Doug Kass: Is AI a Big Tech 'Hail Mary'?
The value of what large language models can deliver relative to the cost of delivering it, is still not close to being economical.
You've reached your free article limit
You've read 0 of 1 free Pro articles.
- Wall Street Journal, "This AI Pioneer Thinks AI Is Dumber Than a Cat"
In tech, and in most things, when the only solution seems to be more of what isn’t working well, you have a problem that really cannot be solved. As we all know, Nvidia NVDA has a new part, Blackwell, which is the "Gerry Cooney" for the large language models, or LLMs. The hyperscalers' hope is the new part will make what doesn’t work well, work well. It won’t. It will likely deliver the same bad results. The whole approach may be flawed. A chip that does the same thing with more transistors does not change anything. If the whole approach is flawed, using the same approach with more transistors does not change a thing. As noted in an earlier "Tales," they can just hallucinate with expensive water-cooled processing cycles now.
Below I include some snippets from an interesting piece of research late last week from Julien Garran at Macrostrategy. His research touches on some big picture stuff and then, importantly, the vendor financing and investing in customers issue. It is clear that business for Nvidia remains good over the near term -- better than most have thought. The stock market is telling us that, as well. In addition to the money recycling issue, which is addressed in the research piece, is somewhat counter-intuitive. Basically, the less well this stuff works, the more money the hyperscalers seem willing to throw at it. They got the money, regardless of how grey the source of the money is, so the brinkmanship continues. They are desperate to solve a problem, and their solution for the time being is to ... throw more money at it.
With respect to the LLMs, the value of what is delivered relative to the cost of delivering it, is still not close to being economical. As stated in this report, the more people who use it, the more money they lose. And what is delivered, in most cases, is still no better -- or even less good -- than doing it the old fashioned way, which is very cheap. But, it is so fascinating.
Not only is the stock telling you business is good for NVDA (people have loose lips), so is the stock market. Everyone is now expecting another massive beat and raise. The market should have been down on last week's consumer price index report that was hotter than expected, and it shouldn’t have been robust as it has been recently.
NVDA was actually up as well on Thursday and Friday. The hyperfocus on one stock, and what it means for the equity markets in total, remains stunning. It is the oddest of the odd. It is one company, with some very company-specific issues, and not a reflection on the broader economy or most of the other 500 companies in the S&P. As they say, for the time being, it is what it is. This research note argues that Big Tech’s investment in artificial intelligence today is a ‘"Hail Mary," a phrase from American football; a desperate pass to a wide receiver, when the game looks dead and buried. Why are Big Tech and the software-as-a-service (SaaS) companies so keen on pumping large language model AI and AI applications? It is because they are running into growth problems, partly due to the lack of new products and content, partly due to physical constraints on their technology, and partly due to the regulator and the courts.
This research note argues that Big Tech’s investment in artificial intelligence today is a "Hail Mary," -- a desperate pass to a wide receiver, when the game looks dead and buried.
For SaaS companies, upselling LLM AIs would be a dream come true, if the LLM AI apps were useful. For Amazon AMZN, Alphabet GOOGL, Microsoft MSFT and the other major cloud providers, the huge processing requirements for a mass market AI would be a boon to their data center businesses. The question is: Can they scale up LLM AI to make up for the loss of growth prospects elsewhere? The key problems facing LLM AIs are, first, finding a killer app. A mass market LLM AI app is notably absent, despite two years of intense development and which, because LLMs were "Built to Fail," may never come. The second problem is taking that app to profitability. That is a major issue for LLMs, as the technology appears to display "dis-economies of scale." That means that the bigger they get, the more loss-making they become. In this note I argue that all the successful Big Tech companies have used "flywheels" to gain dominant market positions. But one could argue that none of the LLM AI businesses will achieve that, and that the AI boom will turn to bust. Microsoft’s failure to sell the Co-pilot app to more than 1% of its customers, despite being perfectly placed to do so, is one sign of trouble. Are there others? Absent causal (rather than correlation-based programming), absent a completely new mass market application, which I can’t see coming, and absent a method to turn dis-economies of scale during inference into economies of scale, there might be no path to profitability for LLM AI.
