When Will the AI Bubble Burst? Watch Credit Default Swaps and Token Prices
The AI bubble will eventually burst, but current valuations are far below the dot-com era, with Nvidia's P/E at just 15x. The real risk lies in the credit markets, where credit default swaps have risen to historic highs, signaling a potential sharp correction ahead.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Token prices down 50% but usage up 2.5x
Dan Niles points out that to judge whether the AI bubble has peaked, you need to watch two core variables: token production and unit price. Since the end of May, due to competition from open-source models, token unit prices have fallen about 50%, but production has grown 2.5x in the same period. As long as production growth offsets price declines, industry revenue is still expanding. Meanwhile, the three major cloud providers (AWS, Azure, GCP) saw revenue growth accelerate from 35% in March to 43% in June, with operating margins up about 2 percentage points, indicating the industry's fundamentals remain healthy.
— Dan Niles4x leverage was the main cause of the fund's blowup
The Situational Awareness fund's blowup was not due to wrong positioning but the use of 4x leverage. If the underlying drops 25%, 4x leverage means a 100% loss, triggering liquidation. This echoes the lesson of Long-Term Capital Management (LTCM) in 1998. After hearing about the fund's problems on July 29, Dan Niles judged that a short-term bottom was near because fundamentals (token usage growth) had not deteriorated. He warns that leverage amplifies risk in both directions, and when long and short positions are simultaneously unfavorable, it can trigger a catastrophic sell-off.
— Dan NilesCloud providers' cash flow turns negative for the first time
Despite accelerating revenue and improving margins at cloud providers, the market worries about deteriorating free cash flow. For example, Google's cash flow turned negative for the first time since going public, due to massive AI capital expenditures. Dan Niles believes this resembles the dot-com era: the technological revolution is real, but stock prices can still crash 78% (Nasdaq 2000-2002). He expects strong industry growth for the next year, but if a hyperscaler cuts capital spending due to debt problems, it could trigger a chain reaction.
— Dan NilesData center political risk is the biggest short-term threat
Dan Niles points out that the biggest short-term risk to AI growth is not demand but political resistance. A Gallup poll shows 71% oppose building data centers in their backyards, higher than the 53% who oppose nuclear plants. Both parties are unusually united against data centers, especially in red states like Texas. Hyperscalers have failed to effectively communicate the jobs and tax benefits of data centers, while anti-business forces have successfully cast them in a negative light. He expects political risk to weigh on the market until the November midterm elections.
— Dan NilesChinese semiconductors will break the cyclicality myth
Dan Niles believes that the view that semiconductors are no longer cyclical is dangerous, and China may be the key to breaking this myth. He draws parallels to Japan in the 1980s and South Korea in the 1990s, which entered the DRAM market with government support and eventually displaced US dominance. Today, China's CXMT (长鑫存储) and YMTC (长江存储) are expanding rapidly: CXMT plans to raise monthly capacity from 300,000 wafers to 500,000, and YMTC aims to surpass Samsung and SK Hynix in NAND. For China, semiconductor self-sufficiency matters more than aircraft carriers.
— Dan NilesAgentic AI will extend the cycle by at least a year
Dan Niles believes agentic AI is a new phase, marked by the launch of the OpenClaw project on January 30 this year. The agentic phase consumes 10-100x more tokens than chatbot AI, which will sustain demand growth. He expects the stock market to have at least another year of upside, but warns of competition from Chinese memory chips. He advises investors to watch the data like a hawk and avoid excessive optimism or pessimism.
— Dan NilesSoftware faces AI disruption; security and gaming are relatively safe
Dan Niles believes AI's impact on the software industry is enormous. OpenAI and Anthropic's combined annualized revenue surged from $29 billion at the start of the year to $105 billion seven months later, and that spending must squeeze other software budgets. He favors three relatively safe areas: cybersecurity, recording systems, and video games. Recent moves like Salesforce's partnership with Anthropic and Workday's acquisition by Silver Lake suggest software stocks may be near a bottom, but security valuations are already high.
— Dan NilesCredit markets are the early warning system for the AI bubble
Dan Niles emphasizes that the credit market is key to judging the AI bubble. He monitors daily the credit default swap (CDS) prices of hyperscalers like Microsoft, Google, and Nvidia, and finds they are higher than the North American investment-grade average and at historic highs. Credit investors are more cautious than equity investors because bond returns are fixed, so they must ensure principal safety. AI's massive capital expenditures rely on debt financing, and the 30-year Treasury yield has risen to its highest since 2007, increasing financing costs and potentially triggering the bubble's burst.
— Dan NilesIn their own words · checked verbatim
There is a number of tokens produced and how much you can charge for each of those tokens. And what you've seen since the end of May is what you can charge for these tokens has gone down by about 50%.
Dan Niles4:03
The problem was that they were using 4x leverage. So, if you do the math and say: if you have a drawdown of 25 percent, but a leverage of 4 times, then you lose 100%, and at that point your fund is liquidated.
Dan Niles7:03
71% were against data centers and only 53% were against nuclear power. That's bad, right? I couldn't believe it when I saw it.
Dan Niles14:19
I think many investors greatly underestimate what China is doing. And I hear these arguments and I tell myself that this is exactly what people were thinking back in the 1980s when Japan entered the semiconductor industry.
Dan Niles21:28
The agent phase uses 10-100 times more tokens than the chat-based AI phase. So, I think you still have a long way to go, at least another year for the stock to go up.
Dan Niles27:18
When you have some of the largest companies in the world that generate the largest cash flow that have credit default swaps trading at premiums to North American investment-grade credit default swaps, that tells you about the ability of the capital markets to absorb that.
Dan Niles41:19
Figures
| Token price decline | 50% | 4:03 |
| Token production increase | 2.5x | 4:03 |
| Cloud revenue growth (March) | 35% | 5:03 |
| Cloud revenue growth (June) | 43% | 5:03 |
| Nvidia P/E ratio | 15x | 3:03 |
| Nasdaq decline 2000-2002 | 78% | 10:03 |
| Opposition to data centers | 71% | 14:19 |
| Opposition to nuclear plants | 53% | 14:19 |
| OpenAI and Anthropic annualized revenue (start of year) | $29 billion | 31:18 |
| OpenAI and Anthropic annualized revenue (7 months later) | $105 billion | 31:18 |
Glossary
- Credit Default Swap (CDS)
- A financial derivative used to hedge against the risk of bond default; its price reflects the market's expectation of default probability.
- Hyperscaler
- A giant cloud provider with massive data centers offering cloud services, such as AWS, Azure, and Google Cloud.
- Token
- The basic unit of text processing in AI models; token usage is a key metric for measuring AI demand.
- Agentic AI
- An AI system that can autonomously execute multi-step tasks, consuming far more tokens than chatbot AI.
How to listen
Fund managers focused on the AI investment cycle, tech stock investors, and entrepreneurs seeking to understand the risks and opportunities of the AI bubble.
The opening pleasantries and the closing segment about games can be skipped; the core content starts at the 4-minute mark.