&P 500 lost 1.95% to 7,659.44, the Dow rose 0.67%, and money moved out of technology (XLK −7.70%) and into consumer staples (XLP +5.02%) and health care. Four of 11 sectors fell. This was a sell-off inside one trade, not a market crash, which is itself a clue about what kind of problem it is.

Why the bond market mattered more than the share price

CoreWeave’s shares get the headlines. Its bonds told the story earlier and more precisely.

On 14 October, CNBC reported that the yield on CoreWeave’s 9.75% bond due 2031 was around 14.9%, and on its 9.625% bond due 2032 about 14.8% (CNBC). The next day the yield to maturity on the 2031 bond reached 15.78%, and the five-year credit default swap spread went above 1,140 basis points before settling near 1,038 (CNBC). Gavan Nolan of S&P Global Market Intelligence told CNBC that a spread of 1,000 basis points or more is generally taken as a sign of credit distress.

Read that again. A company whose shares had risen several hundred percent from their 2025 low was, in the credit market’s judgement, priced like a distressed borrower. The two markets were looking at the same company and disagreeing about what it was worth. In October 2026 the credit market won the argument, at least for a week.

Credit default swaps in one paragraph

A credit default swap (CDS) is insurance on a bond. The buyer pays an annual premium, quoted in basis points of the amount insured, and if the borrower defaults the seller makes them whole. A spread of 1,038 basis points means it costs about 10.4% a year to insure CoreWeave’s debt. Reuters calculated that this level implied a roughly 63% chance of default within five years, up from about 30% a year earlier (Reuters via Investing.com). That calculation depends on an assumed recovery rate, so treat it as a rough gauge rather than a precise forecast. The useful point is simpler: when the cost of insurance rises faster than the share price falls, the people who lend money are more worried than the people who own shares. They are usually worth listening to.

The scale of the borrowing

CoreWeave’s second-quarter results showed total debt of $35.6 billion plus operating lease liabilities of $16.3 billion, against revenue of $4.6 billion for the first half of 2026 and a net loss of $1.4 billion (Fortune). CNBC put total debt and obligations above $51 billion. Against that, the company has $103.7 billion of remaining performance obligations, meaning non-cancellable customer contracts, with 41% of it due through 2028.

That is the whole bull case and the whole bear case in two sentences. The contracts are real and large. The debt is also real and large, and it comes due on a schedule that does not care how good the contracts are.

Flashback: the telecom bust, 1998 to 2002

CTE has covered the dot-com bubble and compared the AI boom with the dot-com era in detail. What we have not covered is the part of that period that most resembles October 2026: not the websites, but the pipes.

The build-out

In the late 1990s, the consensus was that internet traffic would grow without limit and that whoever owned the fibre-optic cables carrying it would own the future. Global Crossing, founded in 1997 by former Drexel Burnham Lambert executive Gary Winnick, went public in 1998 and set out to lay about 100,000 miles of undersea and terrestrial cable, spending a reported $13.7 billion on its network. It was one of dozens. According to Morgan Stanley data cited in a March 2002 US House Financial Services Committee memo, Global Crossing and its rivals raised over $1 trillion in equity and debt from 1998 onwards, and that figure excluded bank loans and vendor financing (House Financial Services Committee).

The fibre was real. The internet did grow, enormously. But everyone built at once, traffic grew more slowly than the forecasts, and by 2000 and 2001 bandwidth prices had collapsed. Billions of dollars of cable sat unused, “dark fibre” in the industry’s phrase, and the revenue that was supposed to service the debt never arrived.

Vendor financing: the part that matters most

Here is the mechanism that made the bust spread. The equipment makers, Lucent, Nortel, Cisco, Motorola and Alcatel, did not just sell to the carriers. They lent the carriers the money to buy their equipment. Vendor financing let a start-up carrier with no cash flow order billions of dollars of switches and routers, and let the vendor book the sale as revenue. Paul Starr’s 2002 essay on the telecom implosion describes how this arrangement meant that when the carriers collapsed, the losses flowed straight back to the equipment makers who had financed them (Princeton).

Keep that structure in mind when you read about Nvidia. The company sells CoreWeave its chips, holds about 10% of its equity, and has agreed to buy any capacity CoreWeave cannot sell through April 2032. Nvidia is supplier, investor and customer of last resort at once. That is not identical to Lucent lending money to Global Crossing, and Nvidia’s balance sheet is in a different universe from Lucent’s. But the shape is the same: the shovel seller is financially entangled with the digger.

