Nvidia CEO Jensen Huang (Photo by Artur Widak/NurPhoto via Getty Images)

“What about the AI bubble?” is the most common question being asked of economists and investment professionals today. Skepticism is appropriate about two issues: the assertion of a bubble, and the valuation of companies in new fields. Bubbles are hard to identify in real time, though underlying demand may be easier to see.

Underlying fundamentals are extremely positive for artificial intelligence companies, with two major economic activities almost certain to prove strong. Companies producing large language models (LLMs) will be a major engine of productivity across a very wide swath of business, non-profits, and government.

In addition, a multitude of companies are developing specialized applications that use LLMs to improve productivity in specific tasks, such as billing or product design. In addition to these two major approaches, small language models for specialized tasks have been developed, though it’s less certain that this approach will beat specialized apps connected to LLMs.

But strong fundamentals do not always justify sky-high company valuations. The trick is to figure out how high the fundamentals will be relative to current valuations. That is incredibly hard to do, except in retrospect.

The AI industry is figuring out the best business model for applying to technology to real-world problems. That combines with human tendency toward speculative frenzies to make valuation difficult. The business model challenge has been under-appreciated in recent discussions.

In 1908, the United States had over 250 companies manufacturing automobiles. Two major trends followed: Cars sales soared to previously-unbelievable levels, and most car companies went out of business. That’s a stunning result.

Car companies had to figure out how to manufacture the product efficiently, something that Ford was good at. And they had to figure out what customers most wanted in their cars, which General Motors was better at. In an industry with large economies of scale, the lower-performing companies lost market-share, causing their costs to exceed that of other companies. So they failed.

Similar results came in the dot-com boom of the 1990s. Many companies were formed in e-commerce, and most went out of business. But online shopping has hit levels that few expected two decades ago. The business concept was great, but most of the specific attempts failed.

Underlying this dichotomy is a concept I call “the trial and error economy.” Businesses need to experiment even though many experiments will fail—or they succeed in the sense that the company learns what will not work. Rarely does a major innovation work perfectly in its first iteration.

In the AI space, we will watch many errors. Some companies will fail. Others will stumble badly but survive. And a few will find the right model for helping their customers achieve huge gains in productivity—and they will end up being worth their lofty valuations.

Market structure is part of the puzzle as to ultimate company success. Sometimes one company dominates a particular space. Who can name the number two company in small business accounting software? That’s because every accountant can load a computer file created with QuickBooks. Small business owners with an alternative accounting program have more trouble.

Some industries have a handful of big players, while others have many small players—think of restaurants. I wrote a while back: “Whereas the large language model will resemble the Airbus-Boeing oligopoly, the applications sector will look like sushi, burgers, pizza and on and on.”

That is, the businesses providing very specific applications will be numerous, using very specific knowledge of a sector. When an electrical contractor wants faster turnaround times for bids, he will turn to an AI product developed with intimate knowledge of that particular niche: estimating costs of electrical installations. That program will not help to the car dealer trying to boost the service department’s sales, but some other app will. However, those application providers may use the same LLM as the engine behind the scene.

As for stock valuation, consider that Amazon stock reached a high in 1999. Two years later, the stock had dropped 95%. It turned out, though, that buying at the 1999 high wasn’t so stupid. The stock is now worth 52 times that old 1999 peak. A comparable investment in the Standard & Poor’s 500 index would, with dividends reinvested, be worth more than eight times the amount invested.

The Amazon story is valuable not only for its eventual monetary value, but because of the wild ride along the way. The stock has had plenty of setbacks along its journey. And investors should remember that many dot-com investments became worthless.

Is the AI industry in a bubble? I honestly don’t know. Every individual stock in the sector is pretty risky at current prices, but some will prove to be worth today’s price and more. Many investments will be total losses. However, the evidence is strong that AI will be a huge force in the business world for years to come.

More By This Author: