The Coming AI Bubble?
A chip maker is the most valuable company on the planet.
WSJ (“Nvidia Becomes First $5 Trillion Company“):
Nvidia became the first company to hit $5 trillion in market value, the latest milestone in an unprecedented surge that reflects the growing influence of artificial intelligence on markets and the economy.
The chip maker’s shares recently topped $210, ahead of the $205.76 needed for a $5 trillion valuation. The stock has been boosted by exuberance for AI’s potential and a recent flurry of deals and partnerships with some of the biggest companies in AI and corporate America, from OpenAI and Oracle to Nokia and drugmaker Eli Lilly. Enthusiasm for Chief Executive Officer Jensen Huang’s speech in Washington, D.C., also helped fuel gains this week.
Nvidia is now larger than AMD, ASML, Broadcom, Intel, Lam Research, Micron, Qualcomm and Taiwan Semiconductor Manufacturing combined, according to Dow Jones Market Data. Its value also exceeds entire sectors of the S&P 500, including utilities, industrials and consumer staples.
That the company that came to prominence making gaming chips and is now at the center of the AI boom is valuable is not surprising. That it’s more than so many blue chip companies combined, though, is. Especially consider how rapid its rise has been:
Its value held essentially flat at $1 trillion for years, hitting $2 trillion six months ago. And it has simply exploded since.
Naturally, this has people wondering how sustainable it all is.
Some investors, technology executives and industry analysts have begun to raise the prospect of an AI bubble akin to the dot-com boom and bust. Technology companies are pouring hundreds of billions into data center development and chips and taking on heavy debt, but current revenue is relatively tiny.
The “remarkable valuation” sets high expectations for the company, said David Kotok, co-founder of Cumberland Advisors. “It is justified only if margins and profits continue on the current trajectory or even get better.”
Indeed, Nvidia’s valuation is among the most elevated in the equities market. The stock had a price-to-earnings ratio of about 33 as of Wednesday morning, based on next year’s projected earnings, compared with an average of about 24 for the S&P 500.
The chip maker still trails some of the more eye-watering valuations of its tech industry peers: Tesla is trading at more than 210 times next year’s earnings. Palantir’s price-to-earnings ratio is even higher.
This, too, is interesting:
Nvidia in September agreed to invest up to $100 billion in ChatGPT maker OpenAI, which would allow the startup to build and deploy at least 10 gigawatts of Nvidia systems for its AI data centers. Nvidia has also backed a host of AI startups. If heavy spending on the AI race wanes, some worry that Nvidia could be hit twice, with less revenue and declining value of its equity investments in customers.
This is the opposite of diversification.
Let me get the pedantry out of the way:
If it is a bubble, it’s already here. What’s coming are the consequences when the bubble pops.
Ok. There seems to be a lot of crossbreeding, incestuous relations between chip makers, data center builders, cloud services, software companies, that look suspiciously as some way of inflating stock valuations by keeping the growth going.
That said, real money is being spent on things like chips, data centers, and electricity. This may be a good thing overall, though it’s also straining the electric grid, giving rise to unregulated on-site gas generators to keep up, and using up prodigious amounts of water to cool things down (why they can’t or wont’ recycle the water, is beyond me).
Revenue, as noted, is low, where there is any revenue at all. I heard a figure of $12 billion as annual revenue for Open AI. It’s insane to think that much revenue is 1) considered small, and 2) not enough to sustain an industry.
And there’s been some chatter that paying compensation to copyright holders for having stolen their work to train LLMs, would turn that revenue negative. This might explain why the director of the US copyright office is being fired.
It is a bubble, IMO.
But I’m also going with one of my simple rules for living: This time it’s NOT different applies to 99% of cases.
Scott Galloway Warns Stock Market Is a ‘Late-Stage Bubble’: ‘America Right Now is a Giant Bet On AI’
Scott Galloway has been saying, for months, that the Stock Market doesn’t make sense right now, given earnings, inflation, interest rates, unemployment, layoffs, tarrifs and lack of profits.
