Could AI Repeat
the Fiber-Optic Bust
of 2000?
Three lessons from the fiber-optic boom, revisited through Big Tech cash flow and today's AI infrastructure buildout
THE QUESTIONFor the first time, combined capex across four Big Tech companies has exceeded their operating cash flow. Even if AI adoption keeps rising, could investors still lose money as they did in 2000?
Snowshagal Research5-minute briefing · About 20 minutes for the full report
The four companies are Microsoft, Alphabet, Amazon and Meta. We define free cash flow consistently as net cash from operations less purchases of property and equipment, using each company's quarterly cash-flow statement. These figures cover the entire companies, not their AI operations alone.
IN THIS REPORT
- 01For the first time, cash left after investment turns negativeCapex +108%; operating cash +10%; long-term debt triples
- 02How the fiber-optic boom unraveled in 2000Hype → investment → oversupply → price collapse → debt
- 03Three questions from the fiber-optic bustContracts, credit and payback: 2000 versus 2026
- 04The weakest link is the customer's creditVendor financing then; guarantees and leases now
- 05Efficiency echoes 2000; power is the constraintTechnology expands capacity, while the grid constrains it
- 06Counterarguments and four possible outcomesPayback, a slowdown without defaults, a credit event, or a 2000-style bust
- 07In 2000, equipment suppliers fell even harderWhat Korean supply-chain investors should monitor
- EEarly-warning dashboardThresholds versus current readings; what is known and unknown
PublishedNovember 5, 2026 (Thursday)
Information cutoffEvening of Nov. 4 (KST) · Hana Securities' Nov. 5 report · bond yields and spreads, Coherent results, and Meta/Alphabet bond issuance updated through the latest prepublication releases
SourcesFor the 2000 episode: contemporary Federal Reserve, FCC and SEC records and academic research. For Big Tech: company quarterly filings. Broker estimates and interpretations are attributed to the firms that made them.
Not yet includedCoreWeave earnings (Nov. 11), Nvidia earnings (Nov. 18), and any Anthropic IPO
INTRODUCTION
On October 30, the Nasdaq fell 2.86% and the semiconductor index dropped 7.2% in a single U.S. session. Reuters cited fears that AI investment was burning through cash as a driver of the selloff. Intel's 28.6% plunge added to the pressure.1 In the trading sessions that followed, the market repeatedly returned to the same question.
- Big Tech Q3 resultsAlphabet, Amazon and Meta each spent more on purchases of property and equipment than they generated in operating cash flow.−$26.4bnFour-company combined FCF
- Debt and equity issuanceMicrosoft's $25bn bond offering · Oracle's $20bn in equity and bonds · Meta's $15bn dollar bond sale and €9bn euro bond debut2, 3, 4€53bnOrders for Meta's €9bn euro bond sale
- KOSPI swings sharplyUp 3.17% early, down 4.02% intraday amid U.S. midterm election counting and data-center regulatory concerns, then closing up 0.62% at 5,683.805−4.02%Intraday low
- AI suppliers report resultsYear-on-year revenue growth: AMD data center +97%, Arista +58.5%, Coherent +50%6, 7, 8+97%AMD data-center revenue
Strong sales at AI suppliers do not establish that their customers will earn an adequate return on investment. That is why the fiber-optic boom of 2000 is so often invoked.
- Back then Telecom operators borrowed heavily to lay fiber on the assumption that internet usage would grow without limit.
- Two years later Roughly $2 trillion in telecom market capitalization and 500,000 jobs had disappeared, according to testimony by the chair of the U.S. Federal Communications Commission (FCC).9
- Yet Internet usage kept growing. The losses fell on network builders and their investors.
This report therefore separates three distinct outcomes.
We turn three conditions behind the 2000 bust, identified by Kiwoom Securities, into questions to ask of both eras, then finish with a monitoring checklist.10
AT A GLANCE · FIVE-MINUTE BRIEFING
- 1The fiber-optic bust did not happen because internet traffic stopped growing.Internet traffic grew roughly twofold each year. But capacity grew even faster: wholesale circuit prices fell more than 90% in two to three years, and operators financed by junk bonds were among the first to fail.
- 2In 2000, the collapse began at the fringes.Back then, the first casualties were new entrants funded through IPOs and high-yield bonds.11Today's vulnerable fringe consists of GPU-rental providers, known as neoclouds. Yields on CoreWeave's unsecured bonds reached 15–16% in October.
- 3Negative Big Tech FCF is a warning about the pace of investment.Capex rose 108% year on year, against just 10% growth in operating cash flow. Combined Q3 FCF was −$26.4bn, although trailing-12-month FCF remained positive at $61.5bn. Debt-servicing capacity is not the same as investment returns.
- 4The weakest link is the end customer's credit.In 2000, Lucent committed $8.1bn in financing to its customers and absorbed the resulting losses. Today, the exposure is distributed across guarantees, leases and joint ventures. Some end customers are AI labs with usage-based revenue but multiyear payment commitments.
- 5Some warnings are visible; important exposures remain unobservable.Of the eight indicators, neocloud bond yields have triggered a warning, while Big Tech FCF is close to its threshold. Three indicators lack published industry-wide data. The next checkpoints are CoreWeave earnings on November 11 and Nvidia earnings on November 18.
Cash left after investment
turns negative for the first time
In Q3, three of the four Big Tech companies spent more on equipment than they generated in operating cash. Demand had not turned down; investment was simply accelerating faster than cash generation. Unlike the new telecom entrants of 2000, these companies still fund most of their spending internally.
Microsoft, Alphabet, Amazon and Meta. Operating cash flow minus purchases of property and equipment; USD 100m.
Derived by converting cumulative amounts in each company's 10-Q and 10-K cash-flow statements into quarterly figures, then summing them. Microsoft's July–September quarter (FY2027 Q1) is aligned with calendar Q3. Source: SEC EDGAR filings.12
Combined Q3 free cash flow for the four companies was −$26.4bn—the first negative quarterly reading since at least 2017 on comparable data, using the same definition as Eugene Investment & Securities.13All four companies still produced positive operating cash flow. The deficit emerged only after deducting capital spending.
TRENDThe first quarter with capex above operating cash flow
Combined four-company figures in USD 100m. Capex rose 108% year on year, versus 10% growth in operating cash flow.
Company 10-Q and 10-K cash-flow statements. Capex is purchases of property and equipment, before proceeds from disposals for Amazon. Assets acquired through finance leases are excluded.12
The quarter was negative, but trailing-12-month FCF is still +$61.5bn. The direction, however, is what matters.
- Ratio Capex as a share of operating cash rose from 45% in Q1 2024 to 61% in Q3 2025 and 115% in Q3 2026.
- 2026 plans Microsoft approximately $170bn (calendar year), Alphabet $195–205bn, Amazon around $220bn, and Meta $135–145bn (including finance-lease principal).14, 15, 16, 17Companies include finance leases on different bases, so these forecasts are not directly comparable to purchases of property, plant and equipment in their cash-flow statements.
- 2027 Alphabet said spending would rise 'significantly'; Amazon said it would invest 'meaningfully more'; Meta said only that it would continue to invest 'aggressively'. Quantified guidance is expected with Q4 results in January–February 2027.15, 16, 17
- Scale U.S.-listed telecom service companies collectively spent $121bn on capex in 2000.18The four companies now spend roughly six times that amount in nominal dollars. The difference narrows when adjusted for inflation and the different geographic scope.
The combined figure captures the industry's investment pace, but the four companies do not share a balance sheet. Excluding Amazon, the other three generated roughly +$8.4bn of FCF in Q3.
Company-by-company detailQuarterly FCF for each of the four companies and three definitions of FCF for the same quarter
Q1 2024–Q3 2026; USD 100m. All four panels share one scale. Navy indicates positive; red, negative.
MicrosoftQ3 2026 +$18.6bn
AlphabetQ3 2026 −$9.1bn
AmazonQ3 2026 −$34.8bn
MetaQ3 2026 −$1.2bn
Company 10-Q and 10-K cash-flow statements; operating cash flow less PP&E purchases. Alphabet's Q3 includes a $5.2bn payment of EU fines and interest. Under the companies' own definitions, Amazon reported approximately −$33.5bn and Meta −$2.1bn for Q3.12, 15, 16, 17
RECONCILING THE FIGURES−$26.4bn, −$26.0bn and −$29.9bn describe the same quarter
- −$26.4bn Operating cash flow less gross PP&E purchases—the definition used in the Eugene Investment & Securities calculation.
- Approximately −$26.0bn Sum using each company's official FCF definition (Amazon nets asset-sale proceeds and incentives; Meta deducts finance-lease principal).
- Approximately −$29.9bn Deducting finance-lease principal repayments (approximately $3.5bn) for all four companies.
All three measures are negative. But assets acquired via finance leases—$7.4bn for Microsoft alone in Q3—do not immediately appear as cash capex on the cash-flow statement. A cash-only measure can therefore understate the economic scale of the investment.14
FUNDINGLong-term debt has tripled in one year
Long-term debt across the four companies tripled in a year. Microsoft alone reduced its balance, while Amazon, the largest debtor, reported trailing-12-month FCF of −$42.4bn.16Assessing repayment capacity also requires looking at cash holdings and debt maturing within the next 12 months.
Long-term borrowings on the balance sheet (excluding current maturities), as of September 30 for each company.
