At the DBS Asia Future Forward Conference in Singapore, Broadcom CEO Hock Tan made a statement that left the entire AI investment community rather embarrassed.
Many investors have no idea what they’re investing in… their expectations are wildly inflated, only to find out it’s almost like being in a casino.
Following suit, Intel's CEO Eugene Q. Lee continued, saying, "If there's a challenge ahead, it's capital and the cost of capital; in the long term, people want to see a return on investment."
Chen Fuyang is qualified to say this. Broadcom’s AI semiconductor revenue in Q3 reached $16.7 billion, a 221% year-over-year increase, and he expects this segment to reach $230 billion by fiscal year 2028. Broadcom’s stock has nearly quadrupled over the past two years, with its market cap briefly surpassing $1.7 trillion. He is not an outsider—he is one of the largest suppliers in this AI infrastructure boom.
Precisely because of this, when he warned that "capital could become the biggest bottleneck for AI," the market found it hard to ignore.
The U.S. stock market reacted on Monday. The Nasdaq fell 0.3%, with memory chip giant Micron Technology dropping 6% during trading, Broadcom shares declining about 1% to around $357, and NVIDIA, Intel, and AMD also under pressure.
The market is reevaluating the weakest link in the AI narrative—financing.
An unprecedented lending game
To understand the warnings from these two experts, first understand how AI companies are raising funds right now.
In the past, tech giants used real money earned from advertising and cloud services to build data centers and buy chips. But that phase has ended.
Morgan Stanley expects global AI-related debt issuance to reach nearly $570 billion in 2026, compared to just $217 billion for the full year of 2025. The five major cloud providers—Alphabet, Microsoft, Amazon, Meta, and Oracle—are projected to spend $769.2 billion on capital expenditures in 2026, exceeding their operating cash flow.
What fills the gap? Borrowing.
The way of borrowing is far more complex than most people realize. The most direct method is bond issuance. Alphabet completed a $25 billion U.S. dollar bond offering in August and entered the Australian dollar bond market for the first time. Goldman Sachs estimates that the five largest AI infrastructure companies may issue up to $250 billion in bonds in 2026, rising further to $400 billion in 2027.
But bond issuance is just the tip of the iceberg; deeper changes are occurring off the balance sheet, namely the adoption of GPU-backed loans and special purpose vehicles (SPVs).
Since CoreWeave pioneered the approach at the end of 2023, GPU-backed loans have rapidly become a mainstream financing tool in AI funding. These loans are typically structured by tech companies and investment institutions through SPVs, which borrow funds from the market to purchase chips and then lease them to AI companies, using the cash flow from lease agreements as collateral for repayment.
For tech companies, the benefits are clear: the debt does not appear on their balance sheets, their credit ratings remain unaffected, yet they can still use the chips.
The deal between Broadcom and Anthropic is the ultimate expression of this logic. Last week, Wall Street banks began distributing a $60 billion debt financing package, of which approximately $42 billion consisted of senior secured loans backed by Broadcom’s credit guarantee, and another $18 billion comprised unsecured subordinate debt. The funds are being used by Anthropic to lease AI chips designed by Google. Broadcom’s A-rated credit rating significantly lowered the financing cost of the loan, while Broadcom secured a long-term chip order from a cutting-edge AI lab. More subtly, Broadcom may also receive up to $42 billion in convertible notes from Anthropic to cover lease payments—blurring the lines between supplier and customer.
This type of operation is being replicated on a large scale.
According to an analysis by the Financial Times, over the past 12 months alone, technology giants have provided residual value guarantees—ensuring a minimum future value for chips or data centers—amounting to $300 billion. Morgan Stanley estimates that seven hyperscale cloud providers and chip manufacturers have issued over $3.1 trillion in off-balance-sheet commitments and credit support. These guarantees essentially mean that tech companies are backing financing structures built around their products with their own creditworthiness. If AI demand falls short, computing capacity becomes oversupplied, or technological advancements rapidly devalue existing chips, the conditions triggering these guarantees could be activated.
Is the profit keeping up?
Chen Fuyang is not bearish on AI. He clearly stated that demand for AI computing infrastructure will continue to far exceed supply in the short term, and this imbalance could persist until 2029 or even 2030. Broadcom’s earnings also confirm this: AI semiconductor revenue reached $16.7 billion in the third quarter and is expected to reach $21.7 billion in the fourth quarter.
But the issue lies on the other end. Chen Fuyang noted that OpenAI’s annualized revenue of approximately $50 billion in September demonstrates strong demand, yet its massive training costs make it nearly impossible to justify long-term profitability.
He used a metaphor: "It's like when you were a kid going to the arcade and your parents gave you ten dollars in tokens—you spent them all in ten minutes and learned nothing. That's how companies operate today."
Data from Morgan Stanley shows that global AI-related debt issuance dropped to $23 billion in September, less than half of August’s level and a significant decline from the historic peak of $113 billion in June. Investment-grade AI bond issuance by U.S. blue-chip tech companies has nearly stalled.
Meanwhile, data center construction is facing increasing public resistance.
Oracle's massive data center campus in New Mexico, which faced difficulties accessing electricity, issued a force majeure notice; the market's private trading price for its $18 billion配套 loan has fallen to 85 cents on the dollar. Concerns among U.S. residents about the impact of high-energy facilities on water supply and air quality are emerging as an underappreciated risk factor in AI infrastructure financing.
The warnings from Chen Fuyang and Chen Liwu point to this tightening cycle: AI companies need to continuously borrow money to buy chips and build data centers, but borrowing depends on investors believing these investments will ultimately generate sufficient returns. When this belief begins to waver, financing costs rise, project progress slows, revenue growth comes under pressure, and investor confidence is further eroded in turn.
Bain & Company has calculated the bill for this feast: by 2030, global data center spending will reach $5 trillion to $6.5 trillion, and the AI market would need to generate nearly $6 trillion in annual revenue to support this scale—but so far, no company has proven this figure achievable.
Author: Bear Cookie
Twitter: https://twitter.com/BitpushNewsCN
BitPush Telegram community: https://t.me/BitPushCommunity
BitPush TG subscription: https://t.me/bitpush