The Year AI Ate the Stock Market

The Demo That Moved Trillions

ChatGPT launched on November 30 2022 at the end of the worst year for tech stocks since the financial crisis. Within two months it had over a hundred million users the fastest adoption of any consumer product in history up to that point. It took a few months for the markets to connect the dots and then 2023 became the year the dots were connected all at once. This entry in the series is about how a research demonstration became the single dominant driver of profitability.of stocks because understanding the year 2023 is essential to understanding every market year since then

The Guidance Heard Round the World

The decisive moment came on May 24 2023. Nvidia which designs the graphics processors that turned out to be the picks and shovels of AI reported earnings and guided next quarter revenue to around $11 billion. Wall Street expected about $7 billion. targeting increase Something of that magnitude more than 50 percent above consensus essentially never happens in a company of that size. The stock rose about 24 percent the next day and within a week Nvidia joined the trillion-dollar market cap club

What that guidance revealed was not a forecast but a fact. Cloud giants Microsoft Google Amazon and Meta were already buying AI chips in quantities that no one had modeled because each of them had concluded that they could not afford to lose the race for AI platforms. Capital spending guidance at the big tech companies increased every quarter afterward and every dollar of that capital spending was income for someone first for the chip makers

The Magnificent Seven

By mid-2023 strategists had a name for the business. Magnificent seven Apple Microsoft Alphabet Amazon Nvidia Meta and Tesla a label coined by Bank of America strategist Michael Hartnett. The S&P 500 is up about 24 percent in 2023 and about two-thirds of that gain came from just those seven names. The Nasdaq is up about 43 percent its best year since 1999 a comparison that excited bulls andterrified everyone who remembered what followed 1999

A concentration like that changes even what an index is. The S&P 500 is weighted market capitalizationThe equal-weighted version of the S&P where all 500 companies count equally gained only about 12 percent one of the widest gaps on record

A Worked Example: What Owning the Index Actually Bought You

Concentration statistics are quoted constantly and are almost never decomposed. Doing the decomposition requires two lines and produces a number that is truly surprising

The performance of an index is just a weighted average of its components. So if the seven names contributed two-thirds of a 24 percent gain they generated 16 percentage points and everything else generated the remaining 8

Now divide by the weights. The Magnificent Seven entered 2023 with about 20 percent of the index having been beaten through 2022. Sixteen contribution points from a 20 percent weighting means those seven stocks returned about 16 divided by 0.20 which is 80 percent as a group

The other 493 companies carried the remaining 80 percent of the index and contributed 8 points implying a return of 8 divided by 0.80 which is 10 percent

Check it with a number we already have. The equal-weighted S&P where the average company counts the same as the largest returned about 12 percent. That's the return of a typical stock and it falls right next to the 10 percent implied for the 493. The arithmetic is internally consistent meaning the decomposition tells the truth

Here's what that really means to an investor. Someone who owned the S&P 500 index fund earned 24 percent. Someone who owned every company except those seven earned about 10 percent. The gap is 14 percentage points and is entirely due to a position that almost no one consciously chose to take

Then run the same arithmetic in reverse which is the part worth internalizing. By the end of 2023 those seven names were approaching 30 percent of the index. If they fell 30 percent while all other companies remained stable the index would lose 0.30 times 30 which is equivalent to 9 percentage points before any other stock did anything. The concentration is symmetrical and does not announce which direction it intends to work

These figures are rounded and the initial weight is approximate so treat 80 percent as an order of magnitude rather than a measure. The structure holds regardless: a handful of names with large weights make the index a different instrument than the one described in the prospectus

Bubble or Buildout

The obvious question constantly asked ever since is whether 2023 was 1999 again. The honest answer from 2026 onwards is that the comparison fails in one crucial place. Dot-com leaders fixed their eyes on hope while AI leaders reported on the real revenue and real profit growth that the topic promised Nvidia most dramatically. Valuations grew richer but wealth multiplied by the flow ofReal cash is a different animal than wealth multiplied by nothing

The harder question is the second-order one. Hyperscalers have become the world's biggest spenders of capital and that spending manifests itself as depreciation that drags down their profits for years. Whether the applications built on all those chips generate returns that justify expansion is as of this writing in 2026 an open question and it is the question for anyone subscribing to these stocks today

Case Study: The Nifty Fifty and Being Right About the Company

The bubble comparison people are looking for is 1999 and it's wrong because it allows AI to trade too easily. Most of the dot-com leaders were out of business. A much more uncomfortable precedent is the Nifty Fifty

In the early 1970s American institutional investors converged on a group of about fifty large high-quality growth companies names like IBM Coca Colaprofits

The bear market of 1973 to 1974 caused many of them to fall by sixty to eighty percent. That part is well known and is used as a simple warning