Sources: company 10-Q balance sheets. Microsoft's $25bn bond issue, which settled November 4, is not included.12
A series of financing announcements accompanied the Q3 earnings releases.
| Company | Financing |
|---|---|
| Microsoft | $25bn bond offering2Priced October 29 · Ten tranches maturing 2028–2066 · Aaa/AAA |
| Oracle | $10bn in bonds and $10bn in equity3November 2 · Bond coupons of 7.00–8.65% |
| Alphabet | €6bn euro-denominated bond sale19Placed Nov. 2 · Seven tranches · More than €17bn in investor orders |
| Meta | $15bn in dollar bonds · €9bn in euro bonds4Nov. 3 dollar deal: five tranches maturing 2031–2056 · Nov. 4 debut euro deal: record €53bn in orders |
Sources: company disclosures and reporting by Bloomberg and Reuters.
- OracleSeparately from the $20bn equity raise it had already completed, Oracle said on October 28 that it would cap additional FY2027–FY2030 financing at $20bn, excluding refinancing. It projected a return to positive FCF in the second half of FY2028.20
- Daishin SecuritiesMeta and Alphabet debt-funding plans, alongside disclosures of Anthropic’s multiyear commitments, renewed concerns over how AI investment will be financed (Nov. 3).21
- Investor demandDespite those concerns, large technology issuers drew heavy demand. Meta attracted a record €53bn in orders for its €9bn debut euro bond sale; Alphabet drew more than €17bn for its €6bn offering. The primary bond market remains open to megacap borrowers.4, 19
- iM SecuritiesInvestors are not rejecting AI investment, but they are less willing to wait. Rising financing costs are shortening that window of patience.22
Brokerages generally expect the peak in capex growth to be near. Most still forecast a period of cash shortfalls.
- Capex growth For the four companies: 107% in Q3 to a projected 52% in Q2 2027 (Yuanta Securities). U.S. technology capex growth slows to a projected 14% in Q4 2027 (Hana Securities).23, 24
- Free cash flow Projected four-company FCF of −$44.1bn in H1 2027 (Yuanta Securities)23 · In FY2027, consensus projections show negative FCF at every company except Microsoft (Bloomberg, cited by Hana Securities; company-level figures appear in Table 2).25
Even if capex growth slows, the absolute spending level can leave cash flow constrained.
Market-consensus estimatesProjected FCF for six companies, years of expected recovery, and capex relative to operating cash flow
USD 100m. FY denotes each company's fiscal year (Microsoft's ends in June; Oracle's in May). Since reporting periods differ, the simple totals are not industry FCF for a common period and are not comparable with the quarterly four-company total in Figure 1. Read the table for each company's trajectory and projected recovery date, not the aggregate.
| Company | FY | FY | FCF turns positive | Capex / operating cash FY |
|---|---|---|---|---|
| Microsoft | +670 | +307 | Stays positive | 63% |
| Alphabet | −180 | −531 | 2028 | 110% |
| Amazon | −427 | −548 | 2028 | 122% |
| Meta | +26 | −394 | 2029 | 98% |
| Oracle | −237 | −347 | 2029 | 174% |
| CoreWeave | −269 | −312 | 2030 | 416% |
| Four-company sum | +89 | −1,166 | — | — |
| Six-company sum | −417 | −1,825 | — | — |
Bloomberg consensus estimates cited by Hana Securities. The projected positive-turn dates refer to annual FCF. Oracle's own projection of positive FCF in H2 FY2028 covers a different period and comes from a different source. Meta is projected to post negative FCF in FY2027–FY2028 before turning positive in FY2029. The four-company and FY2027 totals were calculated by summing the table entries.25
How did the fiber-optic boom
unravel in 2000?
Internet traffic kept increasing. The problem was that demand expectations overshot reality by roughly sixfold, while supply expanded faster than actual usage.
First, an important distinction: usage means the volume of internet traffic, while orders refer to telecom operators' spending on equipment. Traffic kept rising, but equipment orders fell sharply from 2001 onward. It is those orders that matter most to Korean suppliers (Section 07).
STAGE 1 · EXPECTATIONS'Internet traffic doubles every 100 days'
- April 1998 A U.S. Department of Commerce report repeated an estimate from UUNET, a WorldCom subsidiary: 'Internet traffic doubles every 100 days.'26
- October 1998 Research by Coffman and Odlyzko measured annual growth closer to 100%. The truly explosive increase had occurred in 1995–1996.27
- July 2000 An FCC staff study traced the '100 days' figure to a single industry white paper.28
- September 2002 The president of the Federal Reserve Bank of Boston concluded that demand forecasts had been off by roughly a factor of six.29
Doubling every 100 days implies roughly 12.6× annual growth. Actual traffic grew about 2×, while the maximum carrying capacity of a single fiber strand grew faster still—roughly 2.5× annually.
The per-strand figure is the technical maximum, not necessarily capacity lit and sold. Installed supply also reflected roughly 2.2× route-length expansion over four years and more strands per route. The projection annualizes the Department of Commerce claim (2 to the power of 3.65); observed traffic comes from Coffman and Odlyzko, and maximum strand capacity from the Richmond Fed (2.5Gbps in 1996 to as much as 100Gbps in 2000).18, 26, 27, 30
STAGE 2 · INVESTMENTCapacity doubled in five years; twenty long-haul carriers in one city
Telecommunications equipment within private fixed investment, nominal USD 100m. Spending fell 33%, from $123.3bn in 2000 to $82.2bn in 2003.
U.S. Bureau of Economic Analysis (BEA) national accounts. The current revised series differs slightly from data published at the time.31
- Carrier capex Listed U.S. telecom service providers' capex peaked at $121bn in 2000 and fell by more than half to $49bn in 2002.18
- Route mileageU.S. long-haul fiber routes grew from 143,000 miles in 1996 to 320,000 miles in 2000. Chicago and Minneapolis each had twenty long-haul operators, compared with about five per city in 1990.18
- Risks disclosed by companies themselves Level 3 warned in its 1999 filings that it expected to incur substantial capex before meaningful revenue and that competitors' capacity could materially depress prices.32
STAGE 3 · CAPACITY SURGECapacity increased without digging another trench
Dense wavelength-division multiplexing (DWDM) sent multiple wavelengths of light through a single fiber. Upgrading the equipment at each end could multiply the capacity of cables already in the ground, while 30operators raced to lay new routes. The FCC chairman cited estimates that network capacity had grown by as much as a factor of 500.9
The Richmond Fed identified three causes of the long-haul glut: excessive entry early in the industry's development; larger-than-expected gains in per-strand capacity; and, most importantly, demand that fell short of expectations.30 Supply exceeded expectations even as demand failed to meet them.
STAGE 4 · PRICE COLLAPSEMore than 90% in two to three years
- By July 2000 The Band-X bandwidth-exchange price index stood about 63% below its October 1998 level. The researcher cautioned that the index was not representative of the entire market.28
- By mid-2002TeleGeography estimated that wholesale circuit prices had fallen more than 90% on most routes within two to three years. In October 2001, some trans-Pacific cable capacity was offered for resale at prices far below the estimated cost of construction.33
- The regulator's explanation The FCC chair testified that a brutal price war had driven down industry revenues and those of individual firms.9
STAGE 5 · DEBT AND BANKRUPTCYJunk-bond-funded operators failed first
- Credit spreads surge Telecom junk-bond spreads widened by 1,129 basis points (11.29 percentage points) between January 2000 and August 2001.34
- Defaults concentrate Telecom accounted for more than half the value of U.S. and Canadian high-yield defaults in 2001–2002.35
- Debt overhang The FCC chair estimated industry debt at 'around $1 trillion by some estimates'.9
Detail: defaults and the order of failuresDefault shares by data provider, troubled bank loans, bankruptcy chronology and revenues inflated by reciprocal transactions
- Default concentration Telecom accounted for 55% of the $63.6bn of U.S. and Canadian high-yield defaults in 2001, and 52% of the $96.9bn total in 2002.35
- Bank lending A 2002 shared national credit review of large syndicated loans classified 27% of telecom and cable credit commitments as criticized credits, up from 3.9% a year earlier.36
Share of defaulted dollar volume attributable to telecom; two datasets with different coverage are shown side by side.
Moody's figures are quoted in an FDIC report and cover all rated corporate bonds. Altman's dataset covers U.S. and Canadian high-yield debt; its 12.8% default rate in 2002 was a record at the time.35, 37
- 2000.03Nasdaq Telecommunications Index peaks 1,230 Down to 136 by May 2003
- 2001.04Winstar files for bankruptcy
- 2001.06360networks seeks bankruptcy protection March-end assets $6.3bn; liabilities $3.6bn
- 2002.01Global Crossing files for bankruptcy Filing lists $22.4bn in assets; $12.4bn in liabilities
- 2002.01McLeodUSA enters a prepackaged bankruptcy
- 2002.04FLAG Telecom and Williams Communications Williams plans to reduce debt by roughly $6bn
- 2002.06XO Communications files for bankruptcy
- 2002.07WorldCom files for bankruptcy Second-largest U.S. long-distance operator; pre-fraud-disclosure book assets $103.8bn, long-term debt $29.3bn
- 2002.11Genuity files for bankruptcy; agrees to sell assets to Level 3 for $242m
Assets appeared to exceed liabilities at book value. But networks whose market prices had collapsed could not be sold for their carrying amounts. Sources: company announcements and SEC filings; Global Crossing amounts from U.S. House testimony; Nasdaq index from the Richmond Fed.30, 38
Some firms also overstated financial performance. In Q2 2001, reciprocal purchases of similar value with other carriers accounted for 91% of Global Crossing's sales of long-term indefeasible rights of use (IRUs). Excluding these transactions, the company's claimed adjusted EBITDA would have been −$43m, not +$472m. The SEC sanctioned the company for inadequate disclosure.39 Qwest improperly recognized more than $3.8bn in revenue through reciprocal arrangements and other transactions and paid a $250m fraud settlement.40 Reciprocal deals concealed the revenue damage caused by collapsing prices.