I’m in cash, bonds, and T-Bills.
If AI is as useful and as successful as its proponents hope, it will defeat itself by causing mass unemployment. More AI = fewer wage earners = a shrinking market of paying users. AI doesn’t just reduce the number of workers, it also reduces the number of consumers.
It’s the problem of the cure for cancer. Find a cure for cancer and you’ll be worth a trillion dollars. In the first year. After that, the demand for the cure falls off a cliff and your cancer-curing company is junk.
Yup, it’s a bubble and it will be ugly when it breaks. The build out of data centers is similar to the build out of telecom networks in the late 90’s, early 00’s. The developers of those data centers will feel the pain, but in the long run they’ll be an asset. They’ll be bought up for pennies on the dollar and will provide lower cost computing power for the surviving and future companies. It only for a few years for the excess telecom capacity to be utilized.
Beyond the data center developers, the ones that are going to be crushed are those pursuing LLM AI.
@Michael Reynolds: Thinking in same line. How does AI become a revenue generating machine? I’ll use it, but I ain’t paying for it. And the kids who do use it will not have jobs to earn money to pay to use it. And what is it used for that would require locked in recurring subscription revenue long term? It’s not like it is an Apple or Msoft OS….Right now it is used for what? Writing code faster, making dead celebrity meme videos, doing your homework, replacing human writing…I mean maybe it replaces human jobs etc but at what cost? One needs the other, and humans will just unplug it and sure as heck will not pay the price tag that it will need to be set at to recoup all of this investment, energy consumption and infrastructure and cooling etc…
I mean maybe its a generational thing. Being older I have used AI, but if you charge me X, I’ll just go back to doing it myself.
The CEO announced $500B in new orders this week. That despite (at least) Amazon and Microsoft committing to designing and using their own chips. I think the place where the bubble breaks is when everyone realizes that TSMC is the only actual manufacturer — for all of NVidia, Apple, Amazon, Microsoft, etc — and they don’t have nearly the capacity to produce that many.
Another thing that bugs me about the great AI debate, is how casually people accept the need for humongous data centers and brobdingnagian use of electrical energy.
What this means is the LLMs do not run on your PC or phone, but on the “data” centers. That’s why you need an internet connection to use them in the first place.
Ok. But so what’ you need an internet connection just to access your shopping list in the cloud, right?
No. You can store your files in your PC or phone and back them up in the cloud. That’s what I do. I think it’s what most people do, given how cheap hard drives are these days.
The point is these things are very energy intensive. Add to this they are not that useful for most things that people do. Their best known output seems to be AI slop and hallucinations (ie made up BS).
So tons or money are being spent (“invested”), copyrights have been violated, data has been mined, ludicrous amounts of electricity and water are being consumed, in order to get not very much that’s useful in return.
This is somewhat reminiscent of the dot.com bubble, when many companies with a high valuation had little to no income, and no real business plan beyond “being acquired by a larger company.” But at least then these companies offered something useful or desirable, at costs far lower than the AI companies do today.
So, are you going to pay a ChatGPT subscription to make full use of the LLM capabilities? Right now, I’d demand payment to use the things regularly.
@Sleeping Dog: @Jc:
The issue with LLMs is obvious: once they’ve consumed all human-generated data, where do they go to feed? They will inevitably be training more and more on AI product. A narrowing and eventual closed loop, a dead-end.
@Michael Cain:
And TSMC stands for what? Taiwan Semiconductor Manufacturing Company. Taiwan, as in the country China is actively planning to invade as soon as 2027. Two years. Which I’m sure will be more than enough time for us to design, fund and build a domestic equivalent with all the competent employees needed. Right?
@Michael Reynolds:
The oligarchs will trade among themselves, and hire some peasants for security against the rest of the peasants.
Until Skynet kills them all.
