The Greatest Financial Market Bubble In History Is Forming A Top

Velocity, Not Valuation Defines A Bubble

We’re living through the biggest financial mania in history.

On April 24, the Semiconductor Index (SOX) closed higher for the 18th consecutive session. That’s the longest daily winning streak in the 32-year history of the index.

The move eclipsed a 15-day run in 2014. But that move only produced a 7.8% gain. During the April rally, the SOX rose 47%. The move was led by Intel (INTC), which has gone up 129% in the last six weeks.

Forgive the math, but I need to quantify these extraordinary moves.

The Semiconductor Index (SOX) has annualized volatility of 30% to 35%. Thus, the expected volatility over any 18-day period is approximately 1.2% to 1.5%. A 47% gain in 18 days is 4.1 standard deviations away from the mean. In a normal distribution, this kind of four-sigma event occurs less than 0.01% of the time, or roughly once every 40 to 50 years of trading days.

Something is happening in the markets that is not normal. And what comes next will not be normal, either.

On May 4, the SOX made an all-time high – and then closed below its opening price for the session. A daily reversal in overbought territory – with bearish relative strength index (“RSI”) divergence and a stochastic line-cross at extremes – is a classic sign of a top.

Then, yesterday, May 12, the index dropped 3.01% in a single session to close at 11,717. But the Nasdaq is still reaching new highs. That divergence – leadership rolling over while the broad index extends – is not a bullish setup.

It is the precise dynamic that preceded March 2000.

And guess what? On May 5, the weekly ratio of the VanEck Semiconductor ETF (SMH) to the Nasdaq-100 (QQQ) hit a 26-year peak – its highest level since May 2000, exactly two months after the dot-com bubble peaked in March 2000.

There are two other factors you must understand.

1. This top formed as the economy is experiencing rising inflation and an oil-price shock. April consumer price index (“CPI”) rose 3.8% year-over-year (“YOY”), with energy increasing 17.9% and gasoline 28.4%. Higher inflation will lead to higher bond yields. Falling bond prices will trigger a panic out of financial assets and cause a recession. As I’ve been forecasting all year, the 10-year U.S. Treasury yield will go above 5% and, when that happens, stock prices will fall.

2. What enabled this massive bubble and what will make its collapse far worse than most people expect, is this: credit has been mis-priced for far, far too long. Credit is way too cheap and far too much has been borrowed. The ICE BofA U.S. High Yield Option-Adjusted Spread closed May 11 at 2.79% – roughly half its post-1996 average. The margin debt of FINRA member firms – FINRA oversees brokers and trading in the U.S. – hit $1.22 trillion in March, up 38.7% YOY. Money is virtually free.

Market conditions like these: extremely cheap credit, wildly surging markets, and extremely overvalued stocks have appeared together before.

But only three times.

The U.S. cyclically adjusted price-to-earnings (“CAPE”) ratio currently sits at 40. Three prior comparison points exist in the modern record. In 1929: 21x. In 1972: 21x. At the dot-com peak in March 2000: 35x. The historical mean is 17.

The 2026 reading is the highest sustained CAPE in the history of U.S. equity markets.

But that’s not why it’s a bubble.

Most investors think a bubble is defined by the absolute level of valuation. It isn’t. It’s defined by the velocity of the deviation from trend. When prices move two standard deviations above their normalized growth path – not their nominal high, their trend – the asset has reverted to the trend line every time.

The trend, not the prior high. The arithmetic: a reversion from CAPE 40 to a historical mean CAPE of 17 implies a price decline of roughly 58%.

These mean reversions aren’t pleasant. They are violent because the leverage that drove the move up must be unwound, quickly. The selling isn’t conviction selling. It’s margin selling. Which brings us to credit.

The financial press treats the artificial intelligence (“AI”) rally as an earnings story. It is not. It is a credit story, and the credit cycle just turned.

The four largest hyperscalers – Amazon (AMZN), Alphabet (GOOG), Meta (META), and Microsoft (MSFT) – have collectively guided to between $610 billion and $725 billion of capital expenditure in 2026. Total data center spending is currently projected to reach $3 trillion by 2029.