WHAT COULD INVESTORS HAVE KNOWN?Warning signs already public by 2000
- 'Doubling every 100 days' was one company's projection; academic measurements put annual traffic growth closer to 100% (October 1998 paper and November 2000 follow-up).27, 41
- Level 3's 1999 disclosures warned of major capex before meaningful revenue and possible price pressure from competitors' capacity.
- A November 2000 Bank for International Settlements (BIS) report found that telecom received roughly a quarter of global syndicated loans in January–September 2000. European operators' 3G spectrum payments accounted for a substantial share.42
Default rates, bankruptcies and accounting scandals became evident later. Some early warnings were publicly available before the wave of failures.
SETTING THE RECORD STRAIGHT'Only 2.7% of installed fiber was being used'
- The source This is not an official government statistic. Citing a Merrill Lynch report from April 2001, the Dallas Fed noted that long-haul backbone utilization had fallen from 15% in 1988 to below 3% in April 2001—and cautioned that the figure did not account for peak-hour usage.43
- What it measures Average utilization of equipped, or 'lit,' capacity—not the share of all fiber strands installed in the ground. This was Merrill Lynch's description as reported in the trade press; telecom executives disputed the claim.44
- Official data The FCC's published series ended with 1998 data. In 1996, 53.1% of AT&T's fiber strands were equipped and in service.45
Odlyzko attributed the bust to overinvestment, not a slowdown in traffic growth.46SK Securities argues that demand weakened first in both the dot-com bust and the global financial crisis, but 47the fiber-optic industry collapsed even as internet usage kept growing. Strong usage alone does not establish that a repeat of 2000 is impossible.
Three questions from
the fiber-optic bust
We convert three conditions identified by Kiwoom Securities into diagnostic questions, comparing contemporaneous records from 2000 with today's filings. This is a framework for knowing where to look, not a formula for declaring a bust.
2000
2026
- CoreWeave: more than 98% of 2025 revenue came from committed, take-or-pay contracts. That describes revenue already recognized, not the proportion of facilities under construction that have been pre-sold.48
- The three cloud providers report roughly $1.829tn in backlog (Microsoft's figure covers its entire commercial business), about three times the four companies' trailing-12-month capex. That backlog represents revenue expected over multiple years.13
- The critical issue is not merely whether a contract exists, but whether the counterparty can pay (Section 04).10
2000
2026
- At the core: Microsoft's new 10-year notes yielded 5.60% at issue (Aaa).2
- At the fringes: Oracle's 20-year notes yielded 8.67% (Baa2, negative outlook); CoreWeave's unsecured bonds yielded 15–16% in the secondary market.3, 10
- Even GPU-backed loans diverge with end-customer credit quality. IREN's financing against a Microsoft contract cost about 6.0%, versus 9.0% for exposure to a below-investment-grade customer (Kiwoom Securities).10, 49, 50
2000
- Wholesale circuit prices fell more than 90% in two to three years (TeleGeography); some capacity was offered below construction cost.33
- Fiber cables were long-lived assets, but their economic value collapsed first.
2026
- Kiwoom's model for 1 MW of IT load (server power): approximately $26m in equipment, a four-to-six-year GPU useful life and a five-year customer contract.
The payback calculation is EBITDA-based, so actual cash recovery may take longer. Pre-tax profit assumes depreciation over five years and an 8% annual interest rate.10Annual revenue Simple payback Pre-tax profit $20m About 1.5 years About $10m $10m About 3.7 years Near zero - Microsoft says roughly three-quarters of capex is on short-lived assets.14
Kiwoom Securities' assessment
‘Not yet’If the cycle falters, the trigger is more likely to be a counterparty credit event—a customer failing to pay—than an immediate collapse in equipment values across the board.10A fresh look at official disclosures likewise finds insufficient evidence of a systemic, 2000-style collapse.That does not mean profitability has been established.
- Even with contracts, low pricing can mean poor returns.
- Companies can service their debt while shareholders still lose if equipment values fall and repeated equity issuance dilutes existing holders.
- Even uninterrupted access to financing cannot prevent returns from deteriorating as payback periods lengthen.
AT THE FRINGESWhere the warning signs appear first
In 2000
New entrants funded through IPOs and high-yield bondsAs the Richmond Fed described it, a flood of IPOs gave way to a flood of bankruptcies, with the crisis eventually engulfing WorldCom amid accounting fraud.30Today
Neoclouds and OracleThe four Big Tech companies plan to spend $720bn–$740bn this year. Oracle alone plans $90bn–$95bn in FY2027 capex (Kiwoom Securities).10The weakest borrowers, rather than the sector average, provide the earliest warning signals. Kiwoom's four warning indicators appear alongside their thresholds in the final checklist.10
A financing-rate ladder for AI infrastructure. Collateral, seniority and maturities differ, so not every yield gap reflects customer credit alone. Issue yields and secondary-market yields are also mixed, as are observation dates. SOFR is the Secured Overnight Financing Rate, a benchmark for short-term U.S. dollar lending.
Microsoft and Oracle figures are final offering yields to maturity (Microsoft: 5.595%, or 65bp over the comparable Treasury). CoreWeave's secured loan (March 30) and IREN figures come from company announcements; CoreWeave's unsecured secondary-market yields are Bloomberg data cited by Kiwoom Securities. Historical comparison: telecom junk-bond spreads widened 1,129bp in nineteen months around 2000.2, 3, 10, 34, 49, 50, 51
Higher interest rates would translate that gap into a larger debt-service burden. In Hana Securities' scenario with the 10-year U.S. Treasury yield at 6–7%, interest expense would remain at or below 5.5% of operating cash flow for the four Big Tech companies, but would rise sharply for Oracle and CoreWeave (Figure 10).25
Hana Securities interest-rate stress test. Solid bars show base assumptions; faded bars, conservative assumptions.
Parentheses denote the conservative case. Calculations use Bloomberg operating cash-flow forecasts and hold issuance spreads constant. Source: Hana Securities.25
The weakest link is
the customer's credit
In 2000, equipment makers extended credit to customers to keep equipment orders flowing. Today, financing exposure is spread across guarantees, leases and joint ventures.
2000Vendors financed customers to sell equipment
- LucentCustomer-financing commitments had reached $8.1bn by September 2000, which Lucent said was often a prerequisite for winning contracts or bidding. The company recognized approximately $1.8bn and $765m in customer-financing losses in FY2001 and FY2002, respectively.52
- Nortel At year-end 2000, undrawn customer-financing commitments reached approximately $4.1bn; in January–September 2001, Nortel reserved about $1.07bn against customer receivables and financing losses.53
- Cisco Customer loan commitments were about $1.9bn in July 2001. When demand suddenly fell, Cisco wrote down $2.25bn in inventory.54
Outstanding customer-financing commitments. Lucent FY2002 annual report (10-K exhibit); fiscal year ended in September.52
Equipment vendors were selling to cash-constrained competitive local exchange carriers (CLECs). Aggregate capex by these customers fell nearly 80% from its 2000 peak by 2002.18 When customers failed, vendor loans became vendor losses.
2026Guarantees, leases and joint ventures replace vendor loans
- Nvidia lease guarantee Disclosed in August: Nvidia provided a guarantee with a cumulative cap of $105bn for a 20-year lease under which SB Energy would lease facilities in Portsmouth, Ohio, to an OpenAI affiliate, covering roughly 4.25GW of IT load. This is not a guarantee that activates only at lease maturity: Nvidia may have to cover shortfalls during the term if OpenAI becomes unable to pay (Table 3).55
| Item | Provision |
|---|---|
| What is guaranteed | Contractually specified portions of lease and power payments for land, electrical infrastructure and building shells—not the resale value of GPUs, all costs of the site or every liability of the tenant. |
| When it can be called | If OpenAI becomes insolvent, fails to perform the lease, or defaults on rent. The guarantee covers the difference between the contractually specified minimum lease value and recoveries from re-leasing or asset sales. |
| When it becomes effective | In stages as each of nine data centers enters service, with facility availability expected from 2028. |
| How it declines or terminates | The guaranteed amount declines as OpenAI pays rent and ends if OpenAI obtains a sufficiently strong credit rating. |
| Remaining exposure | OpenAI has agreed to indemnify Nvidia, but that promise itself depends on the credit of the same company that may fail to pay the rent. |
Sources: Nvidia's August 17 8-K and Q2 10-Q.55
- Meta's Hyperion project In October 2025, a joint venture owned 80% by funds managed by Blue Owl and 20% by Meta agreed to develop a data-center complex costing approximately $27bn. Meta guaranteed a capped portion of residual value over its initial sixteen years of operations.56
- SpaceXOn October 26, SpaceX disclosed a commitment for roughly $40bn in GPU finance leases with an Apollo-managed entity. Each equipment tranche carries a term of approximately five to 5.25 years. Customer credit risk rests with the lessor and its investors.57
| Dimension | 2000 | 2026 |
|---|---|---|
| Who bears the risk? | Equipment vendors (Lucent, Nortel, Cisco) | Equipment vendor (Nvidia guarantee); funds and leasing vehicles managed by private-capital firms (Apollo and Blue Owl affiliates); and the tenant itself (Meta) |
| Mechanism | Vendors lent to customers or deferred payment—vendor financing. | Finance firms and funds build or lease facilities; suppliers and tenants guarantee contractual lease value or residual value. |
| Scale | Lucent $8.1bn committed; Nortel up to about $4.1bn; Cisco about $1.9bn | Nvidia guarantee cap $105bn; SpaceX lease commitment about $40bn; Hyperion development costs about $27bn |
| What may not appear in the primary balance-sheet totals | Undrawn customer-financing commitments | Disclosed joint-venture and leasing structures that build and lease the assets, with guarantees described in footnotes |
| When losses become visible | Customer bankruptcy → equipment vendors set aside loss reserves (2001–2002) | Guarantees become effective after leases commence (expected from 2028 for Nvidia). Nvidia's guarantee can produce losses during the lease term if the tenant becomes insolvent or fails to pay rent, net of recoveries through re-leasing or asset sales. |
These amounts cannot be compared directly because the obligations are different. Lucent's $8.1bn was a commitment to lend; Nvidia's $105bn is a cumulative cap on specified guaranteed shortfalls. Its ultimate exposure depends on what can be recovered through re-leasing or asset sales after a tenant default. Sources: company filings.52, 53, 54, 55, 56, 57
The end customers fall into two broad groups: Big Tech, with $676.8bn of trailing-12-month operating cash flow, and a handful of AI labs dependent on capital raises and IPO proceeds. The weaker link is the second group.