The cancer cure bit, you have it wrong. People will continue to be born and some will continue to get cancer and they will continue to want the cure. Same if it’s a vaccine, except that vaccines are far, far, far, far cheaper than most drug, radiation, or chemotherapy treatments.
But in the firs place, there’s no such thing as “cancer” as if it were one thing. There are many types of cancer. So there’s never going to be a “cure for cancer.” There may be a cure for lung cancer, one for skin cancer, one for liver cancer, one for leukemia, etc.
I want to recommend a book I read a few years back (we have a family tradition that everyone receives a new book at Christmas). It was called The Resisters by Gish Jen. Rather than me summarizing it here is the Goodreads synopsis:
@Kathy:
The economics are inescapable: In Year One a cancer preventative can be sold to 100% of humans. In Year Two it can be sold only to newborns. Further, a cancer cure is valuable because cancer is terrifying. Once a cure is available, cancer will be less of a threat and defense against same will decline in value.
It’s the same as the common cold problem: curing the common cold would be worth far, far less than treating the symptoms of the cold.
And I would gently point out that as a 71 year-old man who drinks whiskey, smokes cigars and cannabis, and eats red meat, I am acutely aware that cancer is a hydra with many heads. (The very Rocky Patel currently clamped between my teeth may do for me.) I’ve had a sarcoma which obviously didn’t kill me. But if I’m diagnosed with pancreatic cancer I do understand that that is a whole different beast.
Finally, a topic I am qualified to comment on. Does anyone remember the 1990’s, when everyone was confused as to how Google was going to make money? For no reason Google stock kept going up and up and up. Then around the year 2000 (and it was considered a pretty big scandle at the time) we learned Google was tracking everyone and selling data to companies of all sorts – political, marketing, etc. We learned of a HUGE data center in nevada that was constantly getting data sent to it about everything you do. Now people understood why the valuations were so high. Sadly, without any real regulations, except a few ok ones in CA, everyone has adjusted to this way of life and accepts it as part of the convenience of the internet.
I’m a CIO for a healthcare company. We have been testing AI that automatically writes citations for medical facilities that violate CMS standards. We have 80 nurses all over the country that do this manually. After teaching the models through ambient listening to do what the site visitor does, AI writes the citations more accurately, consistently and quickly than a nurse. Basically, we don’t need the nurse anymore for this line of work. But we will need to replace 80 nurses with a 3 or 4 engineers.
As CIO, what I don’t like about it most, is we are running this info through engines out of our control. Our vendor has all kinds of misleading BA language that leads you to believe your data is yours and they won’t profit off of it or use it in ANY OTHER WAY. After sending it to our attorney, he agrees that language is ambigous and unclear. Additionally, even if they did violate a BA there are not overarching laws we can use to sue them anyway.
I been sayin’ , I been sayin’, I been sayin’.
AI has been so oversold, so overpromised, so ridiculously hyped by every mark consumed with FOMO! that there’s no ground below it.
It’s a masturbatory fantasy of the technuts.
Where AI will succeed, will be in specific applications and that is where it is having its first successes today.
@HelloWorld: The Google reference is good, though Google became wealthy because they cornered the market for online advertising to go along with the monopolistic dominance in search.
As a CIO you’re likely familiar with the term “shelfware,” that will affect AI apps as well as the implementations move up the corporate ladder. Lots of AI companies will fail waiting for the day that corporations fully implement the technology.
So, use cases.
I’ve mentioned an online course in the new acquisitions law, where uses for AI in assembling proposals were demonstrated. I was not impressed, and using it the way they showed requires a paid subscription for ChatGPT. I’ve tried something close to it with Gemini, and for what I do it does nothing useful Not.One.Thing.
They also claimed this would reduce prep time by some percentage. Maybe, but I doubt it. One other suggestion was to use a cloud server, which we already do, claiming this, too, reduces prep time. I can say there are advantages to cloud servers, but not that they’ve saved us as much as a second in prep. It’s worth saying we had a local file server anyway for the past ten years or more. So it’s not much different.
