AI Investment Bubble: Debt, Hardware Collateral, and Market Risks
The current investment frenzy surrounding Artificial Intelligence (AI) is frequently compared to the dot-com bubble of the late 1990s and early 2000s. Despite widespread agreement that it is a bubble, it has yet to burst. This article explores the evolving dynamics of the AI industry, the financial mechanisms sustaining it, and potential parallels with AI market corrections in other countries.
The Shifting Financial Landscape of AI
Initially, the AI hype cycle was sustained by the solvency of major players. Many large tech companies were either generating substantial revenue from selling "shovels" (e.g., hardware) or had robust core business operations providing ample cash flow. Decades of accumulated cash reserves further enabled these companies to fund AI commitments without incurring significant debt.
However, the situation has drastically changed in recent months. Even companies with strong core revenue streams, like Google, are experiencing negative cash flow for the first time in decades and are resorting to debt to continue AI development. In the first half of the current year, AI companies reportedly borrowed $489 billion, primarily through high-interest bonds exceeding 10% annually. This amount, adjusted for inflation, surpasses the total subprime lending originated in 2006 during the peak of the real estate bubble.
This new wave of lending is arguably riskier, as it often lacks the tangible collateral seen in real estate. Consequently, hundreds of billions of dollars worth of these bonds, even from top-tier companies, are now considered "junk status." To mitigate this risk, financial instruments like credit default swaps from major insurance firms are being employed, raising concerns about a potential financial bubble.
The Unintuitive Rebalancing of Company Balance Sheets
AI development is inherently expensive, particularly concerning chip procurement and data center construction. While these investments are substantial, they are often recorded as asset reallocations rather than direct expenses. For instance, when a company like Meta purchases a data center, its net balance sheet theoretically remains unchanged, as cash is converted into property, plant, and equipment.
The process is further complicated by circular financing and intricate deals. Companies often invest in complementary or competing labs, and cloud providers build data centers to lease to hyperscalers. Increasingly, debt is being taken on to sustain these investments.
A critical issue arises from the rapid depreciation of computer hardware. Historically, computer components lose value quickly as newer, more powerful alternatives emerge. This means that companies whose assets primarily consist of such technology risk losing purchasing power faster than if they held cash.
Depreciation Practices and Controversy
A significant point of contention has been the depreciation schedules adopted by major AI companies, particularly cloud providers. They have been depreciating computer chips over as long as six years, a period much longer than the typical one to three years for cutting-edge components. This practice makes their paper expenses appear lower and their balance sheets healthier.
However, the unprecedented demand for computer chips, especially RAM, has led to an unusual phenomenon: some outdated hardware is actually appreciating in value. For example, SK Hynix, a memory maker, reported an 80% profit margin in its best quarter ever. Similarly, secondhand Mac Studios, even three-year-old models, are selling for more than their original price due to demand for local Large Language Models (LLMs). Even commercial-grade GPUs, like Nvidia's H100, have seen their rental rates increase by 13% over the past year, despite the hardware not getting newer.
While this might seem beneficial, it could be delaying an inevitable reckoning.
The Interconnected Web of Debt and Collateral
Most new data centers are financed with debt, much of which goes to cloud companies that will not use the infrastructure themselves but will rent it out to other AI companies. This creates a complex web of interdependencies, with even theoretical competitors like SpaceX, Meta, Amazon, Microsoft, and OpenAI renting compute infrastructure from each other.
A key development is the use of graphics cards themselves as collateral for loans. In March, Coreweave secured an $8.5 billion loan backed by Nvidia chips, receiving an investment-grade rating. The surprising resilience in hardware value, coupled with contracts from well-capitalized tech companies, has made lenders more willing to fund these ventures. Coreweave's borrowing costs, for example, dropped from nearly 15% in 2023 to just 2.25% above the risk-free rate in March of this year.
This creates a potentially fragile system: * Lenders provide money, secured by depreciating hardware. * Hardware appreciates due to strong demand from AI companies. * AI companies maintain demand through external funding. * External funding often comes from the same lenders or hardware manufacturers.
Furthermore, a significant amount of this hardware remains in boxes, awaiting installation. Companies are desperate to acquire any available hardware, believing it will hold or appreciate in value, even if they don't immediately need it. The primary risk they perceive is a competitor gaining access to these tools.
This intricate chain, however, is only as strong as its weakest link. A slowdown in user growth or a decrease in demand for AI services could lead to fewer ongoing leases, prompting companies to offload hardware. A flood of hardware onto the open market would cause prices to drop, rendering the "secure" loans suddenly insecure.
The Impact of AI on Core Business Operations
The financial strain on large tech companies is not solely due to AI development costs; the technology is also impacting their core businesses. Google, for instance, has reported negative free cash flow for the second consecutive quarter, a first in its 22-year history as a public company. Google Search, now often providing AI summaries, is seeing ad revenue squeezed, and the cost of hosting content has surged. Search ad growth has slowed from 19% to 12% in two quarters.
The rising costs of hardware and data center maintenance affect platforms like YouTube, which is experiencing an influx of AI-generated content, further increasing hosting expenses. These risks to previously robust core businesses are causing jitters in the market. While stock prices have seen significant corrections, they often only revert to valuations from six months prior.