- Kiwoom Securities The ability of a few major buyers, including Anthropic, to meet commitments depends on IPO proceeds and guarantees from hyperscalers.10
- DS Investment & Securities As Anthropic expands its direct model—leasing chips and installing them in dedicated facilities—the AI developer bears the risk of paying lease charges even when construction is delayed.58
In its amended registration statement on November 2, Anthropic described the mismatch itself. Most revenue is usage-based rather than locked into long-term contracts, while compute and infrastructure commitments run for years. Many of its largest customers can also reduce or stop spending at any time.59 Revenue depends on usage. The bills are committed for years.
Total commitments as of September 30. Almost all relate to compute and infrastructure.
- Approximately $69.8bnWithin twelve months of September 30
- Approximately $447.5bnBefore year-end 2030 (including the next twelve months)
- Approximately $447.9bnRemaining commitments, mostly 2031–2045
The remaining amount is total commitments less those due before year-end 2030. Approximately $213.8bn is subject to pre-commencement termination rights or conditions involving counterparties' funding, facilities or permits. Source: Anthropic's amended registration statement (November 2).59
- Revenue and operating resultsRevenue totaled $33.2bn for January–September. Anthropic posted $970m in operating profit in Q2, but reported a $570m operating loss in Q3 after incurring $5bn in costly short-term compute expenses and $2bn in stock-based charitable donation matching (a noncash expense).59
More from the filingLeases not yet commenced, funding sources, collateralized borrowing, and Broadcom equipment leases
- Leases not yet commenced These increased from $50.2bn at year-end 2025 to $451.2bn as of September 30. Most are expected to commence in 2026–2028, with a weighted-average term of roughly ten years. They are included in the commitment total above.
- Sources of funding Operating cash flow was $6.6bn in January–September, while financing activities provided $110.7bn, largely from issuances of redeemable convertible preferred stock.
- Secured borrowing Anthropic increased its revolving credit facility from $3.25bn to $15bn in September. Substantially all assets secure the facility; collateral and guarantees may be released if at least two of three rating agencies assign investment-grade ratings.59
- Broadcom equipment leases Anthropic will pay approximately $161.2bn, excluding interest, for TPU equipment acquired under five-year leases. In October, Broadcom transferred some rights and obligations under equipment leases to a new special-purpose vehicle (SPV). Anthropic stated that the transfer did not increase its total lease payments.59 Reuters reported that Bank of America, Citi, Morgan Stanley and Blackstone had committed $63bn in financing, with Broadcom providing residual-value support to lenders on some equipment.60 Hana Securities noted that splitting the exposure among suppliers, SPVs, lenders and bondholders does not eliminate the underlying risk.25
In 2000, equipment makers bore most of their customers' credit risk. Today, the exposure is shared among equipment suppliers (Nvidia and Broadcom), lease vehicles and funds, the banks and private-capital managers financing them, and lessees themselves. The structure is examined in greater detail in Research No. 12.
Efficiency echoes 2000.
The power grid changes the equation.
In 2000, an equipment upgrade could multiply capacity on cables already in the ground. Today, projects can stall because power is unavailable. Yet there is a parallel: technology can rapidly expand effective supply.
WHAT IS DIFFERENT
- Who funds the buildout Then, startup telecom operators relied on IPO proceeds and debt. Today, cash-generative Big Tech finances most investment.
- The bottleneck Then, endpoint equipment upgrades multiplied capacity, and grid interconnection was not the primary constraint it is today. Now, access to power is the bottleneck.
- Credit quality Then, telecom accounted for over half of high-yield defaults. Today, the largest investors borrow at investment-grade rates—Microsoft is rated Aaa.2, 35
WHAT RHYMES
- Rapid growth in usage Google's monthly AI token volume rose more than 300-fold in two years, according to figures cited by Kiwoom Securities. Internet usage also surged in 2000.10
- Efficiency expands supply Just as DWDM expanded capacity per fiber strand, a given GPU can produce more tokens, while the cost of generating an equivalent answer is falling rapidly (Table 6).10
- A surge of new entrants Then, Chicago and Minneapolis each had twenty long-haul telecom operators.18 Now, neoclouds (emerging GPU-rental providers) and former Bitcoin miners that have pivoted into data-center leasing have entered in force.10
Large loads of at least 75MW within the Electric Reliability Council of Texas (ERCOT) system. Requested capacity comes from the governor's August release; energization approvals and observed usage from the September operating report. These are different stages and observation dates.
Requests include long-range proposals, so these figures cannot be treated as a conversion funnel or utilization rate for the same set of projects. ERCOT regards observed load as evidence of actual operation among approved loads. Since August, it has paused energization approvals for data centers of 75MW or more pending verification. A review report is due December 10 and a Public Utility Commission of Texas (PUCT) meeting is scheduled for December 17.61
Power and siting constraints cut both ways: they may prevent a 2000-style surge of excess capacity, but they also delay the payback on equipment already purchased.
| Constraint | How it limits oversupply | How it delays payback |
|---|---|---|
| Grid interconnection | Capacity cannot all come online at once. | Purchased GPUs can sit idle without grid power. Kiwoom estimates a one-year delay in commercial operations costs $7m–$19m per IT MW.10 |
| Policy and local opposition | Slows the pace of new construction. | Texas energization pause; heightened regulatory concern after the midterms5 ; only 18% of counties net supportive of new construction; one-year pause on major project permitting in New York State25 |
| GPU useful life of 4–6 years | Unlike long-lived fiber, GPUs do not depress market supply for decades. | Shorter economic lives leave less time to recover capital, while collateral values fall quickly. |
Whether regulatory concerns result in actual restrictions remains uncertain. Demand indicators, however, are strong: CBRE reports a 1.4% vacancy rate across eight major U.S. markets and 80.4% preleasing for capacity under construction in H1 2026. Hana Securities argues that such constraints will affect commissioning dates and investment returns more than demand itself.25
Three different prices must be tracked.
| Price | Current signal |
|---|---|
| Cost to produceCost of generating an answer of equivalent quality | Reported to be falling more than tenfold annually (cited by Kiwoom Securities)10 |
| Price charged to AI usersWhat AI service providers charge their customers | Small-model API prices down 56–90% in six weeks (Kiwoom Securities)10 |
| GPU rental priceWhat an AI company pays to rent GPU compute | Hourly Nvidia B200 rental rates rose 14% in October (SK Securities)47Kiwoom argues falling end-user prices will feed into renegotiations and that short-term contracts yielding about $40m per IT MW per year are unlikely to persist.10 |
If the cost of providing compute falls faster than usage-based prices, providers can cut prices and preserve margins. If prices fall faster than costs, customers have greater leverage to renegotiate rental rates.Whether GPU rents break, as wholesale circuit prices did in 2000, is the key to the third question.
Counterarguments—and
four possible paths
The case for why this time is different is substantial. But each counterargument depends on evidence that has yet to be verified.
A cash-flow recovery, as in 2019
Following the 2017–2018 cloud investment cycle, FCF rose in 2019 to $38.3bn at Microsoft and $31.0bn at Alphabet. Yuanta Securities argues that cash recovery can begin once capex growth has passed its peak.23
WHAT TO CHECKThe same firm still forecasts negative combined FCF for the four companies in H1 2027 (Section 01).
Bubbles burst when end demand falters
In SK Securities' reading, dot-com demand and then housing prices weakened first; vendor financing and accounting manipulation amplified the damage. It sees funding concerns today but no evidence yet of weakening AI demand.47
WHAT TO CHECKAs Section 02 showed, the fiber-optic industry collapsed while usage kept rising. The rate of supply growth and pricing must be examined together.