I’ve found it useful to summarize relatively small amounts of text. Once I fed one LLM a whole story, and the summary sucked. You might think it’s my fault. maybe the story wasn’t clearly written. Except it skipped whole sections, made up events, and claimed there was no ending. really, it was eager to know what happens next (as if).
I still think it can do specific searches for some thing. Not that I read its interpretation of what I ask, but that it directs me to what it uses as online sources, and I can read it myself.
I’ve also had it do images. It has improved over time. But at most I’d use such images as way to give a human artist a notion of what I want.
Some more later. I’m supposed to be working.
@Sleeping Dog: So, the vendor who built our system did it for free. Basically, they came to us saying they can already do what we were looking for. Actually, they lied. They knew in theory they could build it and then they did. Now, we are negotiating price to use the service. While it makes financial sense – and serves as a solid example of why valuations are so high (and as an example of why companies might fail, too) – I have a call with CMS (after the govt opens) to determine if we can even use it, because the law says an RN must be present.
However, it cannot be overstated how well AI does some things. My main point is that I am super paranoid that AI valuations are so high for factors that we do not know about yet.
@Michael Reynolds:
Have drug companies gone broke making antibiotics that cure a lot of bacterial diseases to this day?
@Kathy: This surprises me. We have CoPilot built into our systems and it has transformed how well we are managing institutional knowledge. Our IT helpdesk is able to resolve 97% of our support tickets without a human engineer. Our marketing department is writing blogs and sending them to medical facilities without anything but a marketing review of the communication. I could go on, but AI is a HUGE productivity booster. Still, I’m not sure humanity is ready for 2027…do some research if your interested, I don’t want to appear as a paranoid old man.
@Michael Reynolds:
TSMC has a presence in Phoenix. From memory, they started production from a new fab there this year and are getting acceptable yields. It’s a couple of generations behind the leading-edge chips they manufacture only in Taiwan. TSMC has suggested the Arizona facility will always be behind because they can’t duplicate the R&D supply chain for leading edge materials and ancillary equipment that exists in Taiwan.
I’m less concerned about exactly where TSMC’s fabs are than that the AI giants have all their eggs in the one corporate basket.
@HelloWorld:
I don’t work at the IT desk. I don’t solve report tickets, I file them 🙂
What I do is, mostly, take tables of products arranged in columns labelled like: code number, product, description, capacity, minimum amount, maximum amount, etc. These come in Word files. I then transfer them to an Excel file that has to follow a format found in the Word file. Then I add things like brand, price, minimum total, maximum total, etc. And add these up.
Attempts to have Gemini do this ended in disaster. Copilot did a bit better, running off Word, but did not get the format right, and couldn’t account for formatting discrepancies in the Word file.
I figure I could then spend some more time refining the prompt, or some time fixing the mess myself. Or do the whole thing myself in less time.
It’s possible other things involved in such projects, which I mostly don’t handle, might be more amenable to LLM ministrations.
All this before adding brands, prices, etc*. I don’t think it could do this, and I’m not even willing to try. A mistake can cost us considerable income, if not the project.**
* The etc. here includes things like country of origins, and the name of the manufacturer or distributor.
** There was the time we did not notice the 50 kilo sack of sugar had to be priced per kilo rather than per sack. Our offer was ridiculously high as a result.
@HelloWorld: I have found that AI is great when you don’t care about the quality of the result.
Anything that hasn’t been answered many times before in the training data results in mirages and hallucinations and gibbers and squeaks. Which, if you don’t care about the quality of the result, can be great.
Otherwise, ask an LLM to explain how the song “hobo humpin’ Slobo babe” is about the collapse of Yugoslavia, and it will just start spewing things citing lyrics that don’t exist.
But, if 97% of support tickets can be closed with “reboot your router” or “fuck off, leave me alone, and poke at it some more”… I can see how it would be very handy.