April 2024, Island Resilience

The AI Bubble and its Needles

By Marc J. Elzenbeck

Hot as the Dot-Com bubble may have been, the AI bubble is blazing. There are strong connections between now and then. At its height in 1999, a smallish company named Sun Microsystems made computers key to powering the graphical internet, and its stock was bid to a staggering 10x price to sales.

If Sun Microsystems valuation was staggering at that valuation, Nvidia’s current 40x sales figure is eye-popping. For perspective, the historical price/sales average for Standard and Poor’s 500 companies is only 1.5x. 

Like Sun, Nvidia is an innovative, smallish company, a maker of computer graphic cards with an up-and-down price history. It sells 80% of all AI “cloud” server farm chips, designing the best metaphorical shovels and picks for miners in an AI Gold Rush that’s in everything, everywhere, all at once. By volume, its product is worth more than gold and easier to transport than uncut diamonds. 

Since OpenAI rolled out ChatGPT’s public beta in November of 2022, the AI sector has absorbed trillions in investment, led by pension funds, big tech, and Congress members, driving United States stock markets into their biggest bull run ever. A few weeks ago, Nvidia hit $2.4 trillion in market cap, coming within a good trading day of being the world’s second most-valuable company. Only Microsoft and Apple still remain above it. 

Starry-eyed speculators like Google (Alphabet), Microsoft, Apple, Amazon, and Facebook (Meta) are splurging on tents in Deadwood and filing their claims. They see the enormous potential impact on their future business, so are compelled to establish an advantage. Collectively, they’re funding AI-focused spin-offs and start-ups at a furious rate of one or more per day.  

One potential problem for new prospectors, large and small, is that, just like in the Dot-Com era, this Gold Rush doesn’t yet have a clear or established business model, and graphically-oriented Generative AI requires huge amounts of computing power. Another potential issue is supply disruption: 90% of AI chips are now made in Taiwan, which is politically subject to China and just experienced a major earthquake.

I’ve helped make solutions and core AI technologies for over 30 years, so am just an old foot soldier doing his usual homework. Nvidia does have a real business model; its recent performance was propelled by sales growth and they reported record top-line revenues last quarter. If you dig a little, however, there are interesting signals in the details. In 2023, Nvidia made many “strategic investments” in its customers. These customers took money from Nvidia, then used it to buy chips, which Nvidia then booked as service-connected revenue, immediately round-tripping the sales. The customers then typically used the chips as collateral to raise more money from other investors to fund their electricity-hungry server farm operations. 

If you’ve got the hottest tools in the world, why would you need to lend new customers money? On the one hand, Ford and Toyota do this all the time to unload inventory. At the Microsoft-hosted “Davos of AI,” Nvidia just announced its new Blackwell 200 chip, said to be 30 times more powerful than the H100. So, with a better shovel on the way, it made sense to clear the old ones out. Chips go down in value, not up. On the other hand, we’ve seen former stock market darlings Enron and Cisco use similar sales accounting and practices with disastrous results.

One of its newest start-up customers, HyperCloud Nexxus, got a $1.1 billion loan from Nvidia. It is valued at $100k and founded by a Norwegian who went to jail for money laundering – then changed his name and moved to Dubai. Earlier this year, HyperCloud Nexxus also announced a $300 million merger with American Cannabis. So … why would a server farm with a billion in brand-new H100 chips want to buy a pot company? Some market observers questioned the strategy and the proposed deal was called off on March 1st

This is not to say that former international felons don’t make fine executives, and that mixing chips and cannabis isn’t a match made in heaven. Or that it’s illegal to loan money to somebody to buy your product and call it revenue. What it’s saying is that there are speculative red flags gathering and that millions of Americans now have a substantial portion of their retirements riding on one giant bet. Wall Street consensus is that Nvidia’s stock price is expected to double or triple over the next year or two.

As it happens, one of my first AI projects was for a government that asked, “How can we predict avalanches in the Alps?” There are a lot of factors – temperature, motility, declension, wind – but the simplest way to predict an avalanche is to go to where one already happened and consider how much new material has accumulated. Avalanches, like bubbles, are cautionary tales. Their energy potentials swell until they can’t be contained anymore. Then it lets loose all at once.

Consider Sun Microsystems again, that little company that once powered up the internet. While it was indispensable its stock rose like a rocket, to $200+ billion. As the internet was built out, lead times for its orders shrank, inventory ballooned, and the stock lost over 96% of its value. Sun’s CEO, Scott McNealy, later publicly told investors they were crazy to have paid so much for the stock. He kept right on making great computers and eventually Oracle, a database company that adapted to and enhanced the internet, bought Sun for $7 billion in 2010. 

The selected S&P 500 list of important publicly traded stocks was created 67 years ago. During that time 148 companies on it have temporarily been valued at over 40x sales. None stayed so for long. On average, over the following year the price per share of those blessed 148 fell by negative 38% compared to the market average. History may not repeat exactly, but there are obvious parallels for the current crop of AI flyers with euphoric expectations.

April 8, 2024

About Author