That’s roughly 10% of U.S. GDP!

Some of that is being paid from cash flow. But most is not.

Data center debt issuance hit $625 billion in 2025, 4x the $166 billion issued in 2023. Oracle (ORCL) alone has accumulated roughly $100 billion of debt. It closed a $16 billion financing on a single Michigan data center in April. Meta is preparing as much as $25 billion in new investment-grade bonds.

CoreWeave (CRWV) is the cleanest case study of how financing has built this bubble. It has raised $28 billion in equity and debt in the last 12 months.

The company closed an $8.5 billion delayed-draw term loan in March. That was the first investment-grade financing in history secured by GPU hardware and customer contracts, rated A3 by Moody’s. Before March 2026, no one had ever managed to convince Moody’s to give an investment-grade rating to a loan backed only by GPU chips. Historically, GPUs were considered too volatile, too short-lived, and too easily made obsolete to support an investment-grade rating. And Moody’s gave this line of credit an A3 rating – that’s three notches into investment grade. That’s the kind utilities and railroads get.

The bond matures in March 2032. That’s a six-year maturity against assets (Hopper-generation GPUs) that Nvidia (NVDA) is already several development cycles past. This is a six-year loan against an asset with a two-to-four-year life span. The cash flows depend on payments from a company (OpenAI) that, by its own backers’ projections, is going to lose $35 billion in 2027.

OpenAI was valued at $30 billion in January 2023. Today: $750 billion. Anthropic: $4 billion in April 2022, $350 billion today. Total AI venture capital raised in 2025 alone: more than $200 billion – 60% of all U.S. venture capital deployed that year.

Where did all of the money come from…?

The Federal Reserve has held real interest rates suppressed for the better part of two decades. Cheap credit creates the illusion that capital is infinite. When capital appears infinite, every project looks accretive at the margin – including the marginal data center that will never earn its cost of capital. This is malinvestment in the precise sense that Austrian economist Ludwig von Mises used the word.

The signal price – the price of money – is broken. So entrepreneurs allocate capital toward projects whose returns are visible only under the false price. When the price normalizes, the projects don’t just become marginal. They become uneconomic in bulk.

And nothing creates malinvestment – aka, bubbles – like revolutionary technology!

Let’s compare the AI bubble to the railroad bubble. Total U.S. railroad track mileage more than doubled between 1850 and 1857. Guess how?

The number of chartered American banks doubled over roughly the same period. Outstanding bank loans jumped 36% in a single year, 1854. The contemporary press identified the cause with characteristic frankness. The New York Herald in June 1857:

Paper bubbles of all descriptions, a general scramble for western lands… What can be the end of all this but another general collapse like that of 1837, only upon a much grander scale?

The trigger arrived August 24, 1857: a single bank branch – the Ohio Life Insurance and Trust Company’s New York office – told depositors it could not pay them their money. The news traveled by Samuel Morse’s telegraph, which had not existed a decade earlier. By the end of 1858, U.S. GDP had fallen 2%, and a generation of railroad lines – Michigan Central, Erie and Pittsburgh, Fort Wayne and Chicago – were in bankruptcy.

The railroads were not a fad. They were the most important infrastructure of the 19th century. They remade the geography of American commerce. They were also, in 1857, dramatically overbuilt against the actual demand for commercial and passenger traffic. The technology was correct. The credit-financed pace of deployment was not.

That is the definition of malinvestment. The underlying productive innovation is real. The capital allocation against it is not. The error is one of velocity, not direction.

We saw this play out again with radio. (For the kids reading who have never used a radio, this was an early form of wireless social networking.)

RCA was the Nvidia of the Jazz Age. The stock traded at $5.825 in 1921 and at $420 by 1928, before a 5-for-1 split. Post-split, RCA peaked at $114.75 in September 1929. By May 1932, it traded at $2.625 – a decline of 98%. The 1920s gain was 200-fold. The collapse erased it entirely.

