Control Panel: AI Bubble?

Investment Analysis Clubs Macro Economic Trends and Risks
1

https://www.wsj.com/finance/stocks/the-eerie-parallels-between-ai-mania-and-the-dot-com-bubble-f99be6fe?mod=hp_lead_pos9

The Eerie Parallels Between AI Mania and the Dot-Com Bubble

Bulls deny that there’s a 1990s-style bubble in AI. There are a few striking similarities, and some notable differences.

By James Mackintosh, The Wall Street Journal, Dec. 13, 2025

Valuation

There are lots of ways of valuing stocks, and pretty much all of them make U.S. shares look the most expensive since the dot-com bubble. The forward price-to-earnings ratio, price to cash flow, the “Fed model” calculation of the extra reward offered by stocks compared with bonds and the cyclically adjusted PE ratio all scream that stocks are expensive.

image
image1097×864 60 KB

…

Investment

…
The [fiber optic buildout] numbers in 2000 were immense, with well over $100 billion being sunk into new telecom networks in the late 1990s. There was so much fiber that much of it ended up mothballed for a decade before internet traffic expanded enough to justify using it.

The race to build data centers is even more extreme, with investment figures in the trillions thrown around by leading AI developers. Spending is so large that economists say it’s making up a significant share of gross domestic product growth…

Retail trading

Individuals are dominating stock trading, again betting big on tiny loss-making companies. In both the 2000 bubble and the 2021 bubble in SPACs, clean tech, crypto and cannabis, loss-making small stocks far and away beat profitable small stocks, which are much less exciting.

image
image1164×744 64.7 KB

… [end quote]

I have discussed the high valuations and over-investment before but the last point is new to me. I didn’t know that there is an S&P 600, a small-cap index which requires companies to show profits before being included.

The point is that retail traders are being influenced by exciting stories and speculating on stocks without caring whether they are profitable. This carries over from the larger AI story to smaller speculations.

The record amount of margin debt that is driving the stock market continues to rise rapidly. (These are debit balances in customers’ securities accounts. I made a nifty chart but it wouldn’t copy into this post.)

09/01/20 654,324
10/01/20 659,313
11/01/20 722,118
12/01/20 778,037
01/01/21 798,605
02/01/21 813,680
03/01/21 822,551
04/01/21 847,186
05/01/21 861,626
06/01/21 882,103
07/01/21 844,324
08/01/21 911,545
09/01/21 903,117
10/01/21 935,862
11/01/21 918,598
12/01/21 910,021
01/01/22 829,637
02/01/22 835,255
03/01/22 799,660
04/01/22 772,940
05/01/22 752,944
06/01/22 683,445
07/01/22 696,781
08/01/22 687,787
09/01/22 664,009
10/01/22 649,618
11/01/22 643,783
12/01/22 606,659
01/01/23 641,228
02/01/23 624,379
03/01/23 645,429
04/01/23 631,949
05/01/23 644,170
06/01/23 681,228
07/01/23 709,834
08/01/23 689,185
09/01/23 680,846
10/01/23 635,276
11/01/23 660,887
12/01/23 700,774
01/01/24 701,975
02/01/24 742,963
03/01/24 784,136
04/01/24 775,464
05/01/24 809,431
06/01/24 809,322
07/01/24 810,835
08/01/24 797,162
09/01/24 813,211
10/01/24 815,369
11/01/24 890,852
12/01/24 899,168
01/01/25 937,253
02/01/25 918,144
03/01/25 880,316
04/01/25 850,558
05/01/25 920,960
06/01/25 1,007,961
07/01/25 1,022,548
08/01/25 1,059,723
09/01/25 1,126,494
10/01/25 1,183,654
11/01/25 1,214,321

When a bubble bursts the margin calls often force the sale of all assets, both good and bad. The baby gets thrown out with the bath water. This is what makes bursting bubbles so dangerous to stock investors.

The Price-to- earnings ratio based on average inflation-adjusted earnings from the previous 10 years, known as the Cyclically Adjusted PE Ratio (CAPE Ratio), is 40 compared with the long-term median of 16. This aligns with the other valuation metrics discussed above.

The stock indexes are continuing their rising trend despite some noise. VIX is low. The Fear & Greed Index is still in Fear but it is rising. The trade is risk-on as stocks and junk bonds have been rising relative to the 10 year Treasury.

USD has stabilized (with noise) at the lower level it reached in April 2025. Gold, silver and copper are rising strongly. Bitcoin has stabilized after a steep drop.

The Federal Reserve cut the fed funds rate by 0.25% as expected last week. There was lack of unanimity since the economy is still growing strongly and inflation is rising faster than the Fed’s target.

