Not A Second AI Winter, But a Bubble
In the late 1950s, Frank Rosenblatt introduced the perceptron, a pioneering neural network model designed to recognize patterns and learn from data. This breakthrough seemed to herald the dawn of a new era in artificial intelligence. The perceptron could learn to identify shapes, letters, and even simple images. The excitement around this technology was palpable, with predictions that perceptrons would soon rival human intelligence.
However, this optimism was abruptly curtailed by the incisive critique of Marvin Minsky and Seymour Papert. In their 1969 book *Perceptrons*, Minsky and Papert highlighted the limitations of the perceptron, notably its inability to solve real life problems that were not linearly separable. They argued that perceptrons, in their current form, were fundamentally limited and could not achieve the lofty goals set for them. an excellent summary of the percepton and what happened to it can be found in this youtube video at around minute 3:40
This critique had a profound impact on the field. Funding for AI research, which had been abundant, began to dry up. The enthusiasm that once drove the field forward was replaced by skepticism and caution. This period of diminished interest and investment in AI is now referred to as the "AI Winter." It was a time when the promise of AI seemed to have vanished.
Fast forward to today, and we find ourselves in a vastly different landscape. AI has made remarkable strides, fueled by advancements in computational power, data availability, and sophisticated algorithms. Machine learning, particularly deep learning, has demonstrated capabilities that were once the realm of science fiction. AI systems can now beat world champions in complex games, drive cars, translate languages in real-time, generate human-like text, and create videos out of simple text prompts.
Yes, the hype is real, and most of us realized it since the introduction of ChatGPT 3.5 in November 2022. Since then, a lot has happened, and the adoption rate has increased in a phenomenal way, most notably observed in the soaring valuations of GPU value chain providers such as NVIDIA, AMD, TSMC, and others. These companies have seen their stock prices skyrocket as demand for AI-related hardware has surged. NVIDIA, in particular, has become a market leader in AI hardware, with its GPUs being the backbone of many AI applications. AMD has also seen significant growth, leveraging its competitive products to capture a share of this booming market. Meanwhile, TSMC, as a leading semiconductor manufacturer, has benefited from the increased demand for advanced chips required for AI computations.
To see this hype in numbers, since November of 2022, NVIDIA has returned over 790% in capital gain to shareholders, crossing the $3 Trillion market cap (it was $250 Billion in November 2022). To give you even more perspective and comparing on the same time frame, NVIDIA outpermored Bitcoin which rose 230%!
Yet, despite these impressive achievements, there is a growing concern that we may be entering an AI bubble. The rapid commercialization of AI, the hype surrounding its potential, and the substantial investments pouring into the field have created an environment reminiscent of past technological bubbles. Many companies are rushing to integrate AI into their products and services, often without a clear understanding of its limitations and potential pitfalls.
This post aims to explore the parallels between the current AI boom and historical technological bubbles. While we are not on the brink of another AI Winter, the unchecked enthusiasm and speculative investments in AI could lead to a period of disillusionment. It is crucial to recognize the signs of an emerging bubble and take steps to ensure that the field continues to advance in a sustainable and realistic manner.
The biggest problem with this upcoming bubble is that its being targeted from multiple angles, spanning from short to long terms ones:
Valuation
Geopolitical risk
Infrastructure asymmetry
1. Valuation
During COVID (remember it?) I was lucky to discover Aswath Damodaran known as the dean of valuation, and took his full course (totally free on youtube). I still remember when he explained the importance of valuation, drawing a nice analogy of the lemmings (a rat like animal) that moves in herds and suddenly face a cliff. The first lemming falls into the ocean, followed the second one and so on.
The lemming in the back is seeing whats going on in front of it, its brain left and right is begging not to continue or else you will end up in the bottom of the ocean. But then in the back of its heads the most dangerous seven words come out ….
“They must know something that I don’t” and it continues till it meets the wonders of fate.
That’s exactly what valuation is all about, knowing where you stand and far away you from that, so that you when you need to hit the brakes instead of relying on “They must know something that I don’t”
If you want a good cliff reminder, remember the days of CISCO before the dot-com bubble. The craze of the internet back on those days put CISCO under the spot light being the “NVIDIA” of networking hardware supply. If you needed the “internet”, CISCO name had to be mentioned, putting the stock at a market cap of around $560 billion in March 2000, making it the largest company in the world at the time. After the dot-com bubble burst, Cisco's stock price plummeted 88% from its peak, taking over 20 years to recover.
Lets go back to NVIDIA and get some sense
by becoming the most valuable company on search on Jun 17th, NVIDIA is trading at a next twelve month PE ratio of 46.7 and with PEG ratio of 1.46 (at the time of writing this) means that current market expectations are for NVIDIA to have earning growth of 31.7% over the next year!!
Doing a proper valuation for a company like NVIDIA is no easy task given how early we are in the AI age. But we know a couple of things:
Over 50% of NVIDIA revenue is coming from the US
Estimations are that nearly 40% of revenue is concentrated around META, Alphabet, Amazon, and Microsoft, followed by big purchases coming from TESLA and OpenAI. Most of those highly concentrated customers are reportedly working on their own alternatives to stop relying on NVIDIA GPUs.