RCA had real cash flows. Total U.S. radio equipment sales rose 14-fold from $60 million in 1922 to $843 million in 1929 – an expansion justified by consumer adoption. The technology was, again, correct. But the financing and the velocity weren’t.

The 1929 mania was financed by what was then referred to as “call money” – broker margin loans that could be revoked at any time. When the call came, the leverage worked in reverse. The parallel to today’s margin debt – $1.22 trillion, up 38.7% in 12 months – isn’t an analogy. It is the same phenomenon.

And as I’ve told you before, the era that I believe is most like our current macroeconomic situation: the Nifty Fifty of 1972.

The Nifty Fifty were not speculative junk. They were the most successful operating businesses in America. Through 1972, the group averaged over 22% annual earnings growth across the prior five years. They were good businesses: average return on equity was over 22% These were IBM, Kodak, Coca-Cola, Johnson & Johnson, Xerox, Procter & Gamble, and others – the dominant franchises of their era. They were “one decision” stocks. All you had to do was decide to buy them and your fortune was assured.

At year-end 1972, the average P/E of the Nifty Fifty was about 43. They had compounded 28% per year for five years. Then OPEC raised oil prices in October 1973.

From 1973 through 1977, the Nifty Fifty compounded at negative 4.4% annually. And within the Nifty Fifty, the stocks with the highest starting P/E ratios produced the lowest subsequent returns. Value matters. Maybe not this week. Maybe not this month. But over time, it is the most important variable. The shareholders who paid 1972 prices did not recover for a generation.

Finally, there’s the tech bubble most of you lived through.

In March 2000, Cisco Systems (CSCO) briefly became the most valuable company in the world, with a market capitalization north of $500 billion at peak. P/E ratio: near 200. By 2001, Cisco had lost 77.4% of that market capitalization. It took 25 years – until December 10, 2025 – for the stock to close at a new record high.

Sun Microsystems peaked at roughly $200 billion in market capitalization. Oracle bought what was left for $7 billion in 2010. Scott McNealy, Sun’s co-founder, said this in a 2002 Business Week interview:

At 10 times revenue, to give you a 10-year payback, I have to pay you 100% of revenue for 10 straight years in dividends. That assumes I can get that by my shareholders. That assumes I have zero cost of goods sold, which is very hard for a computer company. That assumes zero expenses, which is really hard with 39,000 employees. That assumes I pay no taxes, which is very hard. And that assumes you pay no taxes on your dividends, which is kind of illegal… What were you thinking?

Sun’s investors at 10x revenue lost 95% of their money. Palantir Technologies (PLTR) today trades at over 100x sales. Tesla (TSLA) trades at over 300x earnings – and earnings are down 61% with negative revenue growth.

What led to the collapse of the 2000-era tech bubble: velocity and credit led to enormous amounts of unused capacity.

In late 2002, consulting firm TeleGeography estimated that of the long-haul fiber installed during the boom in Europe and North America, only 10% was carrying any signal at all. And among the fibers that were lit, only 10% of available wavelengths were in use. Roughly 99% of the installed capacity was unused.

The credit-financed buildout – premised on the claim, repeated by every analyst on the Street, that internet traffic was doubling every 100 days – was off by almost two orders of magnitude. Actual internet traffic doubled every 12 months. The claim was wrong by a factor of 30.

The technology worked. And it changed the world over the next 20 years. But the capital deployed against that growth was off by 30x. The picks-and-shovels providers – Nortel, Lucent, Cisco, JDS Uniphase, Sun – collapsed.

Now overlay that on what’s happening today.

Investors have stretched server depreciation schedules from the three years they used in 2020 to six years today. That accounting choice added billions to reported earnings.

Meta has extended the useful life of its servers three times in three years – from 4.0 to 4.5 to 5.0 to 5.5 years – at each step booking a non-cash reduction in depreciation expense. In January 2025 alone, Meta’s extension reduced reported depreciation by $2.9 billion.

In the same quarter, Amazon did the opposite. It shortened server life from six years back to five, took a $700 million operating income hit. Amazon’s CFO cited “the increased pace of technology development, particularly in the area of artificial intelligence and machine learning.”

