The AI Bubble Question Deserves Better Than a Yes-or-No Answer
Richard Coffin of The Plain Bagel wades into one of the most contested debates in financial markets right now: whether the artificial intelligence boom constitutes a genuine bubble. His analysis lands in a measured middle ground that, while perhaps unsatisfying to those looking for a definitive call, reflects the genuinely uncertain state of play. The strongest contribution here is his insistence on distinguishing between a full-blown financial crisis and the more mundane possibility that markets have simply gotten ahead of themselves.
The Scale Problem
The sheer magnitude of capital flowing into AI is difficult to overstate, and Coffin does an effective job contextualizing numbers that have become almost meaninglessly large. OpenAI alone has committed to roughly $1.5 trillion in deals spanning data centers, compute power, and chip purchases. As Coffin frames it:
OpenAI is planning to consume the power of 26 [nuclear power plants] simply to run its software. And that's just Open AI's own investments.
McKinsey estimates the broader data center buildout will require nearly $7 trillion in capital expenditures over five years, which amounts to roughly a fifth of total American capital expenditure concentrated in a single industry. For companies that are collectively hemorrhaging billions annually, this represents an extraordinary bet on future demand that may or may not materialize.
The revenue gap is perhaps the most sobering element. Bain & Company estimates that AI companies will need $2 trillion in annual revenue by 2030 just to reach profitability. Coffin notes this figure exceeds the combined 2024 revenue of Microsoft, Meta, Alphabet, Amazon, Apple, and Nvidia. Meanwhile, Gartner projects AI revenue reaching only $780 billion by 2030, representing remarkable growth but less than half of what breakeven would require.
The Circular Financing Problem
The most striking visual in Coffin's analysis is the now-infamous chart showing circular financial relationships among AI companies. Nvidia invests in OpenAI, which buys Nvidia chips. Nvidia invests in CoreWeave, which sells compute to OpenAI. OpenAI pays Oracle for data center capacity, and Oracle buys Nvidia chips for those data centers. AMD gives OpenAI stock warrants in exchange for chip purchases, and the resulting stock price increase effectively subsidizes the deal.
It's led to concerns that the so-called pick and shovel seller is actually just paying gold miners to buy its picks and shovels and effectively propping up the space.
This observation carries real weight. If Nvidia is investing in its own customers so they can afford to buy its products, then the company's profit margins may be overstated in ways that traditional accounting does not capture, since investments are not factored into margin calculations. Coffin rightly notes the parallel to vendor financing schemes that became popular with companies like Nortel toward the end of the dot-com bubble, a comparison that should give investors pause.
However, the counterpoint deserves attention too. Nvidia generated $77 billion in operating cash flow over the trailing twelve months while deploying $24 billion in total investing cash flow. Vendor financing represents only a fraction of that investing activity. Bank of America has estimated that vendor financing accounts for a small share of overall AI spending. The circular dynamics are real and concerning, but they are not yet the primary engine driving the market.
Why This Is Not Simply Dot-Com 2.0
Coffin's most valuable contribution is his refusal to lazily map the current situation onto the dot-com template. Several structural differences matter enormously. The trailing price-to-earnings ratio for the S&P 500 sits at roughly 30 times, which is elevated historically but well below the 46 times reached at the dot-com peak. More importantly, as Coffin highlights:
S&P 500 companies are generating three times the cash flow as a share of their valuations versus what they were generating leading up to 2000. And on a relative basis, there are nearly half as many unprofitable technology companies when compared to 2000 when 36% of tech companies were losing money.
The big tech companies underwriting much of the AI buildout hold enormous cash reserves and carry relatively little debt. If an AI startup fails, the fallout is contained in ways it would not have been during the dot-com era, when overleveraged companies with no revenue were going public on a prayer. The macroeconomic backdrop also differs: the dot-com bubble deflated partly because of rising interest rates, while the Federal Reserve is currently on a downward rate trajectory.
There is also the matter of fraud. The dot-com bust was accelerated by widespread accounting manipulation as companies cooked their books to meet investor expectations. Reporting standards have improved materially since then, though Coffin wisely leaves room for the possibility that fraud could still emerge.
The Uncomfortable Middle Ground
Where the analysis could push further is on the question of what "bubble" actually means. Coffin acknowledges this distinction but does not quite commit to it. There is an enormous difference between a systemic financial crisis that wipes out trillions in value and a sector-specific correction where overvalued companies come back to earth while the underlying technology continues to develop. The dot-com crash destroyed the NASDAQ by 80 percent, but the internet went on to become the backbone of the global economy. Many of the best internet companies were founded after the crash, not before it.
Assuming that all these players are being perfectly irrational is just as dangerous as assuming that they're all being perfectly rational.
This is a genuinely useful framing. The 500-plus AI unicorns currently in existence will not all survive, just as thousands of dot-com startups did not. Consolidation is inevitable. But that consolidation does not necessarily mean the technology itself is overhyped. It may simply mean that the market for AI applications is large enough for dozens of winners, not hundreds.
One area that deserves more scrutiny than Coffin gives it is the electricity bottleneck. Nuclear power plant approvals take five years or more, construction another five beyond that. If data center demand outpaces grid capacity, it creates a hard physical constraint that no amount of investor enthusiasm can overcome. Companies like xAI are already skirting regulations to set up on-site generators, which suggests the problem is more immediate than the market is pricing in.
Bottom Line
Coffin delivers a responsible and well-sourced overview of the AI bubble debate that avoids the extremes of both breathless optimism and doom-laden pessimism. The analysis is strongest when it highlights the structural differences between today's market and the dot-com era, particularly the cash-rich balance sheets of major tech companies and the lower proportion of unprofitable firms. It is weakest when it treats OpenAI's self-reported revenue projections at face value without interrogating the company's track record on financial forecasts. The honest answer, as Coffin essentially concludes, is that nobody knows whether this is a crisis in waiting or simply a period of exuberance that will self-correct. Alan Greenspan warned about irrational exuberance three full years before the dot-com bubble burst. Timing markets is a fool's errand, but understanding the risks is not, and on that score, this analysis delivers.