Not mentioning AMD recent MI300X that competes against the H100 and getting closer to it in terms of performance. But even NVIDIA customers are working on their chips designs to avoid the high cost coming form NVIDIA chips.
Again, so much to talk about valuation here, but to me, so much expectations have already been factored in the share price making companies like NVIDIA one mistake away or earning away from being missed from analyst expectation from tumbling. and even if the stock keep getting higher highs, I just don’t want to see it worth being a lemming in this context.
2. Geopolitical Risk
For this, we will jump from Chip design firms like NVIDIA, and go chip manufacturers, zooming more into Taiwan Semiconductor Manufacturing Company (TSMC).
TSMC is considered a “Foundary”, a pure-play semiconductor manufacturing facility that does nothing but manfucaturing chips to design companies which are considered “fabless” (like NVIDIA).
TSMC is by no far a new name in the foundary world, and for sure one way or another someone heard about. But what most people won’t realize is as of today, TSMC might be (if not already) the most important company on this earth.
TSMC has a market share of over 60% market share in the manufacturing segment, and once we zoom in the leading-edge technologies (10 nm and 7 nm chips), TSMS goes over 90% market share
and given how gruesom this industry in terms of required CAPEX, speedy innovation, TSMS was one of the few that managed to thrive over the years. from 26 players back early in 2002 to only 3 major ones.
with that said in mind, lets explore the geopolitical risk in bulletpoints:
TSMC is Taiwan, and Taiwan is TSMC. The company on its own contributes to around 4-5% of the country GDP, and the whole sector in Taiwan contributes to 15% of its GDP.
The US government has implemented a series of export control measures to restrict China's access to advanced AI chips and chip-making tools. These measures aim to safeguard US national security interests and prevent China from acquiring advanced technologies that could enhance its military capabilities.
As tensions between China and Taiwan escalate, the potential for a Chinese invasion has become a critical concern for global security and economic stability. Taiwan’s strategic importance is underscored by its dominance in the semiconductor industry, particularly through Taiwan Semiconductor Manufacturing Company (TSMC). if such possibility is realized, the backdrop from it could be taking the world back in time for years due to sudden shortage of semiconductors, reverting nations to older technologies to run basic necessities instead of aspiring for further technologies innovations in the realm of AI, 5G, autonomous vehicles and other cutting edge technologies.
3. Infrastructure asymmetry
It becomes quite a long road to create the value chain till your ChatGPT gets answered. From hardware layer design to manufacturing, moving up the ladder to data infrastructure layer, and then to the application layer.
When this value chain gets examined, we can conclude that the Theory of Constraints (TOC) by Eliyahu M. Goldratt can applied, which emphasizes that in any manufacturing process, the overall production speed is determined by the slowest (or most constrained) step in the process. Such constraints (or bottlenecks) can be identified at multiple places in the value chain.
I believe the software development aspect is the least worrying while expanding manufacturing capabilities & power grids (being multi-year processes and capital-intensive processes) could be worrying one. And with the rising demand of AI computing, such a gap keeps widening, hence the Infrastructure Asymmetry.
Let’s examine the potential gap in power consumption demand
To get a view of how much power is needed to run AI, Goldman Sachs in one of their deep dives on AI & power demand dynamics reported that a single ChatGPT query requires 2.9 watt-hours of electricity compared to 0.3 watt-hours for a single Google search (that’s ten times fold).
Though data centers have been around for a long time, with an increase in their power demand, they also increased their level of power consumption efficiency, this is why we witnessed a flat line between 2015 and 2019. But that tug of war between demand and consumption reached an end by 2020 and efficiency didn’t have much more room to improve, adding AI-hungry computes on top it, and we see predictions of increasing demand in the order of 200 terawatts between 2023 and 2030.
This requires lots of investments ahead of time. Europe on its own requires more than $1 trillion to prepare its power grid for AI demand, either through new grid development, or modernization of 40-year-old grid systems.
I made my point in an earlier article how the bull thesis to invest in electrical utilities, and I still remain at my point.
Eventually, power grid development will keep pace with AI overflow, but I doubt over the next couple of years it will see the light. which poses a challenge to the current AI demand and potential meltdown in the short to medium term.
Conclusion
In conclusion, while the advancements in AI are undeniably transformative, they also bring with them significant risks reminiscent of historical technological bubbles. The current euphoria around AI technologies, characterized by skyrocketing valuations and rapid adoption, could be overshadowing critical issues such as unsustainable valuations, geopolitical instability, and infrastructure bottlenecks.
As we witness the remarkable achievements in AI, it is imperative to stay careful about these underlying risks. The field's rapid growth must be matched with a thoughtful consideration of its limitations and potential pitfalls. Avoiding the mistakes of past technological bubbles requires a balanced perspective—one that appreciates the extraordinary progress while also acknowledging the challenges that lie ahead.
Those who invested early have made quite the gains, no denial about it. But those who are seeking to put up their capital into such sphere should be cautious as the risk/reward dynamics will be different.
Thanks!