Note from Tom: I’ve started a new blog on Substack
called “Tom Alrich’s blog, too”. After August 11, Substack will be the only
source for my new posts, where they will only be available to paid subscribers.
Before that date, new posts will be available for free on Substack and Blogspot. A subscription
to the Substack blog costs $30 per year (or $5 per month); anyone who can’t pay
that should email me. To have uninterrupted access to my new posts, please open
a paid account on Substack or upgrade your free account to a paid one.
Last October, I wrote a post
that described an incredibly powerful presentation at the just-concluded annual
NERC GridSecCon conference in Minneapolis. That presentation was by a meteorologist
named Sunny Wescott, who works for CISA. You can read that post; I’ll just
point out that she showed how far we’ve already traveled on the road to climate
disaster. However, she also showed there’s a lot of activity today that can
still save us, since many people are working in many innovative ways to rectify
the situation.
The point of my post was that the humongous amount of power
needed to train large language models is now working against combating climate
change – or more correctly, AI’s power requirements are on net accelerating
climate change. This might be surprising to you, since the large AI processing players
like Microsoft, AWS, Meta, and Google have until recently loudly
touted their commitment to renewable energy. They have backed their words with
commitments for energy from wind and solar farms, and even nuclear plants.
Unfortunately, there are only so many large renewable energy
sources to go around (plus recent political developments are, if anything, working
to decrease future availability for all renewables sources except nuclear power[i]).
In the above-linked post, I also mentioned that, at the conference, a friend who
works for a large electric utility pointed out that a 1,000MW coal plant they
used to own in North Dakota (and which I had visited with him 7 or 8 years ago)
had just been purchased to power a huge greenfield data center nearby.
Some other large coal plants in the US have also recently
been purchased, or at least removed from the list of plants to be
decommissioned. I wrote in the post that it was likely that most US coal plants
that aren’t already far down the path to being decommissioned will be given a
similar new lease on life. And other countries like Saudi Arabia and India,
which already have many large coal plants, are almost certain to use those
plants to attract AI investment, also setting back their climate change
efforts.
Thus, the drive for AI has already started to set back North
America’s efforts to delay climate change. However, it turns out there’s
something just as important that AI is also setting back: the effort to set the
US on a low-inflation and low-interest rate path to prosperity.
The inspiration for this statement was someone else who I
greatly admire, although he’s in a completely different field from Sunny
Wescott. This is Greg Ip, who writes about capital markets for the Wall
Street Journal and is IMHO the best economist of the many that work for
WSJ. I’ve never heard anybody rave about him like I’ve heard raves about Sunny,
but I’ve always been impressed by his ability to analyze well known facts and
come to important conclusions that have otherwise been overlooked.
Greg started his column in the August 2 edition of the
Journal (the link is here,
but it’s behind a paywall. I can send you a PDF if you email me) with these
three paragraphs that precisely summarize what he said:
In the past two weeks one big tech company after another
reported blowout earnings amid a wholesale embrace of artificial intelligence.
Look a little closer, and a more unsettling side to the AI
boom emerges. All the spending on chips, data centers and other AI
infrastructure is draining American corporations of cash.
This underscores the hidden risks from the AI boom. No one
doubts its potential to raise growth and productivity in the long run.
But financing that boom is straining the companies and
capital markets.
He continues:
Since the first
quarter of 2023, investment in information processing equipment has expanded
23%, after inflation, while total gross domestic product has expanded just 6%.
In the first half of the year (2025), information processing investment
contributed more than half the sluggish 1.2% overall growth rate. In effect, AI spending propped up
the economy while consumer spending stagnated.
Much of that investment consists of the graphics-processing
units, memory chips, servers, and networking gear to train and run the large
language models at the heart of the boom. And all that computing power
needs buildings, land and power generation.
He goes on to explain that big tech companies used to be
considered “asset-light”. That is, most of the investment they made was in
relatively low-cost intangible assets like intellectual property and software.