Incremental returns already look positive
For the three major cloud providers, each additional dollar of depreciation was accompanied by $2.10 of incremental cloud operating profit in Q3 2026, up from $0.50 a year earlier. That is above the upper end of iM Securities' 1.6–2.0× benchmark, although Amazon alone stood at 1.5×. Even if the roughly $200bn FY2027 funding shortfall across five companies including Oracle (dividends included) were fully financed with debt, net debt/EBITDA would remain below 1.2× for all four Big Tech firms; Oracle alone would reach 4.4×.62
WHAT TO CHECKDepreciation follows capital spending with a lag, so the ratio can look favorable while investment is accelerating. Return on invested capital (ROIC) has fallen close to the cost of capital. iM Securities also cautions that funding capacity and the profitability of AI services must be evaluated separately.62
Credit stress may remain company-specific
Daishin assigns a 60% base-case probability to AI credit problems remaining confined to individual companies. It would watch for broader contagion if the U.S. 10-year Treasury yield exceeds 5.30% or high-yield spreads reach 450bp.63
CURRENT READINGSU.S. 10-year yield: 4.90% (Nov. 4); high-yield spread: 340bp (Nov. 3).64, 65
There is also a more cautious calculation. Yuanta Securities finds that justifying $1.1tn of capex in 2027 would require revenue above market expectations even if the assumed required return fell to 5% (Figure 14).66
Source: Yuanta Securities.66
These are not mutually exclusive forecasts. Big Tech FCF may recover even as some suppliers and neoclouds lose money and their share prices remain weak.
| Path | What it looks like | Signals pointing this way |
|---|---|---|
| ① A 2019-style payback | Capex growth moderates and operating cash flow catches up. | Four-company quarterly FCF turns positive; 2027 capex growth decelerates; renewal rates hold up. |
| ② A slump without defaults | No widespread defaults, but poor returns and falling expectations keep stocks depressed for a prolonged period. | Capex continues to rise but misses expectations; supply-chain prices and margins weaken; Big Tech FCF recovery is delayed; credit metrics remain calm. |
| ③ Credit event at the fringes | One or two neoclouds or AI labs lose access to funding; pricing for uncontracted capacity breaks. | Neocloud bond yields stay above 12%; collateral backed by below-investment-grade customers rises; IPOs or financings are delayed. |
| ④ A 2000-style bust | GPU rental prices collapse and losses reach the core of the investment cycle. | Lower AI-user prices flow through to GPU rents; investment-grade collateralized loans remain at SOFR + 400bp; trailing-12-month Big Tech FCF turns negative; high-yield spreads reach 450bp. |
We do not assign probabilities to these paths. Daishin's base case is closest to Path ③. Path ② underscores that stocks can underperform even when credit indicators remain tranquil.63
In 2000, the steeper losses
were in equipment stocks
Korean equities now sit on the equipment side of the AI cycle—the supply chain. What equipment makers' share prices anticipated in 2000 offers a relevant lesson for Korean investors.
S&P sector indexes, compiled by the Dallas Fed using Bloomberg data.43
Equipment stocks fell before capital spending contracted. Telecom equipment investment peaked in 2000 and began to decline in 2001, but the equipment index had already peaked alongside the Nasdaq in March 2000.31, 43
- Concentration Daishin Securities argues that investors have bought the AI supply chain rather than the companies doing the spending. It forecasts that semiconductors will account for 78% of the KOSPI's projected KRW 1,344tn in operating profit for 2027.63
- The current shortageMemory supply remains tight, and 2027 output of high-bandwidth memory (HBM) has already been subject to advance contracting (Kiwoom Securities).10But as Cisco demonstrated in 2001, a sudden stop in orders can turn accumulated inventories into losses.54
- Growth rates Both Yuanta Securities and Hana Securities expect capex growth to slow next year.23, 24But if spending moves from 100 to 200 to 260, the growth rate slows from 100% to 30% even as the amount spent continues to rise.
That makes total spending and growth rates insufficient on their own. Four additional indicators deserve attention.
| Metric | Why it matters |
|---|---|
| Actual spending versus expectations | Share prices may react more to whether announced spending beats or misses expectations than to the growth rate itself. |
| Where the money goes | The same capex total can benefit very different suppliers depending on the allocation among buildings, power systems, servers and chips. |
| Orders, deliveries and commissioning | There is a lag between announced investment plans and revenue recognized by suppliers. |
| Pricing, inventory and margins | Orders may grow without a proportional increase in profit. |
How that spending is financed matters too. The more customers rely on debt and guarantees, the more their suppliers' orders depend on financing conditions.A low probability of default is not the same as an attractive expected return for shares in the supply chain.
Early-warning
checklist
It is hard to declare a bubble in a single sentence. We can, however, measure which warning signals have appeared. This is a risk-monitoring dashboard, not a model that turns the number of red lights into a crash probability or a trading signal.
Clear (gray); approaching threshold (amber); triggered (red); not observable (dashed, no public aggregate). Unobservable does not mean safe. Thresholds are those cited by brokerages, except that we define the FCF warning as two consecutive negative quarters or a negative trailing-12-month total. Data measured against a different denominator or population are included only as context.
| Indicator | Trigger / threshold | Latest reading | Status |
|---|---|---|---|
| Construction without committed customers | Uncontracted share starts rising (Kiwoom Securities) | No industry-wide figures · Context: (1) over 98% of CoreWeave's recognized 2025 revenue came from committed contracts; (2) 80.4% of the 7.48GW under construction in eight major U.S. markets was preleased (CBRE, H1 2026; cited by Hana Securities). | Unobservable |
| Rate on loans backed by investment-grade contracts | Sustained SOFR + 400bp or more (Kiwoom Securities) | Earlier transactions: SOFR + 213–225bp (IREN, June); SOFR + 225bp (CoreWeave, March). New financing terms since the October selloff have not yet been disclosed. | ClearBased on March–June deals; later terms unavailable |
| Collateral backed by below-investment-grade customers | Share exceeds 10% of GPU-backed loans (Kiwoom Securities) | No industry-wide data. IREN financed equipment for below-investment-grade customers at a fixed 9.0% rate; the share is undisclosed. | Unobservable |
| Multiyear renewal pricing | Below $20m per IT MW per year (Kiwoom Securities) | No public aggregate · Context: IREN's recent three-year new contracts around $22m per IT MW annually; short-term contracts about $40m (Kiwoom estimate), both above the threshold. | Unobservable |
| Neocloud corporate bonds | Normalization: yield at or below 12% and CDS below 1,000bp (Kiwoom Securities) | CoreWeave's 2031 unsecured notes yielded 15–16% in mid-October (cited by Kiwoom); its CDS reached 1,146bp on October 23. iM Securities said the spread had eased below its previous high by early November · Context: 2032 notes yielded 16% (Hana Securities; observation date not stated). | Triggered |
| Four-company free cash flow | Two consecutive negative quarters or negative trailing-12-month FCF | Q3 −$26.4bn; trailing 12 months +$61.5bn. Yuanta forecasts −$44.1bn for H1 2027. | Approaching threshold |
| High-yield spread | 450bp (Daishin contagion threshold) | 340bp (November 3) | Clear |
| U.S. 10-year Treasury yield | 5.30% (Daishin contagion threshold) | 4.90% (November 4) | Clear |
Capex growth is not listed as a single warning light. Spending can keep rising even as the growth rate slows, and the implications for Big Tech FCF and Korean suppliers can point in different directions (Section 07). Sources: Kiwoom, Hana, Yuanta, Daishin and iM Securities; company filings; ICE BofA/FRED; U.S. Treasury.10, 12, 23, 25, 48, 49, 50, 51, 63, 64, 65, 67
Upcoming checkpoints Korea Standard Time (KST)
- Fri., November 6, morning IREN results: financing secured by the Microsoft contract, non-investment-grade customer facilities, and grid-connection timetables.68
- Wed., November 11, morning CoreWeave Q3 results: energized capacity, funding rates and backlog.69
- During November (date unannounced) Anthropic investor presentation and IPO; the amended filing does not yet specify a price range or date.59
- Wed., November 18, morning Nvidia Q3 results: data-center revenue and the scale of customer-financing support.70
- December 10 and 17 ERCOT verification report and PUCT meeting: whether Texas resumes energization approvals for data centers.61
- Late January–early February 2027 Big Tech Q4 results: first quantified 2027 capex outlooks.15, 17
KNOWNCONFIRMED
- The fiber-optic bust unfolded even as internet usage rose.
- Q3 combined FCF for the four companies was −$26.4bn; capex grew 108%, operating cash flow 10%.
- Four-company long-term debt rose from $136.6bn to $410.2bn in one year.
- Nvidia's lease guarantee has a cumulative $105bn cap and can be called upon tenant insolvency or failure to pay rent.
- Anthropic disclosed approximately $895.4bn in commitments, including about $69.8bn due within twelve months.
ESTIMATEESTIMATES
- EBITDA-based simple payback of 1.5–3.7 years per IT MW (Kiwoom Securities).
- Capex growth slows to 52% in Q2 2027; four-company H1 FCF −$44.1bn (Yuanta Securities).
- $3.9tn of revenue needed to justify 2027 capex (Yuanta Securities).
- 60% probability that AI credit problems remain company-specific (Daishin Securities).
- Simple sum of company FY2026 FCF estimates: −$41.7bn for six companies, +$8.9bn for four (different fiscal years; not the same reporting period; Bloomberg estimates cited by Hana).