The same physical asset – a GPU rack – is being depreciated over five years at one hyperscaler and 5.5 years at another. The accounting choice produces a $3.6 billion earnings quality gap. The asset itself is identical.

The reality is that Nvidia is on an annual product cadence now: Hopper 2022, Blackwell 2024, Rubin 2026, Rubin Ultra 2027. Hopper rental rates have already fallen 70% in 18 months – from roughly $8 to $10 per GPU-hour in 2024 to $2.85 to $3.50 today.

Hyperscalers are reporting earnings based on the assumption that today’s GPU will be earning revenue in 2030. The market is paying valuations based on those earnings.

Meanwhile, Nvidia’s product roadmap explicitly assumes the same GPU will be functionally obsolete for its primary workload by 2028. Both things cannot be true. The reconciliation will be a write-down.

And then there is the cash flow itself.

Microsoft’s remaining performance obligations – the contracted future revenue commitments on which its 26% cloud growth narrative depends – total $625 billion. Of that, 45% is concentrated in a single customer: OpenAI. OpenAI will lose $17 billion in 2026 and $35 billion in 2027. Microsoft has invested billions in OpenAI, so that OpenAI can pay Microsoft for Azure. Likewise, Nvidia has committed to invest $100 billion in OpenAI so OpenAI can buy Nvidia products.

See the problem?

This is the same vendor-financing structure that destroyed Nortel and Lucent in 2001. They sold equipment to telecoms by lending the telecoms the money to buy it. The revenue was booked immediately. When the telecoms went under, the receivables were uncollectible and the inventory was worthless.

Total AI revenue is estimated at less than $50 billion annually. Total AI investment is north of $1 trillion. The ratio is 50 to 1, financed in part by six-year debt against two-year assets.

This is not going to end well.

The Austrian credit-cycle indicators are warning us that the bubble is deflating.

Bitcoin hit an all-time high of $126,198 on October 6, 2025. On May 11, 2026, it traded at $81,224 – down 36% from the high and 22% YOY.

Bitcoin tends to lead the credit cycle on the way up and on the way down. Gold tends to lag credit on the way up and to lead defaults on the way down. The divergence over the last six months – gold making new highs while Bitcoin sells off 36% and equities ripped vertically into a key reversal – is not normal.

It is what the leading edge of a credit contraction looks like.

The 1857 panic produced a 2% GDP decline in 1858 – and eliminated half a generation of railroad capital structure. Tracks that survived were bought for cents on the dollar. Railroad investors who held on the way down lost almost everything. Investors who showed up with cash bought the productive infrastructure of the country for two decades of compounded returns.

The 1929 mania produced an 89% decline in the Dow over 34 months. RCA fell 98%. The technology survived. The shareholders did not.

The 1972 mania broke on the OPEC oil embargo and produced a 50% two-year decline in the highest-quality businesses in America. The franchises continued. The investors who bought at 43x earnings did not recover their purchasing power for a generation.

The 2000 mania produced a 78% decline in the Nasdaq Composite and required 15 years to recover. Cisco took 25 years to recover. The internet was real. The capital allocated to the internet via Cisco’s stock was not.

In every case, the technological premise of the bubble was correct. The price paid for it was not. And in every case, the reversion was not a “breather” or a “pullback.” It was a violent, leverage-driven realignment to a normalized earnings stream.

The trigger is always the marginal piece of credit that someone declined to roll.

In 1857, it was a single Ohio insurance company’s branch office in New York. In 1929, it was call money. In 1973, it was OPEC. In 2000, it was the failure of vendor-financed telecoms to make their interest payments.

Here’s a picture of this bubble that you should put on your wall:

• A 18-day vertical run in semiconductors

• 30% concentration of the index in seven stocks

• $725 billion of capex, in one year, from four companies

• $625 billion of data center debt in a single year

• Six-year A-rated loans against two-year assets

• The 22% YOY decline in Bitcoin into new all-time highs in gold

• CAPE of 40

• The high-yield spread of 2.79%

That’s what a bubble looks like.

And the top is in.