Looking at the crisis 25 years ago to understand the current risk of a bubble in the AI ​​field

Author: GMF Research

【Editor's Note】Around 2000, the US telecommunications infrastructure industry, represented by Lucent and Cisco, experienced a transformation from high growth to collapse. In this report, we conduct an in-depth review of the collapse of the telecommunications industry in 2000, especially Lucent, and compare it with today's AI ecosystem in terms of risk dimensions.

We believe that the current AI ecosystem shares three commonalities with that of 2000, and its vulnerability has begun to emerge.However, supply chain finance is an amplifier of the industry cycle rather than a trigger; a decline in capital expenditure demand is the fatal blow. In the short term, considering the Federal Reserve's relatively loose stance and the current supply shortage of GPUs, market enthusiasm is expected to continue for some time.

In the long run, based on the most optimistic current expectations for AI capital expenditure, it means that AI must achieve the greatest success since the Industrial Revolution. This article consists of six parts: Part One reviews the three causes of the 2000 telecom bubble; Part Two briefly describes the three stages and market indicators after the bubble bursts; Part Three discusses in detail the Lucent supply chain finance collapse case; Part Four deconstructs the current circular financing ecosystem in the AI ​​field and analyzes its three major similarities with that of 2000; Part Five discusses four reasons for optimism; and Part Six concludes the discussion. I. The Causes of the Telecom Bubble in 2000 On March 10, 2000, the Nasdaq index climbed to 5048.6 points, surging 40% in just one quarter. At that time, industry giant Cisco's market capitalization once exceeded $500 billion, while Lucent's market capitalization once reached $240 billion. However, technology stock prices subsequently peaked and plummeted. By March 2001, the Nasdaq had fallen by 60% in just one year, and by its lowest point in 2002, the cumulative maximum decline reached 78%, marking the bursting of the telecom and internet bubble. Individual stocks fared even worse. Cisco's stock price plummeted from a high of $80 per share in 2000 to around $15 by the end of 2001, while Lucent's fell from a peak of $65 in 1999 to less than $1 in 2002, a loss of over 98%. The entire telecommunications equipment industry laid off more than 500,000 employees between 2001 and 2003. In retrospect, this bubble was a result of regulatory reforms, technological advancements, and loose monetary policy. Figure 1: Nasdaq Index Data Source: Haver, GMF Research 1) Regulation: The Telecommunications Act Introduced Excessive Competition The 1996 Telecommunications Act was the most direct policy catalyst for the formation of the bubble. The core objective of the bill was to break the monopoly of the Regional Bell Companies (RBOCs) formed after the breakup of AT&T in local telephone services, explicitly declaring that "any company can enter any communications business, and any communications business can compete with any competitor in any market." To achieve this goal, the bill required incumbent operators (ILECs) to allow new entrants to interconnect at "any technically feasible access point," resell services to competing local switching operators (CLECs) at wholesale prices, and provide "non-discriminatory network element unbinding access." These provisions were intended to lower the market entry barrier, but in practice, they reduced it to almost zero, with any company with capital believing it could challenge traditional operators. Within five years of the bill's passage, more than 300 CLECs emerged in the United States, vying for a share of the local telephone market. These new entrants raised tens of billions of dollars from capital markets, promising to defeat traditional carriers with new technologies and better services. However, the problem was that the market size was smaller than the new supply: the U.S. local telephone service was a mature market with annual revenue of approximately $100 billion, with limited growth potential, yet hundreds of players flocked in, trying to grab a slice of the pie. More seriously, most CLECs lacked genuine technological or cost advantages; their business models were built on regulatory arbitrage rather than value creation—leasing incumbent carriers' networks at wholesale prices below retail and then reselling them to end customers at a markup. This model was difficult to profit from due to intense competition. By 2002, more than two-thirds of the CLECs had filed for bankruptcy or been forced out of the market, with the industry accumulating losses exceeding $50 billion. 