The part that makes it genuinely instructive came later. The group's long-term studies most notably the work of Jeremy Siegel published in the 1990s found that the Nifty Fifty as a portfolio roughly matched the broader market over the next twenty-five years. The thesis was substantially correct. They were extraordinary businesses and they grew. Investors were not wrong about the companies

They were wrong about the price and the punishment for it was twenty-five years of not making excessive profits and being right all the time. There was no moment of vindication no crisis to prove that anyone was foolish just a very long period in which excellent businesses grew to early valuations

If we compare this to AI trading the argument that there are real profits this time seems less protective than it seems. The real profits were exactly what the Nifty Fifty had. The question was never whether the business was good. It was about whether the price already contained a decade of good results and that is a different question that the presence of profits does not answer

Where the Real Earnings Argument Is Weakest

I find the construction argument more persuasive than the bubble argument and there are three respects in which I think it is soft

Income is unusually circular. Nvidia's demand came overwhelmingly from a handful of hyperscalers that funded capital expenditures with profits earned elsewhere. That's real revenue in every accounting sense and it's also a small number of customers spending against a strategic fear rather than a measured return. A vendor whose growth depends on four buyers deciding they can't afford to lose a race is exposed to those four buyers changing their minds simultaneously which is exactly what happens when a race is resolved

Depreciation is doing a silent job. Several large buyers extended the assumed useful lives of their server fleets during this period reducing annual depreciation and increasing reported profits without earning a single additional dollar. If the actual economic life of an AI accelerator turns out to be shorter than the accounting life current earnings are overstated and the correction comes as a wave of writedowns rather than a loss of revenue

The telecommunications precedent is closer than that of dot coms. The development of fiber in the late 1990s was not a story of companies without revenue. The operators had real customers real cash flow and they built real infrastructure that we still use. They also built much more capacity than short-term demand justified the vast majority of the fiber installed was turned off for years and the capital was destroyed anyway. Global Crossing went bankrupt in 2002. Nortel collapsed. The infrastructure was really valuable and it was valuable to whoever bought it secondplace at a fraction of the cost of construction. That a construction is real is not the same as a construction that is profitable for the people who financed it

My own position is that the AI ​​investment cycle is producing genuine value and that the distribution of who captures it is much less defined than current prices imply. That is an opinion not advice and I hold it with the humility appropriate to writing about an unfinished story

What 2023 Set in Motion

Three lasting consequences. First the market's center of gravity permanently shifted toward a handful of mega-caps changing the mathematics of indices options flows and how much one earnings report can move the entire tape. Second an entire physical economy emerged around artificial intelligence data centers power contracts cooling and networking topics covered elsewhere on this site which is where the topic meets real estate and utilities. Thirdevery company in America acquired an AI slide on their investor platform and separating the genuinely transformed companies from those that borrowed the multiple became a core skill of analysts a skill this site spends a lot of time practicing

How I Actually Track Concentration

Since 2023 I have maintained a brief routine for this mainly because the index number stopped being an honest summary of the market

The first thing I look at is the gap between the cap-weighted and equal-weighted versions of the same index. It's a subtraction and tells me immediately whether I'm looking at one market or a handful of stocks. In 2023 that gap was about 12 points which is huge and was visible in January if anyone had bothered to check

Second I do the decomposition from the example above every time a headline cites a contribution statistic. Contribution divided by weight gives the implied performance of the group and turns a vague statement about tightness into a number I can compare over the years

Third I look at the weight of the top ten rather than the top seven because the membership of these groups changes and the label follows performance. Naming a group after stocks that have already risen ensures that the group looks impressive in retrospect and the Magnificent Seven acquired their name in mid-2023 not January

Fourth for anything in an AI supply chain I look at customer concentration in filings before analyzing growth. That's the number the Nifty Fifty and the fiber operators would have wanted their investors to check

None of this predicts anything. It simply prevents me from confusing seven decisions with five hundred

The Bottom Line

2023 took the worst tech bear market in more than a decade and turned it around a single tech theme with an earnings report as the pivot point. Seven companies did most of the market's work the index became a bet on an idea and the largest capital spending cycle in corporate history began

The breakdown is the number to remember. Sixteen of the S&P's 24 points came from seven stocks that began the year with about one-fifth of the index implying roughly an 80 percent return for that group versus about 10 percent for the other 493 and the 12 percent equal-weighted index bears this out. Index investors made a concentrated bet that no one asked them to approve and by the end of the year those names were close to the30 percent of the benchmark which goes both ways

The most appropriate case is the Nifty Fifty and not the 1999 one. Those companies were really excellent the thesis was broadly correct and buying them at the wrong price cost investors twenty-five years of excess profitability without ever having produced the collapse that would have proven them wrong. Real profits answer the question of whether a business exists. They do not answer the question of how much has already been paid for it. Everything related to the markets from 2023 the debates about theconcentration the rise of power the valuation arguments that fill this site go back to the year AI ate the stock market