The collapse

Global Crossing filed for bankruptcy on 28 January 2002, listing assets of $22.44 billion and debts of $12.39 billion, the fourth-largest US bankruptcy at the time. Its creditors included Lucent, Nortel, Cisco and Alcatel. Shareholders who had held stock worth about $55 billion at the peak got nothing (Los Angeles Times).

WorldCom followed on 21 July 2002, listing $107 billion of assets and $41 billion of debt, which made it the largest bankruptcy in US history until Lehman Brothers in 2008. The trigger was the disclosure on 25 June 2002 that it had improperly booked about $3.8 billion of expenses, a figure that later grew to more than $7 billion as the internal audit continued (CNN Money). WorldCom was fraud as well as over-borrowing, which is why it is the less useful parallel. Global Crossing was simply a company that built too much, too fast, with other people’s money.

The total bill is hard to pin down. A Colorado Technology Law Journal paper, citing BusinessWeek, put equity losses at roughly $2 trillion by the end of 2002, about double the losses from the dot-com stocks themselves, with telecom companies losing close to 95% of their total market capitalisation. By March 2003, 24 of the 30 largest public telecom service companies were bankrupt. Estimates in the press at the time ranged from $1 trillion to $4.6 trillion, so “roughly $2 trillion” is the honest way to say it (Colorado Technology Law Journal).

Same pattern, different decade

Where the parallel breaks down

History rhymes. It does not repeat, and a fair piece has to say where the rhyme is weak.

The biggest buyers today are cash-rich. Global Crossing’s customers were other debt-funded carriers. CoreWeave’s customers include Microsoft, OpenAI and other companies with enormous cash flows. Alphabet, Amazon, Meta and Oracle borrowed heavily in 2026, but they borrowed against profits, not against hope.

CoreWeave has contracted revenue. The $103.7 billion of non-cancellable contracts is not the same as the speculative traffic forecasts of 1999. CEO Mike Intrator’s internal memo, obtained by CNBC, said the company was confident its backlog more than supports all of its debt obligations and described the market as acting irrationally. He may be right. The 1999 carriers had no backlog of that kind.

Nvidia is not Lucent. Lucent financed carriers because it had to in order to make sales, and its own balance sheet could not absorb the losses. Nvidia’s backstop of CoreWeave capacity is a small commitment for a company of its size. If CoreWeave failed, Nvidia would take a loss, not a wound.

The chips depreciate faster than fibre did. This cuts the other way. Dark fibre laid in 1999 was still useful a decade later. Gartner’s forecast explicitly rests on the point that AI-optimised infrastructure revenue is too low for most providers to recover their investment before the current generation of semiconductors is replaced. A GPU bought in 2024 may be obsolete before it is paid off, which is a risk the telecom carriers never faced in the same form.

Five lessons for traders

1. Credit often moves before equity

Bond investors are paid a fixed coupon. They do not share in the upside, so they spend all their time thinking about the downside. That makes bond yields and CDS spreads a smoke alarm for the share price. In 2000, telecom bond prices fell months before most equity investors accepted the story was broken. In October 2026, CoreWeave’s bonds were yielding close to 15% before the share price gapped down. If you own a stock whose bonds are being repriced, you are late, not early. (Method.)

2. The shovel seller and the borrower are different trades

Nvidia fell 6.61% in the week. The neoclouds fell 22.29%. Both are “the AI trade” in a headline. They are not the same position. The company that sells the equipment and gets paid in cash sits at a different point in the capital structure from the company that borrows to buy the equipment and hopes to rent it out at a profit. In 2000, the equipment makers suffered too, but the carriers that borrowed went to zero. When you buy a theme, know which end of it you own. (Method.)

3. Real demand is not the same as a profitable business

The internet grew. The fibre got used, eventually, by companies that bought it for cents on the dollar out of bankruptcy. The original investors were right about the technology and wrong about the timing, and in markets, wrong about timing is simply wrong. AI demand can be enormous and a neocloud can still fail if its chips depreciate faster than its contracts pay out. (Mind.)

4. Leverage decides who survives a drawdown, not the quality of the story

This is the lesson of every chapter in Market Mayhem, from the railway mania of the 1840s onwards, and it was the lesson of the 1994 bond massacre too. An unleveraged holder of a good story can wait out a bad year. A leveraged one cannot, because the lender decides when the position closes. Nvidia survived a 90% drawdown in its own history because it owed nobody money; we tell that story in Nvidia: The 90% Drawdown That Built the AI Age. Global Crossing did not survive a bad year because it owed $12 billion. (Money.)