The Debt Spiral and Credit Default Swaps
The accumulated debt in the system makes a gradual, stable deflation of the bubble difficult. While falling stock prices typically don't directly impact a company's operations, there are exceptions: 1. Employee Compensation: Companies with booming stock prices attract talent through stock-based compensation. 2. Capital Raising: Businesses may sell their own stock to raise capital for operations that aren't self-sustaining.
Many large companies, including Google, are now running cash flow negative and are resorting to selling their own stock and taking on more debt. In June, Google announced an $85 billion equity raise, the largest in American corporate history. While Google can still borrow relatively cheaply (e.g., 5.25% for 19-year euro bonds), other companies face much higher rates. Cloud providers like Coreweave, which enjoyed low-risk premiums months ago, are now paying double the interest, with their latest loans priced at 5.5% above the risk-free rate, up from 2.25% in March. Unsecured lending rates for these companies can exceed 15%.
To mitigate these risks, the financial system is employing credit default swaps (CDSs). These insurance contracts guarantee bond repayment in exchange for a premium. The riskier the bond, the higher the premium. Premiums on lending from companies like Microsoft have more than doubled, while Alphabet's have tripled. Oracle, for example, faces a premium of nearly 300 basis points, meaning lenders demand at least 9% interest.
These high interest expenses further increase the risk profile of these companies, creating a feedback loop. Investors are increasingly choosing between high-yield government bonds (over 5%) or direct equity investments in AI companies, leaving fewer lenders in the middle and driving up interest rates. While some investors are willing to take on high-risk lending for short-term gains, this strategy is vulnerable to any interruption in strong, uninterrupted growth, potentially leading to severe downward spirals.
Lessons from China and South Korea's AI Bubbles
The AI bubbles in China and South Korea have already burst, offering potential insights. The key difference lies in the position of their dominant companies within the AI supply chain. While Western AI companies are primarily developing AI itself, Korean companies like Samsung and SK Hynix are "second-order businesses" manufacturing the components.
These component manufacturers are more susceptible to market fundamentals, such as pricing competition and capped demand. For instance, SK Hynix's stock fell 10% after a Chinese rival announced a smaller memory chip design. It is easier for companies like Anthropic or OpenAI to maintain a narrative of impending superintelligence than for a memory chip maker to claim "super RAM" that would undercut its own supply. Their markets are arguably more grounded in the measurable reality of component sales.
From their peaks, the Korean stock market fell 40% from June to October, far exceeding the NASDAQ 100's 10% correction. These declines occurred amidst existing economic malaise from property bubbles. Despite the significant market corrections, the impact has been largely contained so far. Korea's chip exports tripled year-over-year in October, unemployment remains stable, and retail sales improved.
While these markets differ, their experiences suggest that a slow deflation of the bubble might be possible, avoiding a catastrophic bailout. However, the current financial pressures and reliance on debt in Western markets indicate that a "big red bailout button" might be pressed if things become too volatile.
Takeaways
- AI funding has shifted from cash reserves to $489 billion of high‑interest debt, with bond rates exceeding 10% and surpassing the total subprime lending of 2006.
- Companies are extending depreciation schedules for chips to six years, making balance sheets appear healthier while hardware values sometimes rise unexpectedly.
- Lenders are accepting GPUs as loan collateral, creating a fragile loop where hardware appreciation supports risky borrowing that could unravel if demand falls.
- Core business cash flows are turning negative, exemplified by Google’s first‑ever consecutive quarters of free‑cash‑flow loss and slowing ad revenue.
- The bursts in China and South Korea’s AI bubbles suggest component‑focused firms may deflate more gently, but Western AI debt could trigger a large‑scale bailout if volatility spikes.
Frequently Asked Questions
Why are AI companies borrowing $489 billion at over‑10% interest, and how does that compare to the 2006 subprime lending peak?
AI firms are borrowing $489 billion because cash reserves have dwindled and they need funding for expensive chips and data centers, so they turn to high‑interest bonds that now exceed 10% annually. This volume surpasses the total subprime mortgages originated in 2006, indicating a similarly large credit expansion in a speculative sector.
How does using graphics cards as loan collateral create systemic risk for the AI financing bubble?
Using GPUs as loan collateral ties lenders to assets that depreciate quickly yet are currently appreciating due to AI demand, creating a fragile feedback loop where loan security depends on sustained hardware prices. If AI growth slows, hardware values could fall, turning previously safe loans into risky, potentially default‑prone obligations.
Who is How Money Works on YouTube?
How Money Works is a YouTube channel that publishes videos on a range of topics. Browse more summaries from this channel below.
Does this page include the full transcript of the video?
Yes, the full transcript for this video is available on this page. Click 'Show transcript' in the sidebar to read it.
100's 10% correction. These declines occurred amidst existing economic malaise from property bubbles. Despite the significant market corrections, the impact has been largely contained so far. Korea's chip exports tripled year-over-year in October, unemployment remains stable, and retail sales improved. While these markets differ, their experiences suggest that
slow deflation of the bubble might be possible, avoiding a catastrophic bailout. However, the current financial pressures and reliance on debt in Western markets indicate that a "big red bailout button" might be pressed if things become too volatile.
Helpful resources related to this video
If you want to practice or explore the concepts discussed in the video, these commonly used tools may help.
Links may be affiliate links. We only include resources that are genuinely relevant to the topic.