2) Investment: "Laying more fiber optic cables is never too much" "Internet traffic doubles every 90 days" became the creed of the telecommunications industry at the time. From 1994 to 1996, US internet traffic grew from 16.3 terabits per month to 1500 terabits per month (Odlyzko, 2002). This led to unprecedentedly optimistic expectations for the growth rate of bandwidth and fiber optic demand, and "Internet traffic doubles every 90 days" became the most popular saying at the time. It first appeared in industry reports in the mid-1990s and subsequently spread virally through analyst reports and media coverage. If this estimate is accurate, then demand would grow 16 times annually, and any current capacity would be exhausted within months, making the deployment of more fiber optic cables futile. WorldCom CEO Bernard Ebbers declared at an investor conference, "We are building not for today, but for the needs of the next ten years," and Global Crossing's prospectus promised its fiber optic network would be fully operational by 2005. However, numerous studies at the time pointed out that demand growth wasn't as dramatic as claimed. Bell Labs researcher Andrew Odlyzko, through analysis of actual traffic data, found that while there was indeed a short period of explosive growth—"doubling in 90 days"—in 1995-1996, the actual growth rate of traffic on the US backbone network after 1996 was only 100% per year. He further pointed out that the claim of "doubling in 90 days" stemmed from a misinterpretation of partial data and self-serving propaganda by industry participants. However, when his article was published, the bubble was nearing its peak, and the fear of missing out (FOMO) was strong, so the article did not generate widespread discussion. More fatally, technological advancements led to supply expansion far exceeding demand. The maturity of dense wavelength division multiplexing (DWDM) technology enabled a single optical fiber to simultaneously transmit signals of dozens or even hundreds of different wavelengths. Based on technological evolution from 1995 to 2000, the capacity of DWDM systems surged from the initial 4-8 wavelengths to 128 wavelengths, equivalent to increasing the capacity of a single optical fiber by 16-32 times (in conventional commercial systems), with advanced systems even reaching 128-160 times. The result was an astonishing scale of both fiber optic deployment and waste. Between 1996 and 2001, the United States alone laid over 80 million miles of fiber optic cable—enough to circle the Earth 3,200 times or make 170 round trips to the Moon. However, by 2002, industry analysis showed that only 5-10% of this fiber optic cable was actually operational, with over 90% becoming "dark fiber," permanently buried underground or on the seabed. A 2001 report in Wired magazine sharply pointed out: "The silicon economy obeys the law that supply creates demand. Too bad it's not true for fiber." In hindsight, technological progress did occur, but market misjudgments of the pace and commercialization timeline turned the revolution into a financial disaster. The number of internet users increased from 16 million in 1995 to 360 million in 2000, an average annual growth of 70%. However, user growth did not equate to a corresponding increase in bandwidth demand: early applications were primarily text and image-based; high-bandwidth applications such as streaming video did not become widespread until the mid-to-late 2000s. Demand forecasts in the late 1990s contained two key errors: users would grow indefinitely, and bandwidth demand per user would grow rapidly. In reality, user growth slowed after 2000, and the adoption of high-bandwidth applications lagged behind expectations by 5-7 years. 3) Currency: Greenspan's Failure to Curb "Irrational Exuberance" in Time In the mid-to-late 1990s, the loose monetary environment fueled the telecommunications bubble. Since becoming Chairman of the Federal Reserve in 1987, Greenspan's so-called "Greenspan Put" has gradually become known to the market. This means rapidly lowering interest rates to support the economy and stock market during financial market crises, while remaining on the sidelines during asset bubbles. This policy asymmetry led investors to believe that the downside risk would be borne by the Federal Reserve, thus inclining them towards more risky investment strategies. Meanwhile, the US economy experienced its longest post-war expansion cycle between 1991 and 2000, with unemployment falling below 4% and inflation remaining moderate. Against this backdrop of the "Goldilocks economy," Greenspan kept the federal funds rate relatively low: after a brief rate hike in 1993-1994, the policy rate remained stable between 4.75% and 6% from 1995 to 1999, and the real interest rate (adjusted for inflation) was at a historically low level of 2% to 3%. Notably, Greenspan's stance on bubbles was wavering, and his warnings failed to translate into concrete policy. On December 5, 1996, he delivered his famous "irrational exuberance" speech in Washington, questioning whether stock market valuations had become detached from fundamentals: "How do we determine if irrational exuberance has excessively inflated asset prices?" This statement immediately triggered global stock market turmoil. However, he did not change the relatively loose policy stance. After the Long-Term Capital Management (LTCM) crisis in 1998, Greenspan cut interest rates by 75 basis points, but the market immediately pointed out that this was unnecessary, and long-term US Treasury yields rose sharply instead. These interventions repeatedly conveyed to the market the signal that "the Fed will provide support," further reinforcing investors' risk-taking tendencies. It wasn't until the early 2000s, when the bubble had expanded to its limit, that Greenspan initiated a rate hike cycle, but this ironically became the trigger for the bursting of the bubble. The bursting of the bubble around 2000 can be roughly divided into three stages, each with different signals and characteristics.