- Incremental cloud operating profit / incremental depreciation reached 2.1× in Q3 2026 versus a 1.6–2.0× benchmark (iM Securities' calculation and assumption).
UNKNOWNSTILL UNKNOWN
- The amount and growth rate of Big Tech's 2027 capex.
- How far GPU renewal pricing will follow the decline in AI end-user prices.
- Industry-wide uncontracted construction and below-investment-grade collateral exposure.
- Terms of new loans against investment-grade contracts since the October selloff.
- Whether AI labs can generate enough cash after IPOs to meet their commitments without further support.
Internet usage kept rising through the fiber-optic bust of 2000. What undermined investor payback was capacity growing faster than usage, falling prices and the cost of debt. The scale of AI contracts or growth in token consumption alone cannot establish investment success.
The warning presently visible in public data is neocloud debt. Industry-wide uncontracted construction and exposure to weaker customers remain unquantified.From here, track not just AI usage, but whether that usage generates enough cash to recover invested capital, cover its cost and leave something for shareholders.
Glossary
- Free cash flow (FCF)
- Operating cash flow less cash spent on capital assets. This report calculates FCF as operating cash flow minus purchases of property and equipment.
- Operating cash flow
- Net cash generated or used in core operations, distinct from accounting profit.
- EBITDA
- Earnings before interest, taxes, depreciation and amortization. Different from operating profit after depreciation or cash generated by operations.
- Finance lease
- An arrangement economically similar to purchasing an asset while paying in installments through lease payments. The full asset cost is not necessarily reflected immediately in cash capex.
- Lease not yet commenced
- A lease contract signed but not yet in effect. The obligation may be disclosed in footnotes before it is recognized on the balance sheet.
- DWDM
- Dense wavelength-division multiplexing: technology that increases a fiber strand's capacity by transmitting multiple wavelengths of light simultaneously.
- IRU
- Indefeasible right of use: a long-term contract granting use of fiber capacity, often for 10–20 years.
- Syndicated loan
- A large loan provided by a group of banks.
- High-yield debt / junk bonds
- Corporate debt rated below investment grade, generally carrying higher yields.
- Spread / basis point (bp)
- The extra yield above a benchmark such as government bonds. One basis point equals 0.01 percentage point.
- SOFR
- Secured Overnight Financing Rate, a short-term U.S. dollar benchmark. Floating-rate loans typically charge SOFR plus a credit spread.
- CDS
- Credit default swap: a derivative that transfers default risk. A higher CDS premium generally indicates a higher market-implied risk of default.
- Residual-value guarantee
- A commitment to cover a shortfall if resale or re-leasing proceeds fall below a contracted amount. The trigger depends on the contract; Nvidia's guarantee is linked to tenant insolvency or nonpayment of rent.
- Special-purpose vehicle (SPV)
- A legal entity established for a specific transaction, such as buying and leasing equipment or raising debt. It can separate designated assets and liabilities from those of its sponsor, subject to accounting and legal rules.
- Take-or-pay contract
- An agreement requiring payment for a committed amount even if the customer does not use all the capacity.
- Remaining performance obligations (RPO)
- The value of contracted obligations not yet recognized as revenue, typically to be earned over several years.
- Neocloud
- An emerging cloud provider focused on renting GPU compute capacity, such as CoreWeave, Nebius or IREN.
- Big Tech / hyperscaler
- A technology company operating data centers at enormous scale. The four examined here are Microsoft, Alphabet, Amazon and Meta.
- IT 1MW · GW
- One megawatt of power consumed by IT servers in a data center (1GW = 1,000MW). Kiwoom estimates roughly $26m of investment in GPU equipment per IT MW.
- Counterparty credit risk
- The risk that a contractual counterparty cannot make payments when due. Even a signed contract is no guarantee of cash collection.
- Token
- A unit of text or other content processed or generated by an AI model; token throughput is one indicator of usage.
- HBM
- High-bandwidth memory: high-throughput DRAM placed next to AI accelerators to move data rapidly.
Sources, definitions and methodology
Published Thursday, November 5, 2026. Most figures are as of the evening of Wednesday, November 4 (KST), supplemented by Hana Securities' November 5 report. Rates, credit spreads, Coherent's earnings and Meta/Alphabet bond transactions reflect the latest information publicly available before publication. Historical figures for 2000 were checked against contemporary Federal Reserve, FCC, BEA, SEC and Bank for International Settlements records, as well as academic research. Third-party industry data are attributed to their providers. Where the source reports from Korean brokerages are not publicly linkable, we provide the firm, title and date; forecasts and estimates remain those of the cited brokerages.
| Free cash flow | We derive individual quarters from cumulative cash-flow statements in 10-Q/10-K filings. FCF = operating cash flow − PP&E purchases (Amazon: gross PP&E purchases; Meta: net PP&E purchases). Microsoft's July–September quarter is aligned with calendar Q3. The comparable four-company quarterly series begins in Q3 2017. |
|---|---|
| Q3 calculation | Microsoft +$18.63bn + Alphabet −$9.07bn + Amazon −$34.84bn + Meta −$1.16bn = −$26.43bn. Trailing-12-month total: Q4 2025 $60.20bn + Q1 2026 $20.98bn + Q2 2026 $6.71bn + Q3 2026 −$26.43bn = +$61.46bn. |
| Capex / operating cash | Q3 PP&E purchases of $201.92bn ÷ operating cash flow of $175.50bn = 115%. A year earlier: $97.3bn ÷ $159.0bn = 61%. Growth rates: capex +107.6%; operating cash +10.4%. |
| Long-term debt | Noncurrent long-term debt on balance sheets. September 2025: 35.4 + 21.6 + 50.7 + 28.8 = $136.6bn; September 2026: 31.1 + 125.1 + 173.1 + 80.9 = $410.2bn. Amounts in the equation are USD billions. |
| 2026 capex guidance | Microsoft about $170bn (calendar year) + Alphabet $195–205bn + Amazon approximately $220bn + Meta $135–145bn = $720–740bn. Finance-lease inclusion differs by company. |
| Annualized expected demand | Doubling every 100 days = 2^(365/100) ≈ 12.6× per year. Relative to observed growth of 2×, that is roughly a 6.3× overshoot, consistent with the Boston Fed president's 'factor of six' comment. |
| Historical data for 2000 | BEA telecom equipment investment uses the currently revised series. Bankrupt firms' assets and liabilities come from filings or congressional testimony. Moody's default totals cover corporate bonds broadly, while Altman's figures cover U.S. and Canadian high-yield bonds. |
- Reuters (republished by Investing.com), 'Wall Street closes sharply lower as AI cash burn fears spark tech selloff,' Oct. 30, 2026. Nasdaq −2.86% (25,437.00), semiconductor index −7.2%, Intel −28.6%, AMD −10.9%, Nvidia −4.2%. · investing.com
- Microsoft, final pricing term sheet (FWP), priced Oct. 29, 2026; settled Nov. 4. $25bn across ten tranches due 2028–2066. The 2036 notes carried a 5.55% coupon and 5.595% yield to maturity (+65bp to Treasuries); Aaa/AAA. · SEC
- Oracle, final bond term sheet (FWP) and preliminary equity prospectus (424B5), Nov. 2, 2026. $10bn in bonds (coupons from 7.00% on 2029 notes to 8.65% on 2046 notes; issue yields 7.062–8.667%, including 7.851% on 2033 notes); Baa2 (negative)/BBB−/BBB. Also $10bn of common equity. · FWP · 424B5
- Bloomberg and Reuters, “Meta Raises €9 Billion as Euro Bond Debut Draws Record Demand” (Bloomberg), “Meta draws record orders for debut euro bond sale” (Reuters), and “Meta Goes to Europe for Debt After Raising $15 Billion” (Bloomberg, republished by Yahoo Finance), Nov. 4, 2026. Nov. 3: $15bn of dollar bonds in five tranches due 2031–2056, with orders reaching up to $58bn. Nov. 4: Meta’s first €9bn euro bond sale across six tranches drew a record €53bn in orders. ·Bloomberg · Reuters · Yahoo Finance
- Yonhap News Agency (republished by Seopul), 'KOSPI whipsaws on U.S. midterms before closing 0.62% higher,' Nov. 4, 2026. Intraday +3.17% (5,828.13) and −4.02% (5,421.72), close 5,683.80 (+0.62%); foreign investors net sold KRW 2.5tn; AI data-center regulatory concerns. Samsung Electronics and SK hynix fell 5–6% intraday before closing −0.68% and −0.14%, respectively. · Seopul
- AMD, Q3 2026 earnings release (8-K Exhibit 99.1), Nov. 3, 2026. Revenue $13.4bn (+45%); data-center revenue $8.5bn (+97%); Q4 revenue outlook about $17bn. · SEC
- Arista Networks, Q3 2026 earnings release (8-K Exhibit 99.1), Nov. 3, 2026. Revenue $3.66bn (+58.5%); Q4 revenue guidance $3.9–4.0bn. · SEC
- Coherent, FY2027 Q1 (July–September) earnings release (8-K Exhibit 99.1), Nov. 4, 2026. Revenue $2.37bn (+50%; pro forma +56%); projected FY2027 pro forma revenue growth above 60%. · SEC
- U.S. Senate Committee on Commerce hearing, 'Financial Turmoil in the Telecommunications Marketplace,' written testimony of FCC Chairman Michael Powell, July 30, 2002. Approximately $2tn of lost market capitalization, estimated $1tn debt and 500,000 lost jobs; traffic doubling annually rather than every 100 days; capacity up by some estimates 500×; destructive price competition. · govinfo.gov
- Kiwoom Securities, 'Neo and Beyond: Neoclouds and Powered Shells—AI Compute and Grid Bottlenecks: Business Models and Costs of Capital,' Oct. 28, 2026. Three conditions behind the fiber-optic bust and associated warning signals; unit economics per IT MW; financing-rate ladder; CoreWeave bonds and CDS (Bloomberg data); token prices and GPU rents; costs of a one-year delay.