The Atlanta Fed’s GDPNow model estimate for real GDP growth (seasonally adjusted annual rate) in the third quarter of 2025 was 3.6 percent on December 11. That is a strong growth reading. The Cleveland Fed’s inflation forecast is over 3% for the CPI which feeds into TIPS, I-Bonds and the adjustment for Social Security.

image
image1201×272 5.24 KB

The Treasury yield curve shows an important pushback from the bond market. Even though the fed funds (overnight) rate dropped the longer maturity yields rose. This shows that the bond market is pricing in higher uncertainty regarding inflation and the overwhelming borrowing needs of the government. The bond market isn’t following the Fed in cutting yields.

The Chicago Fed’s National Financial Conditions Index (NFCI), which provides a comprehensive weekly update on U.S. financial conditions in money markets, debt and equity markets, and the traditional and “shadow” banking systems, shows that already loose conditions are getting looser.

However, financial stress has been increasing. As a result, the Federal Reserve announced that it will start buying $40 billion per month of short-term Treasuries to increase liquidity in the banking system. They say that this isn’t Quantitative Easing, but if it walks like a duck and quacks like a duck…This will reverse the oh-so-gradual decline of their prior gigantic QE assets.

Whatever it’s called, pumping more money into banks will increase asset prices.

There are no dramatic news stories that will shift the markets this week.

The METAR for next week is sunny.
Wendy

8 Likes
2

The story that the investment iin AI will be justified by growth in the economy is just to placate the unaware. The major benefit to a company using AI is to provide better customer service/satisfaction in conjunction with lower costs through getting rid of personnel. The only way to get a return on the vast pile of money going into the system is a vast cut in the system’s cost level.

Jeff

8 Likes
3

Yes, this is where the rubber meets the road. Unfortunately a lot of hopes and dreams are going to be squashed on that road … someday.

There’s margin, there’s endless optimism and seemingly endless investment to go with it - a fair portion of it now financed by debt. The debt it “guaranteed” by massive corporations with stunning cash flows, but they’ve chosen to structure it in a way that hides the true nature of the debt from their investors, while at the same time also hiding it from the accountants. (Don’t get me wrong, the numbers guys know exactly what’s going on.)

The Special Purpose Vehicles, or securitization, or dummy corporations as we used to call them make them disappear from corporate balance sheets but the obligations are still there. I remember the MCI/Worldcom debacle (not precisely the same), and the Global Crossing and Nortel and Lucent and 360 Networks and, well, all of them because of happy, cheery forecasts untethered from reality.

There another issue that’s been ringing in my head, and that’s how we expect (née: demand) government ride to the rescue every times the market belches. The SVB bank collapse is the most recent, but there’s the S&L collapse of the 80’s, the excesses of 2008 (Lehman, Bear Stearns, etc) the magic appearance of bond quant active easing and other remedies, and everyone has forgotten that bad stuff happens.

Really bad stuff, sometimes. (I’m not predicting anything like that, just noting the difference in attitude of investors nowadays.)

Anyway, AI is being used to promulgate all sorts of nonsense, and some of it might even come true! I’m one who think a lot of it might not, too.

9 Likes
4

The article below says Microsoft is cutting way back on AI due to slow Co-Pilot sales. I get Co-Pilot pop-ups a lot in Office dcouments, but I haven’t needed it to do anything.

However, AI is still valuable for lots of tasks. We’ll see if it pays off.

https://www.extremetech.com/computing/microsoft-scales-back-ai-goals-because-almost-nobody-is-using-copilot

5 Likes
5

Do you know what bankrupted the fiber optic industry?
Does it apply to data center buildout?
Hint, you can’t go broke as long as you can pay your bills!

Not necessarily. The other way is to generate sufficient cashflow or a combination of both.

The Captain

1 Like
6

This is what I find maddening. How you can get debt and obligations that you are on the hook for “off balance sheet” to make it look like you don’t have the exposure.

2 Likes
7

Yes. Overcapacity based on wildly optimistic projections.

Nobody can say with authority at the moment, but based on the revenue uptake it appears that they are building jet turbine engines to run a Model T automobile. Perhaps there will be enough paying customer demand for videos featuring cats dancing in top hats to the tune of synthetically created music. It’s, hard to say.

5 Likes
8

Yup - and in order to get that cashflow, they have to return substantial benefits to those who rent them. And, since material costs and plant infrastructure are fixed - that leaves a savings in labor.

Jeff

2 Likes
9

The Bulls say the 1990s overcapacity produced dark fiber… And today is different cause TANSTADDC (there ain’t no such thing as dark data centers).