1) Macroeconomic: Monetary Tightening and Initial Deterioration in Aggregate Demand

The Shift in Monetary Policy as a Macroeconomic Trigger for the Bubble Burst. On January 14, 2000, Federal Reserve Chairman Alan Greenspan delivered a speech warning of the asset bubble and inflation risks caused by the "wealth effect." The Dow Jones Industrial Average peaked that day, and the Federal Reserve subsequently raised the policy rate from 4.75% to 6.5% in the first half of 2000, an increase of 175 basis points. In March 2000, the Nasdaq index quickly reversed course. Between March and April, internet star stocks Yahoo fell by 30%, eBay and Amazon fell by 40-50%, and Akamai and Commerce One plummeted by more than 70% in a single month. This stock market crash not only impacted investor sentiment, but more importantly, it "destroyed" the equity financing channels of telecom operators, making it difficult for them to continue supporting network construction investment and forcing them to cut capital expenditure plans. Bond financing costs also increased rapidly, widening credit spreads for telecom operators, making it difficult to issue high-yield bonds, and making the business model reliant on debt expansion unsustainable. From a macroeconomic perspective, the significant decline in the manufacturing PMI in August 2000 provided clear evidence of a macroeconomic slowdown. Since the end of 1999, the ISM PMI had officially peaked and begun to decline. However, at the time, this was seen by the industry as a cyclical fluctuation rather than a structural turning point; market expectations for equipment manufacturers' profits had not yet been adjusted, and the S&P 500 even approached a new high in September 2000. However, the PMI then accelerated its decline in the second half of 2000, falling to around 42 in early 2001, which was considered a "deep recession in manufacturing." Only then did the market realize that the downturn in the telecommunications industry was actually a structural deterioration. Figure 2: Federal Reserve Policy Interest Rates and PMI Data Source: Haver, GMF Research Between 1995 and 2000, capital expenditure in the communications sector grew at a compound annual growth rate of 23%, while it declined by 28% between 2000 and 2003. This overinvestment in the early stages led to a prolonged period of stagnation in overall US corporate investment in communications equipment, which did not surpass the 2000 level until 2015, 15 years later. Figure 3: US Corporate Investment in Telecom Equipment (Bil USD) Data Source: Haver, GMF Research Lucent's earnings warning issued on January 6, 2000, was the earliest micro-level warning in the telecom equipment industry chain. The company announced that its Q1 fiscal year 2000 earnings per share were approximately 30% lower than expected. Its stock price plummeted 28% that day, from $72.38 to $52, wiping out over $20 billion in market capitalization in a single day. CEO Richard McGinn admitted, "We are clearly disappointed with our performance this quarter," but a company spokesperson insisted, "This is not an issue of demand...This is a bump in the road." In fact, it wasn't until the second warning on July 20, 2000, that Lucent's management confirmed the demand decline was a structural trend. Demand had fallen by more than 50% from its peak at the beginning of the year. In retrospect, this January warning was the earliest "coal canary" warning in the entire industry chain. Cisco's first missed earnings forecast in February 2001 marked the fall of the equipment giant. The company reported Q2 2001 earnings per share that were only 1 cent less than expected, with revenue of $6.75 billion, below the expected $7-7.2 billion. CEO John Chambers warned that "business in January was more challenging than we anticipated." Just two months earlier, on December 4th, Chambers had boldly declared, "I have never been more optimistic about our entire industry or the future of Cisco." Chief Strategy Officer Mike Volpi still insisted in November 2000, "We haven't seen any sign of a slowdown." During this period, enterprise IT spending and telecom operator capital expenditures plummeted in 2001. Telecom operators drastically reduced their capital expenditures, abruptly halting the investment frenzy of over $500 per person and several thousand dollars accumulated per person in fiber optics investment during 1999-2000. Operators like WorldCom, burdened by substantial prior debt, faced debt pressure and financing difficulties, forcing them to cut capital expenditures by more than 50%. Cisco's order backlog, which reached a record high at the beginning of 2001, had become excess inventory by March, and Lucent's new orders declined by over 40% in fiscal year 2001.