- SEC, Litigation Release No. 17588, June 27, 2002. Enforcement action concerning WorldCom's accounting fraud, including capitalization of approximately $3.8bn in operating expenses. · sec.gov
- SEC EDGAR XBRL companyfacts, accessed Nov. 4, 2026: Microsoft, Alphabet, Amazon and Meta 10-Q/10-K operating cash flows, PP&E purchases and noncurrent long-term debt. · sec.gov
- Eugene Investment & Securities, 'Global Technology Weekly—The Gap Between AI Spending and Conviction,' Nov. 3, 2026. Combined Q3 FCF −$26.4bn; combined three-company remaining performance obligations of $1.829tn.
- Microsoft, FY2027 Q1 (July–September 2026) earnings release and call, Oct. 28, 2026. Operating cash $58.0bn; PP&E cash spending $39.4bn; capex including finance leases $45.2bn ($7.4bn in finance leases); FCF $18.6bn; roughly three-quarters of capex on short-lived assets; calendar-year capex about $170bn. · microsoft.com
- Alphabet, Q3 2026 earnings and 10-Q, Oct. 28, 2026. Operating cash $42.29bn; PP&E purchases $51.36bn; approximately $5.2bn paid in July for an EU fine and interest. Maintained 2026 capex guidance of $195–205bn and signaled a substantial rise in 2027. · abc.xyz
- Amazon, Q3 2026 earnings, Oct. 29, 2026. Operating cash $42.42bn; PP&E purchases $77.26bn before $1.38bn in asset-sale proceeds and incentives; trailing-12-month FCF −$42.36bn; 2026 cash capex approximately $220bn, with 'meaningfully more' planned for 2027. · amazon.com
- Meta, Q3 2026 earnings, Oct. 21, 2026. Operating cash $32.79bn; PP&E purchases $33.94bn; finance-lease principal $960m; FCF −$2.12bn; 2026 capex guidance of $135–145bn including finance-lease principal. · investor.atmeta.com
- M. Doms, 'The Boom and Bust in Information Technology Investment,' Federal Reserve Bank of San Francisco Economic Review, 2004. Capex at U.S.-listed telecom services providers declined from $121bn in 2000 to $49bn in 2002; long-haul fiber routes increased from 143,000 miles (1996) to 320,000 (2000), with twenty carriers per city; CLEC capex fell almost 80% by 2002. · frbsf.org
- Reuters reporting cited by The Next Web, “Alphabet launches €6B bond sale in first test of AI debt since selloff,” Nov. 2, 2026. €6bn in seven tranches (one floating-rate and six fixed-rate, with maturities of two to 19 years); the order book exceeded €17bn. ·thenextweb
- Oracle, 'Oracle Investor Day 2026' press release, Oct. 28, 2026. FCF projected to turn positive in H2 FY2028; cap of $20bn on additional FY2027–FY2030 capital-market funding ($10bn equity, $10bn debt), excluding refinancing and separate from $20bn of at-the-market equity issuance completed in FY2027 Q1. · investor.oracle.com
- Daishin Securities, 'Quentin Tactical Plan—AI Financing Worries Resurface Ahead of U.S. Midterms,' Nov. 3, 2026. Reported potential $20bn Meta bond issuance in USD and EUR; Alphabet's prospective €6bn European bonds; Oracle issuance and Anthropic commitment disclosures reignited financing concerns.
- iM Securities, 'Market Weekly—The Market's Patience Is Being Tested,' Nov. 2, 2026.
- Yuanta Securities, '2027 ETF Investment Strategy—Cooling Down,' Oct. 29, 2026. Projected four-company capex growth from 107% in Q3 2026E to 52% in Q2 2027; the 2019 FCF precedent; estimated combined H1 2027 FCF −$44.1bn.
- Hana Securities, '2027 Equity Market Outlook and Strategy—When Earnings Change, Market Leaders Change,' Oct. 29, 2026. Projected U.S. technology capex growth slowing from 102% in Q3 2026 to 14% in Q4 2027.
- Hana Securities, '2027 Global Credit Outlook—Through the Cash-Flow Gap,' Nov. 5, 2026. Bloomberg FY2026 FCF forecasts for six companies, totaling −$41.7bn on a simple, non-aligned-fiscal-year basis; projected return-to-positive dates and capex/operating-cash ratios; rate stress tests using 6–7% U.S. 10-year Treasury yields. CoreWeave's 9.625% notes due 2032 rose from 9.675% at issue to a recently cited 16% yield (date unspecified). CBRE's eight-market U.S. data: 1.4% vacancy and 80.4% of 7.48GW under construction preleased; Fraym's 18% net support at county level; New York's one-year moratorium; structured finance and risk distribution.
- U.S. Department of Commerce, 'The Emerging Digital Economy,' April 1998. Reported UUNET's estimate that internet traffic doubled every 100 days. · commerce.gov
- K. G. Coffman and A. M. Odlyzko, 'The size and growth rate of the Internet,' First Monday 3(10), Oct. 5, 1998. Estimated public internet traffic growth around 100% annually. · firstmonday.org
- D. Galbi, FCC Common Carrier Bureau, 'Growth in the 'New Economy': U.S. Bandwidth Use and Pricing Across the 1990s,' July 9, 2000. Traced the 100-day claim; Band-X price index fell about 63% from October 1998 to July 2000. · arXiv
- C. Minehan, President of the Federal Reserve Bank of Boston, speech of Sept. 9, 2002. Traffic doubled annually, not every 100 days; forecasts were off by a factor of about six. · FRASER
- J. Couper, J. Hejkal and A. Wolman, 'Boom and Bust in Telecommunications,' Federal Reserve Bank of Richmond Economic Quarterly 89(4), Fall 2003. Per-strand capacity from 2.5Gbps (1996) to 100Gbps maximum (2000); three causes of oversupply, most importantly weak demand versus expectations; 'flood of IPOs' becoming a 'flood of bankruptcies'; Nasdaq Telecommunications Index 1,230.06 (Mar. 2000) to 136 (May 16, 2003). · richmondfed.org
- U.S. Bureau of Economic Analysis (BEA), National Accounts, 'Private Fixed Investment: Telecommunications Equipment,' nominal annual series, FRED code C275RC1A027NBEA; currently revised figures. · FRED
- Level 3 Communications, Form 8-K/A risk factors, Nov. 9, 1999. Capex ahead of revenue, uncertainty of customer acquisition, and risk of price declines from rival capacity. · SEC
- S. Beckert, TeleGeography, 'Bandwidth price collapse: will it ever end?' Lightwave, Aug. 1, 2002. Wholesale prices down more than 90% on most routes over two to three years; trans-Pacific capacity offered in October 2001 at prices far below estimated construction cost. · lightwaveonline.com
- S. Kwan, 'Rising Junk Bond Yields: Liquidity or Credit Concerns?' Federal Reserve Bank of San Francisco Economic Letter 2001-33, Nov. 16, 2001. Telecom junk-bond spreads widened 1,129bp from January 2000 to August 2001. · frbsf.org
- E. Altman and G. Bana, 'Defaults and Returns on High Yield Bonds: The Year 2002 in Review,' NYU Stern, Feb. 2003. U.S. and Canadian high-yield defaults of $63.61bn (9.80%) in 2001 and $96.86bn (12.80%) in 2002; telecom shares of 55% and 52%, or 31% in 2002 excluding WorldCom. · nyu.edu
- Federal Reserve, FDIC and OCC, Shared National Credit Review results, Oct. 8, 2002. Criticized commitments in telecom and cable reached 27.0%, versus 3.9% a year earlier. · federalreserve.gov
- FDIC, Bank Trends 02-05, June 2002. Moody's estimates of the telecom share of corporate bond defaults: 13.3% in 2000 and 25.8% in 2001. · fdic.gov
- Company bankruptcy filings: Winstar 8-K (Apr. 2001); 360networks 6-K (June 2001; March-end assets $6.3bn, liabilities $3.6bn); Global Crossing (Jan. 2002; assets $22.44bn, liabilities $12.39bn, per Mar. 21, 2002 U.S. House Financial Services testimony); McLeodUSA 8-K (Jan. 2002); FLAG Telecom 8-K (Apr. 2002); Williams Communications 8-K (Apr. 2002; plan to reduce debt by about $6bn); XO 8-K (June 2002); WorldCom 8-K (July 2002) and Q1 10-Q (assets $103.8bn, long-term debt $29.3bn); Genuity 8-K (Nov. 2002; $242m asset sale to Level 3). · 360networks · House testimony · WorldCom 10-Q · Williams · Genuity
- SEC, In the Matter of Global Crossing, administrative order, Release No. 34-51517, Apr. 11, 2005. Reciprocal transactions accounted for 91% of Q2 2001 IRU sales; excluding them, 'Recurring Adjusted EBITDA' would have been −$43m rather than +$472m; cease-and-desist order for disclosure violations. · sec.gov
- SEC, Press Release 2004-148, Oct. 21, 2004. Qwest fraud charges, more than $3.8bn of improperly recognized revenue, $250m civil penalty, and reciprocal IRU transactions. · sec.gov
- A. M. Odlyzko, 'Internet growth: Myth and reality, use and abuse,' iMP, Nov. 2000. Traffic growth returned to about 100% annually after early 1997; described annual growth rates of 70–150% as roughly 'doubling.' · umn.edu
- Bank for International Settlements (BIS), Quarterly Review, Nov. 2000. Telecom syndicated loans totaled $197bn in January–September 2000 (about 25% of all syndicated loans); international telecom corporate bond issuance reached $94bn (27% of corporate issuance). · bis.org
- J. Thompson, 'Is Telecom Disconnected or Just on Hold?' Federal Reserve Bank of Dallas, Southwest Economy, Jan./Feb. 2002. S&P telecom equipment index −86% ($793bn market value), long distance −65%, telecom services −43% from March 2000 to September 2001; long-haul backbone utilization below 3% in April 2001 (Merrill Lynch estimate). · dallasfed.org
- Light Reading, 'Fiber Utilization Figures Challenged,' July 19, 2001. The Merrill Lynch report's 2.7% figure referred to utilization of lit capacity in operators' networks. · lightreading.com
- J. Kraushaar, FCC Common Carrier Bureau, 'Fiber Deployment Update, End of Year 1996,' Aug. 1997. Equipped/in-service AT&T fiber share: 53.1%. · FCC document copy
- A. M. Odlyzko, 'Internet traffic growth: Sources and implications,' Proceedings of SPIE 5247, 2003. The bust reflected overinvestment rather than slowing traffic growth. · umn.edu
- SK Securities, 'Strategy Comment—Bubbles End When Demand Starts to Decline,' Oct. 27, 2026. Nvidia B200 hourly rental rates rose 14% in one month.