Back in the mid 1990s there was no dark fiber. The big boys started making big $ laying fiber.
Then medium boys jumped in and made $ laying fiber.
Then everybody and his dog started laying fiber, at which point a lot was unused (dark). The installers made money, but the owners struggled.

Today, the big dogs (NVDA, AMZN, meta, TSLA, etc) are building data centers, that are being used
Then a couple years ago MARA n other BTC miners started selling compute time, and a year or so ago CRWV showed up, then NBIS.
More recently IREN.
And apparently their data centers are being used.

I’ve been watching for smaller, local data center builder/operators.
Ie the beginnings of “dark data centers”.
Last week I saw TSS, a tiny local central TX IT company that’s adding data center wannabe to its resume.
https://finance.yahoo.com/quote/TSSI/

The dark data centers are coming.
IMO.

:factory:
ralph

5 Likes
10

There’s another reason, besides lack of demand, that could cause “dark data centers”. What if there is some sort of mishap at a large power plant, or some sort of mishap in a portion of the grid, or whatever, and the choice is to keep supplying power to a data center OR to 120,000 homes? If the event is only expected to last a few hours, as happens all the time, the problem can usually be solved (brownouts, blackouts, expensive imported power, etc). But if the event will last longer, then a decision has to be made. And there obviously would only be one decision made due to politics, the homes will get the power and the data center will go dark for a while.

1 Like
11

:slightly_smiling_face:
I think that “disaster induced dark” is a completely different interpretation of “dark” as in the bubble descriptions of “dark fiber” which is fear of no buyers, which cause “no profits”.

BUT. As part of the data centers’ efforts to “be a productive member of the communities in which they are building”, many data centers are promising to PROACTIVELY curtail their non-essential energy use during times when the humans need the energy.
Not just “emergency” situations, but also peak use periods.

This is being marketed, by the data centers, to investors as energy arbitrage opportunities.

:light_bulb::low_battery::high_voltage::flashlight:
ralph

I own some BMNR.
BMNR is part blockchain and crypto miner, part data center management, and DCaaS.
BMNR has specifically written into its business plan flexible energy use and curtailing in times of need so as to not strain the grid.

3 Likes
12

I’m trying to figure out how useful AI will be if it’s not available when a particular application is suddenly, perhaps repeatedly unavailable.

Let’s say it’s controlling a factory process. Or answering questions of search users. Or summarizing data. Or insuring that a robot is properly caring for an elder.

And suddenly, nothing. Does that seem a likely outcome? I think we all understand when there is a general power failure and everything stops, but if/when the AI, whatever it is doing, is unavailable for some period of time (driving a car? Flying a plane?) How often do you think the PTB, or the customers are going to say “Oh yeah, this is tolerable, because Grandma needs her air conditioning.”

Promises made are sometimes different than promises kept.

2 Likes
13

Does that seem a likely outcome?

No.
Let’s unpack the questions:
There first is that the obvious EDGE processes are operating at the local (edge) situation/task:

  • controlling a factory process

  • a robot is properly caring for an elder

  • driving a car?

  • Flying a plane?

  • I’m gonna put summarize at “edge”, too, but I think this is case dependent.

The mothership is not micromanaging the edge AI.

The second is that search queries are routed to data centers that are available. This is already happening - every second.
If right now, you submit a query and data center Tulsa is too busy your query will be routed to data center Utica, or to data center Bakersfield. Wherever there’s available capacity.

Third is essential vs non-essential.
BMNR does crypto mining. It will stop that mining, when electricity is needed for “elder care” or other more essential things.

Essential data center processes will continue.

So, no. I don’t think total stoppage is a likely outcome.

:alien_monster:
ralph

2 Likes
14

The lines blur for cases like these. Take my example above - a large power plant has an emergency and goes offline for a month or two. We have enough power for the data center OR for grandmas home (120,000 homes) so her care robot can charge itself each night, and can communicate with her physicians to determine which meds are needed each day, and can draw blood to send for analysis (or analysis it locally), help her bathe, dress, etc.

Yesterday I asked Grok a question on my laptop and it said “I’m too busy right now, try again later”. immediately I asked the identical question on my phone and it answered.

1 Like
15

That was the sales job. The reality, it turns out AI is best for replacing other forms of internet search.

2 Likes
16

Sure, I get that you can load shift some things sometimes. But there are certainly going to be applications which you can’t, or at least shouldn’t.

Data centers contain both huge data sets and the power to manipulate them (search, coalesce, evaluate, synthesize, respond.) How many times are you going to create the same, or virtually similar data sets in different centers around the country?