3) Collapse: Inventory Write-downs, Soaring Credit Bad Debts, and Exposure of Accounting Fraud

Cisco announced a $2.25 billion inventory write-off in March 2001, marking the official arrival of the disaster..Cisco was hailed as having the most advanced virtual supply chain and real-time IT systems. However, this system was built on the fatal assumption that demand would continue to grow at an indefinite annual rate of 30-50%. Subsequent analysis showed that Cisco "never built a model to predict events of this scale," and the forecasting system did not incorporate downside scenarios at all. Retired executive Selby Wellman recalled the summer of 2000: "Our forecasts were still incredibly high. We wanted to ensure strong growth, so we placed large orders."The "bullwhip effect" of the supply chain amplified the inventory disaster. Cisco's manufacturing partner, Solectron, CEO Ajay Shah explained the mechanism of order duplication: "People see a shortage and intuitively predict higher. Procurement needs 100 parts, but they know that if they want 100, they'll only get 80, so they want 120 to get 100." Internal case studies show that when a product team actually needed 10,000 units, three independent suppliers each produced 10,000 units, resulting in three times the inventory buildup. Even worse, a large number of products were custom-made for customers and could not be resold, and technological iterations caused obsolete inventory to depreciate rapidly, resulting in massive inventory write-downs. Large-scale bad debts in supplier financing wiped out years of accumulated profits in a single day due to customer defaults. Lucent, Nortel, and Cisco collectively committed over $13 billion in credit to downstream companies. The logic behind this supplier financing model was to lend money to cash-strapped operators during a period of capital market frenzy to purchase their equipment, stimulating short-term revenue growth. Furthermore, it was treated as accounts receivable rather than loans in accounting, masking the inherent credit risk. When large clients like WorldCom and Global Crossing fell into financial distress and ultimately went bankrupt in 2001, a large portion of the equipment manufacturers' accounts receivable turned into bad debts. We will discuss the Lucent case in detail in the next section. The large-scale exposure of accounting fraud sealed the fate of this bubble. WorldCom began recording capital expenditures as expenses at the end of 2000, artificially inflating profits. Internal audits discovered the irregularities in May 2002 and reported them to the board of directors. The SEC launched an investigation in June, and on July 21, WorldCom filed for bankruptcy, becoming the largest bankruptcy case in US history at the time, involving $107 billion in assets and $41 billion in liabilities. Enron concealed its debt and fabricated profits through hundreds of related-party special purpose entities (SPEs), using mark-to-market accounting to immediately recognize estimated profits from contracts over the next 20 years as current revenue. In July 2001, Enron's revenue doubled year-on-year while the industry generally contracted. In August, CEO Jeffrey Skilling abruptly resigned and sold $35 million worth of stock. In October, Enron announced a restatement of its financial statements, reducing its 1997-2000 profits by $586 million, approximately 23%. On December 2, it filed for bankruptcy. During the same period, an SEC investigation revealed that Lucent improperly recognized $679 million in revenue in fiscal year 2000, Nortel overstated its revenue by over $3 billion in 2000-2001, and Qwest Communications overstated its revenue by over $3 billion. Statistics show that between 2000 and 2002, more than 30 major telecommunications and network equipment companies were forced to restate their financial statements, cumulatively reducing profits by over $100 billion. III. A Review of Vendor Financing – A Case Study of Lucent First, it's important to emphasize that vendor credit is not a "monster" nor is it always synonymous with bubbles; it is widely used in capital-intensive industries. For example, in the equipment leasing industry, such as heavy machinery, suppliers like Caterpillar support customer purchases through credit; in the agricultural equipment industry, companies like John Deere use this model to help farmers buy agricultural machinery. DataIntelo estimates that the global vendor finance market will be worth approximately $220 billion in 2024, with a projected CAGR of 8.2% over the next 10 years. That said, the Lucent case is still worth reviewing because it highlights the amplifying effect of vendor credit on cyclical ups and downs. At one point, Lucent's supplier credit accounted for over 10% of its revenue. In 1999, Lucent was the world's largest telecommunications equipment company, with revenues of $38.3 billion and 153,000 employees. To win orders from CLECs (Consumer Technology and Electronic Equipment Companies) from competitors, Lucent adopted a highly aggressive "vendor financing" strategy—the company not only sold equipment but also provided loan financing to its customers. This had two advantages. First, it secured downstream customer demand in the long term, guaranteeing its own orders. Second, it solved the problem of cash shortages and financing difficulties for downstream startup customers. By the end of fiscal year 1999, Lucent had committed up to $7.1 billion in supplier financing to its customers, of which $1.6 billion had been drawn down to purchase Lucent equipment, representing 4.2% of its revenue that year. By 2001, the total amount