Therefore, the boatloads of cash that they brought in mostly went to the bottom
line, making them tremendously profitable.
However, the same companies (he cites Alphabet, Amazon, Meta
and Microsoft), even though their established businesses are still bringing in lots
of cash, are investing huge amounts in their AI infrastructure; this
significantly lowers their profitability. Meanwhile, two fast-growing AI
companies that don’t have established businesses, Open AI and Anthropic, are
both losing money.
However, I don’t recommend going to Sam Altman and offering him
$1 for Open AI, while agreeing to assume all their debt – that strategy isn’t
likely to succeed. In a section ominously titled “Dot-com echoes”, Greg points out
that investors are valuing all these newly asset-heavy companies as if the
investments they are making in AI are as likely to be profitable as the
investments they made in the good old asset-light days – when every investment
in say a new version of Windows was almost certain to bring in tons of cash
(does anyone else remember the huge hype over the rollout of Windows 95? That
hype was money well spent, since Microsoft made a lot of money on that version
of Windows – despite the fact that Windows 95 was and is a security nightmare).
Greg continues,
For now, investors are pricing big tech as if their
asset-heavy business will be as profitable as their asset-light models.
So far, “we don’t have any evidence of that,” said Jason
Thomas, head of research at Carlyle Group. “The variable people miss out on is
the time horizon. All this capital spending may prove productive beyond their
wildest dreams, but beyond the relevant time horizon for their shareholders,”
he added.
In the late 1990s and early 2000s, the nascent internet boom
had investors throwing cash at startup web companies and broadband
telecommunications carriers. They were right (that) the internet would drive a
productivity boom, but wrong about the financial payoff. Many of those
companies couldn’t earn enough to cover their expenses and went bust. In
broadband, excess capacity caused pricing to plunge. The resulting slump in
capital spending helped cause a mild recession in 2001.
Greg adds that he’s not expecting a
stock market crash like the dot-com bust of early 2000 or the (mild) dot-com
recession of March to November 2001 (although the September 11 attacks clearly made
that slump less mild than it otherwise would have been). His point in the first
part of his column is summarized in one of the sentences quoted above: “All the spending on chips, data centers and other AI
infrastructure is draining American corporations of cash.”
In other words, I believe Greg is saying that, just like AI
is straining the electric power industry, it’s putting much more strain
on the companies that are supposedly benefiting the most from the AI boom: the
companies developing the software and running the huge data centers that train
the models. On the other hand, I’m not suggesting anybody start a tag day for
Microsoft or Google, since their investment will inevitably pay off; the
question is when.
Shareholders in companies making the big AI investments are likely
to be disappointed if the economic returns show up ten years later than
anticipated. In fact, this is almost exactly what happened with the huge
increase in office productivity that was anticipated after the IBM PC was
introduced in 1981. That increase didn’t happen until the 1990s, in part
because that’s when networking technology became powerful and cheap enough that
all those standalone PCs on desktops could now work together to share databases,
internet connections, etc. (just ask Larry Ellison, founder of Oracle). Until that
happened, there wasn’t much office productivity gain.
The last section of Greg’s article is called “The
interest-rate effect”. This refers to another effect of the huge cash demands
of AI investment. Greg introduces this section by saying,
After the global financial crisis of 2007-09, big tech was
both a beneficiary of low interest rates, and a cause.
Between that crisis and Covid-19, these companies were
generating five to eight times as much cash from operations as they invested,
and that spare cash was recycled back into the financial system, Thomas, of
Carlyle Group, estimates. It helped hold down long-term interest rates amid
high federal deficits, as did inflation below the Federal Reserve’s 2% target
and the Fed buying bonds.