5. Know how to check for leverage risk yourself

You do not need a Bloomberg terminal. For any company inside a theme you own, look at:

• Free cash flow. Is the business generating cash after capital spending, or consuming it? A company spending many times its operating income on capex is funding growth with other people’s money.
• Debt maturities. The annual report lists when debt comes due. A wall of maturities in the next two to three years, in a company that is not yet cash-positive, is a refinancing bet.
• Bond yields. If the company has traded bonds, the yield is public. A yield far above the company’s peers, or above 10% for a company with investment-grade customers, is the credit market telling you something.
• CDS spreads, where available. A move through 1,000 basis points is the conventional distress line.
• Customer concentration. How much of the backlog depends on one or two customers? If one of them cuts spending, what happens to the contracts?
• Vendor financing and backstop deals. If the supplier is also the investor and the buyer of last resort, the supplier’s reported demand is partly demand it created. That was the telecom pattern in 1999. It is worth noticing when it reappears.

The same logic applies to the rising Treasury yields that sit behind all of this. When the risk-free rate is above 5%, a 15% yield on a B+ rated borrower is not as exotic as it would have been in 2021, but it is still a long way from a company the market believes in.

What to watch this week

• Tuesday 20 October, after the close: Netflix. Not an AI story, but the first large-cap result of the week and a read on how the market treats any miss.
• Wednesday 21 October, after the close: Meta, Tesla, IBM, Texas Instruments, Lam Research, SAP. The number to listen for in Meta’s call is capital expenditure guidance. If the hyperscalers keep spending, the neoclouds’ contracts hold. If they signal restraint, the credit market’s view wins.
• CoreWeave’s bond yields. Whether the 2031 bond yield comes back below 13% or stays near 15% will say more about where this is going than the share price does.
• Fed meeting, 27 to 28 October. Futures were pricing a low chance of a hike at this meeting but a high probability by December, according to the CME FedWatch tool as reported by Investopedia on 16 October. Higher rates raise the refinancing cost for every leveraged AI builder.

Frequently asked questions

What is a neocloud company?

A neocloud is a company that buys large numbers of AI chips, mainly Nvidia GPUs, and rents the computing capacity to AI developers and large technology firms. Unlike Amazon, Microsoft or Google, which have other businesses funding their data centres, a neocloud typically does little else and funds its chip purchases with debt. CoreWeave, Nebius, IREN and Applied Digital are examples.

What is a credit default swap, and why does a 1,000 basis point spread matter?

A credit default swap is a contract that pays out if a borrower defaults, in exchange for an annual premium. The premium, or spread, is quoted in basis points. A spread of 1,000 basis points means paying 10% a year to insure the debt, a level that credit analysts conventionally treat as a sign of distress. CoreWeave’s five-year CDS spread went above 1,140 basis points on 15 October 2026.

Is the AI boom the same as the telecom bubble?

No. The structure is similar: a real technology, a race to build infrastructure, heavy borrowing, and suppliers financially entangled with their customers. The differences are also real: the largest AI buyers are highly profitable, CoreWeave has over $100 billion of contracted revenue, and the outcome is not yet known. The lesson from 2000 to 2002 is about reading the warning signs, not about predicting the ending.

What is vendor financing, and why is it a warning sign?

Vendor financing is when the company selling equipment also lends the buyer the money to pay for it. It lets the vendor book sales to customers who could not otherwise afford them, and it means the vendor’s reported demand partly depends on its own lending. In 2000 to 2002, Lucent, Nortel and Cisco all took heavy losses on carriers they had financed. It is a warning sign because it hides how much real, independently funded demand exists.

How can I check if a stock I own carries too much debt?

Read the balance sheet and the debt maturity schedule in the annual report, compare free cash flow with capital spending, and if the company has public bonds, look up their yield. A bond yield far above peers, a wall of debt due in the next two years at a company not yet generating cash, or a CDS spread near 1,000 basis points are the three clearest signals.

This article is educational and is not investment advice. 2026 market data as at 16 October 2026 from the sources linked; weekly index, sector and neocloud figures from Portfolio Terminal’s week in review of 17 October 2026, which compiles its own neocloud basket. Historical figures from the sources linked.