of customer credit withdrawn had risen to $2.96 billion, accounting for 13.9% of revenue. In other words, for every $100 worth of goods sold, Lucent's own "advance payment" accounted for approximately $14. Figure 4: Lucent Customer Credit Balance (Mil USD) and the Percentage of Withdrawals to Revenue Source: Lucent 10-K, GMF Research From a risk perspective, supplier finance is essentially upstream equipment manufacturers using their own money to achieve their KPIs, effectively transferring downstream revenue risks to themselves. Between 2000 and 2003, at least 47 CLECs filed for bankruptcy, including well-known companies such as Covad, Northpoint, and Winstar. After the bubble burst, equipment manufacturers like Lucent suffered huge losses. Specifically, Lucent experienced a catastrophic "triple write-down." The first write-down was for bad debts on supplier financing. In 1999, Lucent's supplier loan provisions were minimal, only $34 million, less than 0.1% of total revenue. However, starting in 2000, with the successive bankruptcies of star clients WinStar and NorthPoint, Lucent's loan loss provisions for suppliers rose rapidly, soaring to $604 million in 2000, equivalent to 2.1% of its revenue that year. By 2001, this figure had surged by a staggering 250% to $2.12 billion, equivalent to 9.9% of Lucent's total revenue that year, resulting in a net loss of $16.6 billion for Lucent in fiscal year 2001. In terms of the ratio of provisions to outstanding loans, between 2000 and 2002, 46.5%, 71.3%, and 86.7% of Lucent's loans were respectively recorded as provisions. In other words, by 2002, 87 out of every $100 of Lucent's loans had become bad debts. In particular, when a client goes bankrupt, the network equipment used as collateral is almost "worthless" due to the industry downturn and excessive upfront investment, resulting in an exceptionally low loan recovery rate. Figure 5: Lucent's Customer Credit Loan Provision Balance (Mil USD) Data Source: Lucent 10-K, GMF Research Figure 6: Lucent's Supplier Loan Provision Balance as a Percentage of Revenue Data source: Lucent 10-K, GMF Research. The second "write-down" was inventory write-down. The illusory order picture created by supplier financing led Lucent management to severely overestimate the actual demand from downstream CLECs. At the end of 1999, Lucent management hinted to the market and investors that approximately $540 million in undrawn financing commitments were expected to be activated and used to purchase its products within the next 12-18 months. Its first-quarter profit for fiscal year 1999 exceeded expectations, and second-quarter profit doubled, with overall revenue growing by 27% to $38.3 billion and net income reaching $4.8 billion, exceeding market expectations. Amid this optimism, Lucent also stockpiled a large amount of customized optical and wireless equipment inventory for future sales. By the end of fiscal year 1999, Lucent's total inventory reached $4.24 billion. However, this inventory was often highly specific and could not be easily resold to other buyers. Starting in 2000, Lucent's inventory provisions climbed rapidly. In 1999, its inventory provisions balance was only US$709 million, accounting for approximately 1.9% of total revenue. However, this rose to US$892 million in 2000, equivalent to 3.1% of its revenue that year, and surged 103% to US$1.814 billion in 2001, equivalent to 8.5% of Lucent's total revenue that year. In terms of the proportion of reserves to total inventory, between 2000 and 2002, 15.7%, 49.8%, and 109.3% of Lucent's inventory were recorded as reserves, respectively. In 2001 alone, Lucent set aside an additional US$2.4 billion in inventory impairment provisions, equivalent to 42.4% of the previous year's inventory balance. Figure 7: Lucent's Inventory as a Percentage of Revenue Data Source: Lucent 10-K, GMF Research Figure 8: Lucent's Inventory Provision (Mil USD) and its Percentage of Revenue Data source: Lucent 10-K, GMF Research. The third "write-down" is bad debt on accounts receivable. In fiscal years 1999-2000, Lucent stimulated revenue growth through sales of equipment and services on credit. By the end of fiscal year 1999, its net accounts receivable reached $8.8 billion, further increasing to $9.56 billion in 2000, representing 23% and 33% of total revenue, respectively. However, with the financial deterioration of major clients such as Global Crossing and WinStar, and a wave of bankruptcies among downstream customers, Lucent's provisions for accounts receivable increased sharply. In 1999, they were only $320 million, accounting for approximately 0.8% of total revenue, but in 2000 they rose to $500 million, equivalent to 1.5% of its revenue that year. By 2001, receivables had surged 26% to $630 million, equivalent to 3.0% of Lucent's total revenue that year. Between 2000 and 2002, provisions were made for 5.2%, 13.8%, and 19.7% of Lucent's accounts receivable, respectively. Notably, these figures underestimated the true losses, as some receivables were sold by Lucent to third-party financial institutions to embellish its financial statements, and Lucent was also obligated to repurchase or guarantee receivables when customers defaulted. Figure 9: Lucent's Accounts Receivable as a Percentage of Revenue Data Source: Lucent 10-K, GMF Research Combining the three types of write-downs mentioned above, Lucent's total provisions surged from US$1.06 billion at the end of 1999 to US$4.56 billion in 2001, an increase of 330% in just two years, and its proportion of Lucent's revenue skyrocketed from 2.8% to 21.4%. This roughly means that, on average, about 