- CoreWeave, FY2025 Annual Report (10-K). Revenue from committed contracts increased from 96% in 2024 to more than 98% in 2025; risks from SLA credits, refunds and termination; weighted-average committed contract term about five years. · SEC
- IREN, '$3.65 Billion GPU Financing for Microsoft Contract,' June 1, 2026. $2.1bn private notes with fixed equivalent to SOFR + 213bp and $1.55bn delayed-draw term loan at SOFR + 225bp; blended cost about 6.0%; Fitch A/DBRS A(low). · iren
- IREN, FY2026 earnings, Aug. 27, 2026. GPU financing for deployments serving below-investment-grade customers, with a 9.0% fixed cost on McKenzie expansion. · globenewswire
- CoreWeave, $8.5bn DDTL 4.0 financing announcement, Mar. 30, 2026. Floating SOFR + 225bp, fixed tranche around 5.9%, loan rated Moody's A3, maturity 2032. · businesswire
- Lucent Technologies, FY2002 Annual Report (10-K Exhibit 13). Customer-financing commitments: $8.1bn (Sep. 2000), $5.3bn (Sep. 2001), $1.3bn (Sep. 2002); customer-financing losses approximately $1.8bn in FY2001 and $765m in FY2002. · SEC
- Nortel Networks, 2000 Annual Report (10-K405) and Q3 2001 management discussion. Undrawn customer financing up to about $4.1bn at year-end 2000; provisions of approximately $1.067bn in January–September 2001. · SEC · SEC
- Cisco Systems, FY2001 Annual Report (10-K Exhibit 13). Customer loan commitments about $1.9bn (July 2001); inventory write-down $2.25bn. · SEC
- Nvidia, Form 8-K dated Aug. 17, 2026. Guarantee of leases to an OpenAI affiliate by SB Energy at Portsmouth, Ohio, covering about 4.25GW of IT load. Covers the shortfall between minimum lease value and re-leasing/sale recovery if OpenAI becomes insolvent, defaults on rent or cannot perform; cumulative cap $105bn. Becomes effective when leases commence, with facility readiness expected from 2028; OpenAI indemnifies Nvidia; termination upon specified credit-rating conditions. Q2 10-Q (quarter ended July 26): credit support limited to specified lease and power payments for land, electrical and building-shell infrastructure, not whole-site costs or all tenant liabilities; amounts increase as nine data centers open and decline as rent is paid; lease term twenty years. · 8-K · 10-Q
- Meta, Hyperion data-center joint-venture announcement, Oct. 21, 2025. Funds managed by Blue Owl own 80%, Meta 20%; total development costs approximately $27bn; Meta guarantees residual value for the first sixteen years of operation. · investor.atmeta.com
- SpaceX, press release and 8-K, Oct. 26, 2026. Commitment to about $40bn in GPU finance leases with an Apollo-managed vehicle (definitive agreement expected by year-end; Nvidia Vera Rubin hardware cited in release); each tranche lasts about 5–5.25 years, with leases expected to commence in H1 2027. · SEC
- DS Investment & Securities, 'Read Together: Anthropic's S-1 and Hyperscaler Earnings,' Nov. 4, 2026. Anthropic Direct and the AI model developer's increased exposure to data-center construction risk.
- Anthropic, PBC, amended registration statement, Form S-1/A (Amendment No. 1), Nov. 2, 2026. As of Sep. 30 and thereafter, commitments totaled approximately $895.4bn (roughly $447.5bn before end-2030; $69.8bn within twelve months); about $213.8bn subject to termination rights or conditions. January–September revenue $33.2bn; remaining obligations mostly 2031–2045. Revenue predominantly usage-based with limited long-term commitments, while compute and infrastructure commitments are multi-year; many key customers lack long-term contracts. Uncommenced leases increased from $50.2bn at end-2025 to $451.2bn at Sep. 30. Nine-month operating cash flow $6.6bn and financing inflows $110.7bn; Q2 operating income $974m and Q3 operating loss $567m. Revolver $15bn, secured by substantially all assets (conditional release possible at the company's option after investment-grade ratings from at least two of three agencies). Five-year Broadcom equipment leases of about $161.2bn excluding interest, with a partial transfer to an SPV in October. IPO price range not stated. · SEC
- Reuters (republished by Khaleej Times), 'Wall Street banks, Blackstone line up $63 billion to finance Broadcom AI chips for Anthropic,' Oct. 27, 2026. SPV to purchase Broadcom equipment and lease it to Anthropic; Bank of America, Citi and Morgan Stanley in senior tranches, Blackstone in junior tranches; $63bn financing; Broadcom residual-value support. · khaleejtimes.com
- Office of the Texas Governor press release (Aug. 3, 2026: about 474GW of large-load requests, about 90% for data centers); ERCOT September Monthly Operations Report (9,456MW approved for energization; 4,132MW observed across approved loads, the sum of each load's monthly maximum, not a simultaneous peak); ERCOT's Aug. 20, 2026 report to the PUCT (pause on energization approvals for data centers of at least 75MW, Dec. 10 verification report, Dec. 17 PUCT meeting). · gov.texas.gov · ERCOT September report · ERCOT·PUCT
- iM Securities, '[Semiconductors] STRESS TEST,' in-depth report, Nov. 4, 2026, pp. 24–32. Three-cloud incremental operating profit/incremental depreciation: 0.5× in Q3 2025 to 2.1× in Q3 2026 (1.6–2.0× benchmark is iM's assumption); if an approximately $200bn 2027 funding shortfall were all debt-financed, projected net debt/EBITDA: Microsoft 0.4×, Alphabet −0.2×, Amazon 1.2×, Meta 0.9×, Oracle 4.4×; hyperscaler ROIC and WACC.
- Daishin Securities, 'November Equity Outlook and Strategy—Valuation Sets the Floor, AI Numbers Drive the Rebound,' Nov. 2, 2026. Assigned a 60% probability to AI credit trouble remaining company-specific; contagion watch levels of 5.30% on the U.S. 10-year yield and 450bp high-yield spreads; semiconductors represent 78% of projected 2027 KOSPI operating profits.
- U.S. Department of the Treasury, Daily Treasury Par Yield Curve Rates. Ten-year rate 4.90% on Nov. 4, 2026 (4.95% Nov. 3; 4.91% Nov. 2). · treasury.gov
- ICE BofA US High Yield Index Option-Adjusted Spread(FRED BAMLH0A0HYM2) · 2026.11.3 3.40%p(11.2 3.43%p) · FRED
- Yuanta Securities, '2027 U.S.-China Equity Market Outlook and Strategy—Extremes Meet,' Oct. 30, 2026. Estimated $3.9tn in revenue needed to justify $1.1tn of 2027 capex.
- iM Securities, 'FX Outlook—Dollar at a Crossroads: Midterms, AI Credit Concerns, France and Iran Risks,' Nov. 3, 2026. CoreWeave CDS below its prior peak by early November.
- IREN, 'IREN to Release Q1 FY27 Results on November 5, 2026.' Results call at 5 p.m. U.S. time November 5 (morning of November 6 in Korea). · iren
- CoreWeave, Q3 earnings date announcement, Oct. 27, 2026. Results after the U.S. close on November 10 (Korean morning November 11). · investors.coreweave.com
- Nvidia, Q3 earnings date announcement, Oct. 28, 2026. Results after the U.S. close on November 17 (Korean morning November 18). · investor.nvidia.com
This report is for informational purposes to support investment decisions. It does not recommend buying or selling any security. Forecasts and estimates attributed to brokerages represent those firms' own judgments.