It can’t be that every time you want to pull a response (whether search, actuation of robot, etc) you pull from hundreds of different data centers around the country (world?). That would be vastly inefficient, but then replicating the same (or nearly same) information over and over so it is “local’ would be as well.

You’re driving a locomotive with AI because the train guys have decided it’s cheaper than humans (or use trucks, or assembly line, or whatever example you want.) Suddenly your data center says “Oh I’m too busy right now, try again later.” This is a Magic 8-ball worthy response, only slightly better than “Grandma needs her air conditioning so we’re shutting down for a while.” Meanwhile a train, truck, or assembly line somewhere in Kansas says “Oop, no data returned. Guess I better shut down.”

Now I’m exaggerating, obvioiusly there will be redundancies, but then those cost money too.

So all of these data centers have the same “library” of data from which to draw to give you the same (or nearly identical) answers? That seems odd, given that AI will be used for such disparate things as healthcare, financial fraud detection, quality control in manufacturing, drug synthesis, sewage treatment facility control, software development, urban traffic flow, X-ray diagnostics, advertising creation and targeting, social media summation, real time training, military operations, product recommendations, inventory management, and a thousand other things.

And while some things will be at the core and others at the edge, at some point you need to have them communicating - instantaneously - because a lot of functions simply can’t be “too busy now, try again later.”

This is a complicated world we’re stepping in to. I can’t help but wonder how it all turns out.

1 Like
17

Data center capacity or local availability will not be a determining factor for edge devices.

Models are developed - and trained - in massive compute data centers. Those rules and a delivery model are sent to the edge for industrial control (local).

Industrial control - put more broadly:

continuous, onboard, connected localized control.

Any process/safety interlocked system will have local directives and local signals/flags which limit actions or support readiness for further actions.

Local control only: Acceleration limits, range of motion, next step preparation, next step readiness, current step status, state and progress/variance against tasks

When “AI” is given permission to act, make inferences, react to flags/alerts/alarms/or other permissive, it will not do so with a cloud connection that is not local to the location.

Cloud connections will be only for OTA updates, status confirmations, etc. There will be no need to constantly tether “datacenters” to edge devices. All fixed or mobile control systems are “edge”.

Data center capacity or local availability will not be a determining factor for edge devices.

1 Like
18

I agree! completely!
Lemme go off on a slight tangent.
Helen, my roomba, is not connected to the internet.
I step over to her, I push her button, I take a deep breath and I get real high and I scream at the top of my lungs “Helen, GET TO WORK!”.

Helen does a ‘random’ search pattern, every time. Cause no internet connection. I assume there’s some ‘updates’ or some such, and likely some ‘security’ patch/es.
It turns out that Helen’s mothership is now in China.
Do you own a Roomba? What iRobot’s bankruptcy means for customers

{ iRobot, the maker of the popular Roomba vacuum cleaner, filed for bankruptcy protection on Dec. 14, saying it would go private after being purchased by its Chinese manufacturer. }

Yessir! Interesting times.

:slight_smile: ralph

1 Like
19

As I understand it, no. But kind of. Back when Google was first mapping the Internet, they couldn’t afford enterprise grade hardware, but the consumer stuff had higher rates of failure. So they broke everything into chunks, and stored copies of the chunks on different hard drives. If a hard drive failed–which they knew was going to happen–or a even couple hard drives failed, they could still retrieve all the data.

What they do now is a lot more sophisticated and I don’t fully understand it, but they still break everything into chunks, and the chunks and copies of the chunks are not only spread across different racks in the datacenter, they are spread across different data centers. So, there isn’t a file somewhere called “Goofyhooy’s Inbox.” Instead there are bits of your inbox all over the place.

So if there is a regional power outage in say, Texas that knocks out five data centers, it doesn’t matter because copies of the chunks are spread all across the world. You can be in Tokyo and log into your Gmail account just fine.

If all this sounds very expensive, you’d be correct.

2 Likes
20

I used to work for Oracle, on SPARC server processor core (from the old Sun Microsystems days), back before they had their big cloud push. Used to be you could run Oracle software on machines you bought. Or on Intel boxes you bought from Oracle. Or on SPARC/Solaris boxes you bought from Oracle if you needed the highest performance and highest reliability. (RIP, SPARC).

Then came the Cloud push, which included the use of SPARC/Solaris boxes in the Oracle cloud. Some of us on the team worried what that meant for us as a team, and we were told “we need 3X the number of boxes our customers would normally need in order to give the up-time we are promising”.

So yeah, it gets expensive.