From Macrostrategy:
The short summary of this is that buyers are all loading up the boat on the new part, and racing to be the first to do it, thinking they can get ahead and something will change. Then they will put it into production, and find out not much changes. But the race continues. It is so odd. And it is so annoying that MSFT keeps all the losses of doing this off their books, too. Real economic accounting would be a big governor on this. The industry is an example of how a government would behave, without any media or regulations holding them accountable.
All these guys for the time being can crap money down the toilet, with little consequence. Therefore this cycle of irrationality can have longer legs than most cycles of irrationality, because the economic losses are hidden. Other cycles have been driven by public companies with real accounting, so you can see what is going on. This cycle is driven by private companies doing the spending, who are being funded by their suppliers (Nvidia) or customers (Microsoft), and the losses are kept out of the public markets, while the benefits accrue to the public companies, given how this is all structured.
Unlocking Another Key Point: Tesla
There is one more related point: Tesla TSLA. Tesla had its big robotaxi event Thursday night. This was the event that was supposed to happen a few months ago, but was delayed, ostensibly because they weren’t ready to go.
Well, it was one huge flop. (We are short).
The stock got hammered Friday to the tune of 9% (about $80 billion in market cap), because there is still nothing there. No matter how much money and time Tesla throws at it, the AI underpinning full self driving still does not work. Tesla was only really able to talk about vaporware, and things to come in a few years from now, which folks are obviously skeptical will actually happen.
It is so odd; what it's doing obviously isn’t working. Tesla has bright people. Normally bright, and practical, people pivot to a different approach when what they are doing isn’t working. But a different approach does not exist yet, nor will it exist for the foreseeable future. For the time being, I guess they can’t say mercy and give up, because they have promised so much, so they keep throwing more money at the same approach that doesn’t work. But that will not go on forever. Eventually economic gravity strikes, and all of these guys pursuing the same thing in different markets, also won’t be able to get enough power because there is a 7-year waiting list in many cases for new power now (that those jerks still want me to pay for).
This stuff really doesn’t work. As far as the mainstream population goes, the only people using LLMs are the tech bros and kids using it to cheat on their homework, because as long as you turn something in with words in it, regardless of what those words say, you are fine.
This really is the oddest cycle in the history of the markets. The first one ever that is opaque. I mentioned before in "Tales" that we have no idea how much money is being lost doing this, because a lot of it is hidden in private companies. Heck, you don’t even know what underlying revenue growth is. All the public markets know is how much money NVDA is collecting from the money losers, who NVDA itself is funding. Quite a racket, every dollar they fund someone with, probably turns into 1,000-times that in terms of NVDA market cap.
How can you even track this stuff? We don’t know the revenue. MSFT makes a cash investment into Open AI, it is a 1x transaction. But then the CAPEX and operating losses are off their books. TSLA – who knows, the accounting there is so opaque, nobody can ever figure out what is going on, and who knows what is pushed to xAI, which is off the books. And on and on, nobody knows exactly what is going on. Then on top of that, the depreciation schedules for the whole industry are baloney too. My guess is at least 2x as long as their practical use.
Anyway, this is where it all seems to be for the time being. Grey, opaque, and one company getting the benefit of something that seemingly doesn’t work, which can happen for some period of time because the losses that fund them are invisible to the markets and self-serving given how all the losses are funded. It really seems to be a big money recycling operation, with the losing end of the trade hidden.
Perhaps AI is a "Hail Mary."
This commentary was orginally posted in Doug's Daily Diary on TheStreet Pro.
At the time of publication, Kass was long AMZN (S); Short NVDA (S) and TSLA (VS)