In other words, during big tech’s asset-light
period (between approximately 2010 and 2020), the big tech companies were
generating so much cash that they helped hold down long-term interest rates in
the US, although two other things also helped: low inflation and the fact that
the Federal Reserve Bank was buying bonds – i.e. retiring US debt (when the Fed
buys bonds, they inject money into the economy).[ii]
Greg now compares that period
(which ended just five years ago) to the situation today, which is completely
different:
1.
Government deficits are larger now than five
years ago, meaning there’s a much greater need for money.
2.
Inflation is now above 2%, the Fed’s target
rate.
3.
Instead of on net buying back bonds from the
private sector, the Fed is now selling more bonds than it’s buying. This has the
effect of restricting the money supply, lowering bond prices and raising
interest rates (interest rates move inversely with bond prices).
4.
Corporations now “face steep investment needs to
exploit AI and reshore production to avoid tariffs.” This further decreases
available cash for corporations, while having no near-term positive effect on
profitability.
Greg concludes this section, as
well as the column, by saying
All this suggests that interest rates need to be
substantially higher in the years ahead than in the years before the pandemic.
That is another risk to the economy, and these companies, that investors may
not fully appreciate.
Here’s my summary of Greg’s column:
1.
In the near term, AI investments by major tech
companies are not improving their bottom lines. At the same time, those
investments are greatly decreasing the amount of cash available for other
investments (surprisingly enough, AI isn’t the only good area for
investment today! For example, I’m receiving a lot of spam emails for burial
insurance. Do they know something I don’t?).
2.
This won’t cause a stock market crash, but it may
lead to a lot of disappointed investors and falling stock valuations,
especially if the anticipated huge returns don’t show up for 5-10 years longer
than expected.
3.
The above means the money supply is tightening
and interest rates are rising. Don’t look for either of these trends to be reversed
for a long time.
Note that Greg doesn’t mention one big cost of AI, probably
because it doesn’t affect the profitability of AI companies (which is his main
concern, since he writes about investments). This is the huge, mostly uncompensated
charges that electric ratepayers will have to pay to build out (or refurbish)
the power generation capacity to support AI in the future.
Because of the scale of this unprecedented buildout, the
easiest way to finance it is to bill the ratepayers for a lot of it. This is
where the likely delay in returns on AI investment will hurt the most, since
some ratepayers in some states will end up paying more for their fair share of
the additional investment than will others. By the same token, some AI companies
will pay for more than their fair share of that investment, although in general
I think the companies’ share, vs. the ratepayers’ share, is more likely to be
too low than too high.
I would like to see some organization – EPRI, NARUC, the US
Congress (!), the ISO/RTOs, etc. – conduct a comprehensive study of the
question of how to fairly allocate the costs of the required grid buildout across
the US (and perhaps in Canada, at least in the provinces like Ontario and Quebec
that sell a lot of power to the US). It won’t be easy at all, but otherwise I think
we’ll end up with a big mess on our hands and lots of bad feelings to boot.
This concludes the first post in this series; it discusses
the costs of the AI buildout. The second (and I believe last) post will discuss
the benefits of the AI buildout and try to weigh costs and benefits. Spoiler
alert: I believe that, in the long run, there’s no question that AI will provide
a huge net benefit, not only to North America but to the whole world. However,
as the great economist John Maynard Keynes said, “In the long run, we’re all
dead.”
If you would like to comment on what you have read here, I would love to hear from you. Please email me at tom@tomalrich.com, or even better, sign up as a free subscriber to the Substack community chat for my subscribers and make your comment there.
[i]
Nuclear energy isn’t technically renewable, since nuclear plants require a
constant supply of uranium. However, nuclear plants don’t produce greenhouse
gases like fossil fuel plants do.
[ii] In
fact, the Fed borrowing money by selling bonds is the main source of the money
supply. Currency is just a small part of money; most of it is in the form of
bank deposits. When the Fed wants to inject money into the economy, they buy
bonds. When they want to withdraw money from the economy, they sell bonds.
These are called “open market operations”, and they’re the key instrument of
monetary policy.
No comments:
Post a Comment