AI has been the buzzword of the past 18 months since the launch of ChatGPT in Nov 2022. Since then companies in the space have seen their valuations skyrocket, none more than NVIDIA. NVIDIA’s unprecedented growth is the result of a mad scramble for companies to build the computing infrastructure that powers the hungry AI models. However, many are starting to question whether the industry is getting ahead of itself and comparisons are now being made to the rollout of internet infrastructure in the 1990s. ABSI this week will dive into the similarities and draw conclusions on whether investors need to be sceptical about the sustainability of the AI rollout.
NVIDIA is a company that needs no introduction. The Company has become the world’s 3rd most valuable company with a market cap of over US$3 trillion off the back of record revenues which includes a ~262% increase YoY in Q1FY24 to ~US$26 billion with ~75% gross profit margins. NVIDIA’s meteoric rise is entirely due to the boom in AI. Its H100 tensor core GPU is designed to accelerate a wide range of applications, including deep learning training and inference, high-performance computing (HPC), and data analytics. The virtual monopoly NVIDIA holds over the market allows them to sell the H100 for between US$25k to US$35k, depending on configuration.
Source: Google
So who are NVIDIA’s customers that are spending tens of billions of dollars building out this AI infrastructure?
That would be the likes of Amazon, Meta, Google, Microsoft, OpenAI, Tesla, and many other major tech companies that have AI FOMO. According to an analysis by Goldman Sachs, tech giants will spend over US$1 trillion on AI capex over the next few years. While this is great news for NVIDIA, which will be the primary benefactor of the spending, it may spell financial disaster for its customers if the investment doesn’t result in the forecasted revenues.
Analysts at Barclays have taken the inferred 2026 capex and crunched the numbers on how much demand this infrastructure can support. Their analysts concluded that AI infrastructure in 2026 would be able to support over 1 quadrillion AI queries which is the equivalent of 12,000 ChatGPTs today. While there are several exciting new AI products on the horizon, it's hard to believe there would be 12,000 of them which is why analysts are starting to draw comparisons to the internet infrastructure rollout of the 1990’s/2000’s.
Taking a look at history, the internet started to take off in the 1990s and companies like Global Crossing and WorldCom spent billions laying over 80 million miles of fibre optic cable across America in the hope of internet traffic increasing feverishly. Raising over US$1.6 trillion in equity, the infrastructure was built under the assumption of internet traffic doubling every 100 days. However, in reality, it was doubling on an annual basis. The day of reckoning came for Global Crossing in Q4 2001 when it lost US$3.4 billion on US$793 million of revenue and ultimately went bankrupt.
Noteworthy, the overbuild was so great that in late 2005, ~85% of fibre lines were still unused. From a glass-half-full perspective, the upside of the overinvestment into internet infrastructure was bandwidth costs dropping ~90% over a four-year period which unleashed an era of innovation. During the early 2000s companies such as YouTube and Facebook were founded due to the excess capacity of being able to economically host photos and videos.
Today AI has a US$600 billion dollar problem. While there are no precise industry revenue estimates, it's generally accepted that OpenAI makes up the vast majority of that number and is currently estimated to have an ARR of ~US$3.4 billion. Even if you were to generously assume that most major tech companies were to generate US$10 billion in revenue in the next year, you’re still left with a roughly US$500 billion hole.
Economist Carlota Perez wrote a book, “Technological Revolutions and Financial Capital” explaining these phenomena of the gap between new technology installation and its deployment into wider society. She states that this gap is often marked by a burst in a speculative bubble that formed around its installation.
There is no argument that generative AI is groundbreaking technology and will play an important role in progressing humanity. Having said that, it is likely there is a bubble and that it will probably burst. Noteworthy, there are a lot of reasons to believe that a bursting of the AI bubble will not have the same impact as the dot com bubble which saw the Nasdaq lose ~70% of its value. Namely, the companies building the infrastructure are very large and diversified tech businesses that can afford to take the hit. The silver lining is that just as the overbuilt fibre networks led to new innovations in the internet, the overbuilding of AI infrastructure will accelerate the industry and democratise access resulting in applications we can only dream of today.
We offer value-rich content to our BPC community of subscribers. If you're interested in the stock market, you will enjoy our exclusive mailing lists focused on all aspects of the market.
To receive our exclusive E-Newsletter, subscribe to 'As Barclay Sees It' now.