20 cents of every $1 of revenue Lucent earned in 2001 was used to cover various asset impairments. Figure 10: Lucent's Total Provisions for Three Types and Their Percentage of Revenue Data Source: Lucent 10-K, GMF Research The first layer is the AI ​​Model Providers layer, primarily involved in the development, training, and deployment of AI models. Representative companies in this layer include OpenAI, Anthropic, xAI, and Mistral AI. These companies focus on creating large language models (LLMs) and generative AI systems. Their main expenditures are concentrated on the procurement of computing resources, including leasing data center capacity from cloud service providers and ordering GPUs from chip manufacturers; these expenditures often account for more than 70% of total spending. Their main revenue comes from API service subscriptions, enterprise-level solutions, and licensing agreements. The second layer is the Cloud Infrastructure Providers layer, which provides computing infrastructure and services to model providers. Representative companies include Oracle, CoreWeave, Microsoft Azure, and Amazon Web Services (AWS). These companies build and manage data centers, with their main expenses being hardware procurement, energy consumption, and maintenance (data center electricity costs can account for 40-50% of total expenditures). Their main revenue comes from cloud computing leasing and hosting services. The third tier is the Semiconductor Manufacturers tier, which focuses on the core manufacturing of AI hardware, providing the underlying support for computing power. Representative companies include NVIDIA, AMD, Broadcom, and Intel. These companies design and manufacture GPUs, accelerators, and processors, with their main expenses being R&D innovation and manufacturing facility investment. Their main revenue comes from chip sales and related ecosystem services. Figure 11 summarizes the product and service ordering and committed investment among AI companies in the three tiers. In this diagram, the solid blue lines represent arrows indicating the provision of goods (mainly chips) and services (cloud computing services), while the dashed red lines represent arrows indicating investment and financing. Based on this analysis, we believe that, from a risk perspective, the current NVIDIA-OpenAI AI ecosystem shares at least three important similarities with the Lucent-CLECs telecommunications equipment ecosystem of 2000. Figure 11: Current Circulating Financing for AI Suppliers Data Source: Internet, GMF Research 1) Expectations: All based on highly optimistic expectations for downstream companies' capital expenditure and revenue capabilities. In the late 1990s, the narrative of "Internet usage doubling in 90 days" convinced investors that there was no "overinvestment" in the fiber optic field. The result was that CLECs over-built $60 billion in fiber optic capacity, while by the time the bubble burst, less than 5% of the installed fiber was being utilized. Today, we hear many optimistic pronouncements in the market, such as "AI is the new electricity that will reshape all industries" and "The global AI market will reach trillions of dollars in the next few years." Jensen Huang's forecast during Nvidia's Q2 2025 earnings call was: "Over the next five years, we plan to effectively scale into a $3 to $4 trillion AI infrastructure opportunity through Blackwell, Rubin, and beyond. We are only in the early stages of this expansion." Driven by this optimism, leading AI companies are now projecting capital expenditures in the hundreds of billions or even trillions, far exceeding the typical spending levels of individual companies. For example, OpenAI's $300 billion cloud computing deal with Oracle, involving the purchase of 4.5GW of computing power, is one of the largest single cloud service contracts in tech history, exceeding the global semiconductor industry's annual capital expenditure in 2024. OpenAI's purchase of 10GW of GPU systems from Nvidia to build its own data center, estimated at a chip cost of $40 billion/GW, could result in total expenditures exceeding $400 billion. In addition, OpenAI has purchased 6GW and 10GW of chips from AMD and Broadcom respectively, with a total investment of hundreds of billions. In total, the value of these agreements exceeds $1 trillion. For comparison, the total investment in information processing equipment in the US GDP in 2024 was only $500 billion, and the total investment of all enterprises (equipment + factory) was only about $4 trillion. These optimistic expectations may ultimately be realized, but the lesson of history is that the risk often lies not in the ultimate feasibility of the technology, but in the robustness of the financial system supporting the technology ecosystem and the vulnerability of downstream customers lacking a profit base once a period of headwinds arrives. 2) Financing: Downstream companies are burning through cash and are highly dependent on supply chain financing. Back then, downstream companies like CLECs generally relied on supply chain financing and other financing methods. In 1999, CLECs' capital expenditures reached a staggering $15.1 billion, while their revenue was only $6.3 billion. This meant that for every $1 they earned, they spent $2.00 billion on construction, with over 60% of construction funding requiring external financing or debt. Today, astronomical capital expenditures far exceed the revenue capacity of large-scale enterprises. Take OpenAI as an example: its projected revenue in 2025 is approximately $13 billion, while its expected capital expenditures over the next 10 years are as high as $1 trillion. This forces it to rely on "supply chain financing" from upstream companies, thus forming a circular financing ecosystem. Currently, the circular financing ecosystem in the AI ​​field can be roughly divided into two types. One is direct investment. Nvidia, as the player with the most abundant cash flow in the AI ​​field, adopts a strategy of directly investing in its customers through equity investments. Its (committed) investments in OpenAI and CoreWeave are as high as $100 billion and $3 billion respectively, to secure priority for hardware orders and provide customers with some of the necessary funding. Simultaneously, it also provided CoreWeave with an additional $6.3 billion in guaranteed computing power, equivalent to purchasing the computing power corresponding to its own manufactured chips. Another approach is stock-for-order. Players like AMD and CoreWeave, with relatively scarce funds and weaker market positions, adopt a "stock-for-order" strategy. To seize market share from Nvidia, AMD issued 160 million low-priced warrants to OpenAI, giving OpenAI approximately 10% of AMD's equity. Essentially, AMD is exchanging future equity dilution for current orders and market position. These warrants are vested in batches, linked to the deployment progress of 6GW GPUs, with the final batch only fully released when AMD's stock price reaches $600. Based on AMD's current market capitalization, this batch of warrants is worth approximately $30-40 billion. Similarly, CoreWeave also offered $350 million in equity as an incentive for its major customer, OpenAI, in the agreement. 3) Concentration: Upstream companies have a very high concentration in a single industry. For the telecommunications industry at that time, although downstream CLECs (Consumer Classified Economies) customers were relatively dispersed, they were actually highly homogeneous. Looking at Lucent's 1999 financial report, its top three customers (AT&T, Verizon, and Sprint) contributed a total of 37% of its revenue. The remaining portion mainly relied on CLECs such as WinStar, NorthPoint, and Covad. Although these customers were numerous, their financial models and industry risk exposures were highly similar, and a wave of bankruptcies would occur if they went bankrupt. By 2002, 23 telecommunications companies had gone bankrupt, with total assets exceeding $100 billion. Currently, the AI ​​field resembles an oligopoly, with computing power demand rapidly concentrating on OpenAI. Taking NVIDIA as an example, its two largest customers account for 39% of revenue, and its four largest customers account for 46%, a concentration significantly higher than that of Lucent in its early days. In particular, recent reports indicate that OpenAI's share of total computing power demand is rapidly increasing. Currently, OpenAI has committed to purchasing over 50GW of computing power from upstream suppliers, including a 10GW GPU deployment agreement with NVIDIA, a 6GW GPU supply agreement with AMD, a 10GW custom accelerator order with Broadcom, 4.5GW of cloud computing services with Oracle, and over 20GW of computing power from CoreWeave. Sam Altman stated in an interview that he hopes to achieve over 250 gigawatts of computing power by 2033, equivalent to one-third of the United States' electricity consumption. In comparison, the power demand of US data centers (including artificial intelligence) in 2024 is approximately 30-50 gigawatts. In other words, OpenAI's committed computing power alone would double the computing power of US data centers. It is evident that whether it's the highly competitive but homogeneous CLECs of the past, or the oligopolistic customers providing huge orders today, once faced with industry downturns, failed technological iterations, or funding shortages, the impact on upstream suppliers will be singular and fatal. V. Four Reasons for Optimism While there are similarities to 2000, there are also reasons to be more optimistic about AI today. First, the financial situation of today's AI industry's model giants is better than that of the CLECs in 2000. CLECs provide standardized fiber optic connections and voice/data transmission services, lacking a competitive moat. The only competitive factors are network scale and price, leading to price wars and collapsing profit margins. As a result, CLECs generally have poor financial health, with an average debt-to-equity ratio exceeding 5:1, and most are experiencing negative cash flow and losses, heavily reliant on external funding to maintain operations. Compared to CLECs, OpenAI, while also operating at a loss, has a deeper moat, stronger monopolistic power, and better revenue performance. Other model giants generally have powerful backers; for example, Anthropic is backed by Amazon and Google, and xAI has Elon Musk's support, making their financial situation relatively more optimistic. Specifically, although downstream AI companies rely on upstream supply chain financing, it's primarily in the form of equity rather than loans like those in 2000 (AMD issued warrants to OpenAI, and NVIDIA made an equity investment in OpenAI). Therefore, it's less likely to see the massive one-time credit provisions seen with companies like Lucent and Cisco. Secondly, we haven't yet witnessed the kind of "predatory lending" seen in 2000. Back then, competition wasn't just fierce among CLECs; upstream manufacturers like Cisco and Lucent also faced intense competition, leading to exceptionally aggressive and high-risk lending policies. Take Harvard Networks, for example, a Boston-based provider of high-speed internet connectivity and fiber optic