AI is data and power hungry
Remember that AI infra is expensive due to the high-cost GPUs and sky-high energy demands (nuclear energy is even needed). Training an LLM or AI deep-neural model requires massive datasets and energy , not to mention the associated R&D and talent costs. Data centers need large-scale cooling systems which require massive energy. Companies are poaching AI talents with multi-million-dollar salaries. They are also building nuclear power plants or leasing energy from them.
Who will be the winner?
Like all tech race, eventually there will only be one or two winners. In the 1980s personal computing era, Apple and Microsoft emerged as the winners and also amongst the largest companies as of today. In the internet era, Google emerged as the sole search engine winner. In social networking, Facebook emerged as the sole winner. You get the idea. What will happen to the losers? They will have to write off/divest their AI infra investments. Startups shutdown or consolidate one by one. The cloud revenues of Mag7 slowly taper off/decline as demand from startups fade away. The entire ecosystem collapses like a domino. There will be an oversupply of GPUs and you know what happens next. And then... the stocks will plummet and the bubble will eventually burst when investors become disappointed and disillusioned.
Is it so simple?
Don't get me wrong. AI is real and does produce real productivity gains. Our lives and work will be completely transformed by AI forever. You can't put the genie back into the bottle. AI is here to stay. However, the bubbly part of this whole story is there are too many players -- OpenAI, Mag7, Perplexity, Anthropic, Mistral just to name a few. In any gold rush, most will be burned miserably at the end. Any new tech will take at least a few decades to play out. Think of the personal computing and internet. After the first wave of AI hype cycle dies down, the winners will survive and continue to grow and dominate from there.
Parallels with the dotcom bubble
History never repeats but it often rhymes. There are many similarities with the dotcom bubble -- investors chasing after AI-related stocks, the Mag7 propping up the entire market, companies spending billions on infra anticipating accelerating demands that may never materialize so soon, sky-high valuations (all-time highs), most companies mentioning "AI" and related buzzwords during their earnings calls, absurd claims like AI will replace 50% of all human jobs. Nvidia today is basically the Cisco and Sun microsystem of 2000s. So much capex but with potentially diminishing returns.
A perfect storm?
With that said, the stocks may go even higher for longer. Nobody will know exactly what happens next. The bubble may burst in the next few years -- either violently or gradually deflate. A macro environment with high interest rate, slowing economy and tariffs may provide the catalysts.
Other risks
The other risk is the invention of more efficient algorithms that require less data, GPUs and energy. Recall the DeepSeek "flash crash" earlier this year. This is a very real risk. Any breakthrough in research might cause another "flash crash". At the core of this AI revolution is the LLM which is basically the transformer deep-learning architecture with reinforcement learning. Before transformers, we have Recurring Neural Network and Long-Short Term Neural Net. So it is expected that active research will produce newer models/algorithms. The main problem with AI is that you'd have to constantly spend billions in R&D and training the models and to replace the GPUs every few years, otherwise you risk losing out to competitors. There is a lot of peer pressure to increase capex each year. As a result, it is very difficult for startups like OpenAI to be profitable. Before startups can even recoup the capex, they'd have to spend billions to train new models/algorithms. It's almost a never-ending game. Most companies are trapped in this AI race where they have to keep spending and spending to chase after finite demand.
Adding fuel
The US government has recently announced billions of investments in AI infra and made deals with nations in Middle East to build out sovereign AI. This further adds fuel to the fire.
Warren Buffet
Is that why he has a record high amount of cash? Is he anticipating a bubble burst anytime soon? People keep saying don't ever time the market and he also preaches the same. But the truth is, he does market timing sometimes.
Conclusion
- There are many big-name startups and smaller startups that drive up demand for AI infra, energy and cloud computing. They are all loss-making and burning cash rapidly to play in this very expensive game. Perhaps the most expensive out of all tech revolutions so far in history. The scale is unprecedented. And Mag7 and other big-names are still increasing capex exponentially.
- When most of these startups die off after a few years, the demand for AI infra / compute may taper off/decline. There might be an oversupply when that happens. Whether the demand for these resources can be absorbed by other businesses remain to be seen.
- To be profitable, a big-name startup needs to be dominant in its market, or at least get a lion's share and control its costs well, otherwise it is hard to compensate for the high burn rate. But given the massive competition, it is difficult to do so. Alternatively, raise prices for its products/services, but that may decrease demand.
Who would be the winners?
I predict the incumbents -- namely the Mag7 will ultimately be the winners, not exactly sure which one but one or two of them, as they are cash-rich and have an existing moat and ecosystem around their businesses. They are better positioned to weather any bubble burst. AI will enhance their existing products which will further improve their unbreakable moats. Microsoft is still the software king that has proven itself capable of integrating AI across its products. Google is also well positioned with deep research, big data and compute resources. It has access to an existing huge amount of data (Internet) for training its models.
No comments:
Post a Comment