Is AI the Next Dot Com Bubble?
Since the launch of ChatGPT, two letters are on everyone’s lips: AI.
CEOs and VCs are frequently and publicly proclaiming that AI will be bigger than the internet. With the sheer scale of investment by venture capital and Big Tech, it’s nearly impossible for AI to have no impact at all. The trillion-dollar question is whether AI is a frenzy fed by overeager investors and CEOs which will struggle to generate meaningful revenue, or a true revolution that will reshape society and markets.
One can quickly find potential warning signs in today’s AI market that – without much squinting – will look familiar to those familiar with the dot-com bust of the early 2000s.
First, to meet an anticipated deluge of demand, there has been a massive CapEx investment in new infrastructure. During the dot-com boom, there was a false (but often repeated) statistic that Internet traffic was doubling every 100 days. With so much perceived demand, the race was on to build out infrastructure. In the decade preceding the bust, cumulative investment in telecom infrastructure was >$500Bn (not adjusted for inflation).
Source: Fabricated Knowledge
As the bubble burst, that fiber went largely unused. In 2002, only 2.7% of the installed fiber was in use. For AI, the Hyperscalers have sprinted through a similar version of that decade-long build out in only a few years. Microsoft, Google, Amazon and Facebook are committing to spend ~$200Bn on CapEx this year alone, mostly on AI infrastructure.
Second, people suddenly care about top-level domains once again, rushing to acquire their dot-ai real estate. The reason we refer to the speculative frenzy of the late 90’s as the “dot-com” bubble is because investors became obsessed with companies that leveraged the internet, and the of-the-moment proxy to demonstrate that was putting “dot-com” in the name of the company. In their earlier avatars, Amazon and Salesforce were branded explicitly as Amazon.com and Salesforce.com.
Pictured: Older co-workers trying to fit in with Zoomers, but calling it SFDC
Arguably the biggest beneficiary of the AI boom thus far has been the Caribbean nation of Anguilla, owners of the dot-ai top level domain. These registrations have more than tripled and now make up 20% of the island nation’s total revenue. It takes work to establish AI as a legitimate part of your business; it’s much easier to buy a dot-ai domain and pretend it is.
Third, public markets and the valuation of anything related to the hot technology have risen dramatically. During the dot-com boom, from the Mosaic IPO in August of 1995 through the local maximum two years later, the NASDAQ increased 75%.
From ChatGPT release to the most recent all-time high two years afterwards, the NASDAQ has risen 63%. Most striking is the comparison between Cisco, who rode the dot-com wave, and Nvidia, who is arguably the big winner so far in the AI boom.
Both companies saw stratospheric increases in stock price as the supplier of crucial infrastructure powering the forthcoming technology revolution. Both companies also briefly became the most valuable company in the world.
Note: “present day” in this chart refers to June 2024
Even though San Francisco named their very successful football team after the year of the Gold Rush, clearly the most enduring fact about boom times in the Bay Area is that the sellers of “picks and shovels” made the most money.
All of this said, the investing and technology community still bear the scars of the dot-com bubble. The AI era will certainly have its fair share of FOMO investment decisions, and we are already seeing some company valuations that are hard to justify. But, for a few key reasons, I believe these surface-level similarities are not the harbingers of an AI bubble poised to burst and drag down the entire global economy as we saw with dot-com.
First, AI CapEx is meeting tangible demand and is being considered in the context of a more sober ROI analysis. Rather than cheering on overbuilding, investors are pushing for tighter linkage between CapEx and demand. On the earnings call of all four Big Tech companies investing heavily in AI CapEx, analysts pressed on the expected returns from these investments and the demand indicators driving them. The clear message is that investment to-date has had high returns as the market has been supply constrained, rather than demand constrained. Microsoft, Amazon and Google are expected to generate ~$15Bn of AI revenue this year. Future investment is being planned but will only be undertaken if demand materializes as planned.
Second, exorbitant valuations for as-yet unrealized companies have been thus far contained to the private markets. It’s never ideal for excessive hype to lavish capital on incomplete ideas or unsustainable business models. However, private markets are better able to absorb losses without creating panic and contagion that spreads into the overall market. This cycle has not yet produced a Pets.com that goes to IPO one year after incorporation, leaving the general public holding the bag when it goes bankrupt shortly thereafter.
Third, share price gains from AI have been supported by more measured expectations of earnings growth. The forward earnings ratio of the NASDAQ only increased from 23x to 29x in the run up following the release of ChatGPT – lower than its level throughout much of 2020-2021. Nvidia’s forward PE reached ~50x when it was the most valuable company in the world. High, yes, but not nearly the ~135x valuation of Cisco when it reached its zenith. This quarter alone, Nvidia generated $14.5Bn of cashflow from operations.
Further, there is ample evidence that AI is already making a tangible economic impact, and the accelerating technology capabilities will compound those gains. Microsoft called out that its generative AI coding assistant GitHub co-pilot is already a larger business than GitHub at acquisition. Surveys, including the below from Stack Overflow, show that an overwhelming majority of developers plan to incorporate AI tools into all work processes.
These budding capabilities are not limited to coding either. To make the jump from “cool technology” to ubiquity, AI models must be able to run on the devices people use most often, namely smart phones. This requires much smaller models than the large language model that powers ChatGPT. Here, too, progress astounds.
Meta’s latest Llama model release included an 8 billion parameter model that is on par with or outperforms GPT 3.5 Turbo on many metrics. GPT 3.5 Turbo is the model that powered ChatGPT when it burst on the scene, dazzled the world and kicked off the AI hype. Meta, Mistral and others are showing that big models push the boundaries of what is possible, then new smaller models catch up to the old boundaries. This is the way that AI will come to permeate more and more of our lives.
The great irony for investors is that past bubbles arising from technology innovation (e.g. the dot-com bubble, automobiles and radio in the 1920’s, railroads in the 19th century) were not wrong in their predictions, only in their timing. By the 2010’s the internet had delivered nearly all the predictions that failed to come true a decade earlier.
If history is any guide, the AI-powered future envisioned will indeed come to pass. What we must wait to see is which investors and entrepreneurs get the timing right.