NextFin News - The U.S. stock market is sitting at valuations seen only twice in 150 years, and the engine that powered its climb - a record artificial-intelligence investment boom - is almost certain to lose thrust in 2027. That is the warning from William C. Dudley, the former president of the Federal Reserve Bank of New York, who argues in a new analysis that a sharp correction is likely before the end of 2027, and that it will not take much to trigger it.
The case rests on a rare alignment of valuation signals. The S&P 500's cyclically adjusted price-to-earnings ratio has climbed to roughly 41, a level exceeded only during the dot-com peak of late 1999, when it touched 44.2 before the index fell 43 percent. The market-capitalization-to-GDP measure known as the Buffett indicator has reached about 238 percent, an all-time high that puts the stock market's value at roughly 2.4 times the size of the entire U.S. economy. Forward price-to-earnings multiples sit near 20 times, well above the 30-year average of 17.2 times, while the 30-year Treasury yield recently pushed to 5.33 percent, its highest level since April 2007.
What makes this moment different from a simple "expensive market" story is the concentration of the rally and the fragility of its driver. The boom is being carried by a narrow band of AI-related megacap companies and their suppliers, and its continuation depends on capital spending growing at an accelerating pace. But as Dudley points out, it is the change in investment - not the level - that drives earnings momentum. Once capex growth decelerates, profit growth expectations for hyperscalers and their supply chain come down with it, and the multiple that investors are willing to pay contracts at the same time.
Valuation Signals Rarely Align This Tightly
The first question any bubble argument must answer is whether the metrics are flashing red in isolation or in chorus. Today they are in chorus, and that is what makes the setup unusual.
The Shiller CAPE ratio, which divides prices by ten-year inflation-adjusted average earnings, was running just above 41 in August 2026, according to Robert Shiller's data. The long-run historical average is about 17. Only once - the 44.2 reading in December 1999 - has it been higher. The Buffett indicator, which Warren Buffett once called probably the best single measure of where valuations stand, has reached 238 percent, compared with a low near 33 percent in July 1982 and a 20-year average around 106 percent. At that level, the market is priced for perfection across the entire economy, not just the technology sector.
"It is probably the best single measure of where valuations stand at any given moment."
Warren Buffett, speaking of the market-capitalization-to-GDP ratio in a 2001 interview, gave the measure its name. Today's reading sits far beyond anything his generation of investors saw outside the dot-com era.
The concentration makes the reading more fragile, not less. The largest ten companies in the S&P 500 trade at a price-to-earnings ratio of roughly 31 times, compared with 21 times for the rest of the index, and they account for about 30 percent of the market's earnings - up from below 20 percent at the peak of the tech bubble. Their combined weighting in the index has nearly doubled since 2015 to a record 40.7 percent. A market that depends on a handful of names to justify its aggregate multiple is a market whose valuation rests on a narrow foundation.
Historically, readings at this altitude have not ended well. After the CAPE ratio last peaked in 1999, the S&P 500 lost 43 percent over the following three years, and it took more than a decade for the index to reclaim its peak in real terms. After Japan's Nikkei peaked at a cyclically adjusted multiple above 60 in 1989, the index fell roughly 80 percent and has never recovered its high. The 1929 peak, the 1966 peak, and the 2000 peak all shared the same ingredients: elevated multiples, narrow leadership, and an investment boom in a transformative technology that investors believed had permanently redefined growth.
The AI Boom Contains the Seeds of Its Own Slowdown
The second question is whether this time is different - whether the AI revolution is a structural shift that permanently rewrites the earnings power of the companies leading it, or a cyclical investment boom that will follow the same arc as the railroad, internet, and telecom buildouts before it.
The answer, on the evidence, is both - and that distinction is what makes the market call so dangerous. AI itself is almost certainly a structural shift. The railroads, the internet, and the telecom networks all transformed the economy permanently, and their booms produced genuine, lasting value. Britain's canal mania of the 1790s, the railway mania of the 1840s, the U.S. telecom buildout of the 1990s, and the dot-com boom all ended in busts even though the underlying technologies changed the world. The pattern is not a coincidence; it is the mechanics of capital cycles. The technology survived; the valuations did not.
The transmission mechanism is mechanical. Hyperscalers are spending at a pace projected to exceed $700 billion this year, up roughly 70 percent, and that spending implies about $2 trillion in future revenue to justify the investment. If that revenue fails to materialize, or arrives more slowly than the spending, profit margins compress. And even if the revenue does arrive, the moment capex growth slows - as Dudley expects it to in 2027 - the earnings momentum that supported the multiple evaporates. The change in investment is what matters for growth rates, not the level of investment. A still-large but decelerating capex program produces slower earnings growth, not faster.
This is the second-order trap that most investors are not pricing. The first-order story is simple and widely held: AI spending drives earnings, earnings drive prices. The second-order reality is that the same spending creates the capacity that eventually suppresses the returns it was meant to capture. The companies building the infrastructure are, in effect, financing their customers' ability to compete with them. When that capacity comes online, pricing power erodes across the supply chain, and the margin expansion that powered the rally on the way up becomes margin compression on the way down.
The revenue math sharpens the point. If four hyperscalers are spending roughly $700 billion a year on AI infrastructure, they need that investment to generate returns well above their cost of capital. At a 10 percent hurdle rate, $700 billion of annual investment requires about $70 billion of incremental annual profit just to break even on new capital - and that is before accounting for the trillions in existing capital that also needs to earn its keep. The $2 trillion revenue figure Dudley cites is the order of magnitude needed to make that arithmetic work. If AI application revenue grows more slowly than the infrastructure buildout, the gap shows up as falling returns on invested capital, and falling returns on invested capital is what multiple compression follows.
The Bond Market Is Asking a Harder Question
The third element is the fiscal and rates backdrop, which removes the safety net that cushioned previous valuation excesses. The 30-year Treasury yield recently surged to 5.33 percent, its highest level since April 2007, before the Treasury Department announced it would at least double its buyback operations for 10- to 30-year debt, from $2 billion to at least $4 billion, effective from September 9 through November 4. The move sent yields lower, but it also signaled that the government is now intervening directly to manage long-term borrowing costs.
That intervention does not change the arithmetic. Public debt has crossed $40 trillion for the first time, and the interest cost of servicing it rises with yields. When the risk-free rate offers more than 5 percent on a 30-year bond, the equity risk premium - the extra return investors demand to hold stocks instead of bonds - gets squeezed from both sides. Either stock prices fall to restore the premium, or earnings yields have to rise through price declines. There is no third option.
"The Fed has to take the world as it is."
Dudley said of the fiscal backdrop complicating the central bank's task. The Federal Reserve cannot wish away a bond market that is demanding higher compensation for duration risk, and it cannot easily cut rates while long yields are being pushed up by deficit financing. That leaves equities without the valuation support that lower rates provided through the 2010s and the 2020-2021 pandemic boom.
The mechanism here is the discount rate. A stock's value is the present value of its future cash flows, discounted back at a rate that includes the risk-free rate. When the 30-year risk-free rate moves from 1.5 percent to above 5 percent, the present value of cash flows ten or twenty years out - exactly the cash flows that justify a 40-times earnings multiple - falls sharply even if the cash-flow forecasts themselves do not change. Higher rates do not merely compete with stocks for capital; they mechanically reduce the value of the growth that stocks are priced for.
The Strongest Case Against the Bubble Call
The counter-thesis is serious and deserves more than a strawman. It runs as follows: today's megacap companies are not the profitless dot-com names of 2000. They earn real cash flow, dominate durable platforms, and trade at forward multiples that are elevated but not absurd relative to their growth. The top ten companies' 31-times earnings multiple is well below the 43-times multiple the largest ten stocks commanded at the 2000 peak, and their earnings share of the market - 30 percent - is backed by actual profits rather than promises. If AI delivers even a fraction of its promised productivity gains, today's prices will look cheap in hindsight.
There is also a momentum argument: markets can remain expensive far longer than fundamentals suggest, and a deceleration in capex growth is not the same as an absolute decline. Earnings can still grow while investment growth slows, and multiple compression can be offset by multiple expansion elsewhere in the index if market breadth broadens beyond the megacaps. The S&P 500 is up roughly 14 percent in 2026, and momentum is a powerful force that has already proved a generation of skeptics wrong.
The rebuttal is that the counter-thesis requires two things to hold simultaneously: that revenue catches up to a $2 trillion investment program, and that the multiple does not contract when growth slows. History says both conditions are fragile. In 2000, the largest ten stocks also had real businesses and real earnings - Cisco, Intel, and Microsoft were not fantasies. They simply could not grow into their multiples once the investment cycle turned. The same mechanism is available today: capacity expansion, margin compression, multiple contraction. The names are different; the sequence is not.
The falsifying signal is concrete. If AI hyperscalers report sustained revenue growth above 30 percent annually for the next four quarters while maintaining operating margins above 30 percent, and the S&P 500 forward P/E holds above 22 times, the bubble thesis is wrong. That combination would prove that earnings are genuinely growing into the valuation rather than being financed by capex that will later become excess capacity. If instead revenue growth decelerates below 20 percent or margins compress by more than 300 basis points, the thesis is confirmed.
What to Watch, and Who Is Exposed
The forward look splits cleanly by time horizon. In the short term, sentiment and liquidity can carry the market higher - the S&P 500 is up roughly 14 percent in 2026, and momentum is a powerful force. The medium-term risk is the earnings cycle: the next two quarters of hyperscaler guidance will show whether capex growth is peaking. The long-term structural question - whether AI transforms productivity - is almost certainly positive, but that does not protect equity holders from the investment-cycle bust that typically precedes the productivity payoff.
The base case is a correction of 20 to 30 percent from current levels before the end of 2027, triggered by a deceleration in AI capex growth or a failure of hyperscaler revenue to track investment. The upside case is that breadth broadens beyond the megacaps, earnings grow into the multiples, and the market grinds higher with single-digit returns. The downside case is a faster, deeper repricing if long-term yields break above 5.5 percent while earnings guidance weakens - a scenario in which the bond market and the equity market reprice simultaneously.
The exposure is not evenly distributed. The most vulnerable are the companies whose valuations embed the steepest growth assumptions: the AI chip designers, the data-center equipment suppliers, and the hyperscalers themselves if their spending fails to convert into revenue. The least exposed are companies with low multiples, strong free cash flow, and earnings that do not depend on continued capex acceleration. A correction driven by multiple compression hits the high-multiple end of the market hardest; a correction driven by an earnings recession would be broader and deeper.
Investors should watch three signals: hyperscaler capex guidance quarter by quarter, the 30-year Treasury yield relative to 5.5 percent, and the S&P 500 forward P/E relative to 22 times. The first shows whether the boom is slowing; the second shows whether the bond-market safety net is intact; the third shows whether earnings are growing into the valuation.
The uncomfortable truth is that a transformative technology and a bursting bubble are not mutually exclusive. The railroads changed the world, and railroad stocks still crashed. The internet reshaped commerce, and the Nasdaq still lost 78 percent. AI will almost certainly do the same - reshape the economy while leaving a trail of impaired investments behind it. The market is pricing the transformation; it is not pricing the bust that historically comes with it.
Explore more exclusive insights at nextfin.ai.