data transmission services. Harvard Networks began offering digital subscriber line services to enterprise customers in 1996. Two years later, its annual sales were less than $5 million, with annual losses exceeding $1 million and negative net assets. To survive and expand, the young company needed new equipment. It had initially ordered equipment from Paradyne Networks, but ultimately switched to Cisco after Cisco offered up to $120 million in credit, even allowing 25% of the loan to be used for non-Cisco products. By the end of 2000, HarvardNet had gone bankrupt. In contrast, while supply chain financing in the AI ​​field is currently large, competition is not extreme and lacks the predatory nature of that era. Thirdly, from a macro liquidity perspective, the Federal Reserve will maintain a relatively loose monetary policy over the next 12 months and will not choose to "actively burst the bubble." The US job market is showing clear signs of slowing down, with monthly job growth plummeting from around 200,000 in 2024 to around 30,000 currently. This is due to a combination of factors: reduced labor supply from immigration and deportation policies, the substitution effect of AI on entry-level workers, and the "pro-independence" effect of extremely loose fiscal policies. However, on the other hand, AI investment and the wealth effect have kept overall consumption and investment strong, resulting in high GDP growth. The US economy exhibits a significant "two-speed economy" of "cold employment, strong growth." Since the Federal Reserve's dual objectives are price stability and full employment, even with strong economic growth, the Fed is more likely to maintain a loose stance if employment remains weak. In this environment, even if investors know there is a bubble, they may continue to hold risky assets for fear of exiting too early and missing out on losses. In fact, I observed a very interesting signal: the well-known technology investment fund Coatue (one of the Tiger Clubs) cited "don't sell too early" as a key reason for remaining bullish on technology stocks in its October report, whereas a few years ago their bullish logic was "AI will change everything." Figure 12: Coatue's latest report in October 2025. Data source: Coatue, GMF Research. Last but not least, today the usage of GPUs and other computing chips is far higher than that of fiber optic cables back then, so the so-called "dark GPU" does not yet exist. One of the core contradictions of the 2000 bubble was the oversupply of fiber optic cable investment. The industry later discovered that over 90% of deployed fiber optic cables were never activated, earning them the moniker of "dark fiber." In contrast, today's data center GPU utilization is much healthier. Due to the continued explosive growth in demand for AI model training and inference, GPUs remain in short supply in the market. Even for deployed computing power, utilization remains relatively high (especially at the cloud service provider level), with market estimates suggesting an overall data center GPU load of approximately 50-70%. In summary, in the short to medium term, the AI ​​cycle may continue its upward trend, and the underlying problems may not be immediately apparent. VI. Conclusion: The Future of AI The most important question determining whether today's AI boom is a bubble is this: Can the current massive capital expenditure on AI achieve its due return on investment? Here, we provide a simple estimate. If we calculate based on Huang Renxun's most optimistic estimate of 3 trillion USD in total capital expenditure for the AI ​​field, assuming a 30% annual depreciation rate for GPUs (the author does not believe that a depreciation period of 5 years or longer is reasonable), and further assuming that the required ROE for AI investment is 30% (this figure is slightly lower than the current ROE of tech giants), then under a 25% corporate income tax, the required annual EBITDA would be approximately 2 trillion USD. If we further assume that the EBITDA profit margin for AI companies is 30% (equivalent to cloud service providers) to 60% (equivalent to Nvidia), then the corresponding total annual revenue required for the AI ​​industry would be approximately 3.3-6 trillion USD. What does this mean? In 2024, the global automotive industry's total revenue is estimated at approximately $4-5 trillion, the advertising industry's at approximately $800-1 trillion, and the mobile phone industry's at approximately $500 billion. This means that the future revenue scale of the AI ​​industry needs to be on par with the automotive industry, five times that of the advertising industry, or ten times that of the mobile phone industry. In other words, AI must become the greatest success humanity has achieved since the Industrial Revolution to justify the most optimistic expectations. Looking at it from another perspective, assuming there are 1 billion active AI users globally in the future, each user would need to contribute $3,300-6,000 in revenue annually; if the number of active users rises to 3 billion, each user would need to contribute $1,000-2,000 in revenue annually. Such a massive per capita revenue contribution and global user base mean that future AI will likely have to permeate every aspect of people's lives, achieving even greater success than the internet. This is not impossible. However, the most important lesson we learned from the 2000 telecom bubble burst is that bubbles and crashes always go hand in hand with technological progress. As investors, the most important thing is to be prepared for the future and survive until the next cycle.