An insightful financial analysis examining Anthropic's preparation for its initial public offering at a reported multi-trillion-dollar valuation. The breakdown explores artificial intelligence stock market trends, valuation methodologies like discounted cash flow models, total addressable market projections, circular financing loops, and competitive price pressures within the tech sector.
The video analyzes the precarious financial state of the AI industry, focusing on the massive disconnect between speculative valuations—such as Anthropic's reported $2 trillion IPO target—and actual financial performance. It warns that the current boom is driven by inflated market projections and circular funding loops, mirroring the conditions of the 1999 dot-com bubble.
Anthropic, the AI company whose CEO has just asked his industry to slow the pace of AI development
and ask the government to regulate everyone to make sure that they do, is expected to go public
in the next few months, hoping to raise more money more quickly than any other company in history.
Anthropic is reportedly seeking a valuation of around 2 trillion dollars, which The Economist
points out is about as much as the 10 biggest tech IPOs of all time added together. So this is
quite a big one, and on paper they've picked the perfect moment. The Nasdaq, an index of mostly tech
stocks, is at a record high. A survey of purchasing managers has just shown US business output growing
at its fastest pace in five years, and yet IPOs are getting pulled. Jay Ritter, who runs the IPO
initiative at the University of Florida, told Bloomberg that this is especially surprising
with the Nasdaq at a record. Anthropic's public filing was expected to be released as early as
late August, but still hasn't appeared. OpenAI has pushed its listing out to next year. Meanwhile,
Nvidia, the company making all of the money from AI, is the most valuable company in the world,
with its share price up more than 1600% over the last four years. But relative to the profits it's
expected to make, Bloomberg points out that it's the cheapest it's been in over a decade. So today
I wanted to work out why the company selling the shovels looks cheap, while the company digging
with them is hoping to be valued at 2 trillion dollars. To be clear here, I'm not one of those
people who claim that AI is useless. People clearly find it useful, but a great technology
can still be a terrible investment if you pay too much when you buy in. And that's what this video
is about, the price. So first up, how do you get to 2 trillion dollars? Well, normally investment
bankers have two main ways of valuing a company. They can forecast its future cash flows and
discount them back to today, which is called a discounted cash flow model, or they can look at
what similar listed companies trade at relative to their sales or profits or some other key metric
and apply an industry multiple. Then they take the company on a roadshow, ask the big investors
what they'd be willing to pay and set the price wherever the orders are. The spreadsheets are
there to make that number look somewhat grounded in reality. Anthropic poses challenges for both
of these methods. There aren't many similar listed companies to compare to and forecasting the cash
flows of a company in a new industry whose revenue has grown more than tenfold in a year
is going to be a struggle. Its revenue was running at around 65 billion dollars a year in August,
but as the FT's lex column points out, these unofficial numbers should be handled with kid
gloves. Now, a discounted cash flow model needs an interest rate as the time value of money matters.
And this is the one dark cloud over that perfect picture I painted earlier. The same survey that
showed that the economy is booming sent the 10 year treasury yield to 5.23 percent on Thursday
evening, the highest it's been since 2004. That is roughly what you'd expect when an economy is
running hot, inflation is sticky and the government is running a large deficit. The Fed raised rates
earlier this month. Matt Kennedy of Renaissance Capital told Bloomberg that rising rates are a
double whammy for companies like this. They reduce the value of profits that are years out in the
future and they make it more expensive to borrow money to build data centers. This might help
explain the strange IPO cancellations we've been seeing. Holtec, a nuclear power company,
with nuclear being the power source the tech industry is counting on to run its data centers,
postponed its IPO last week, citing market conditions. SB Energy, SoftBank's data center
developer, filed before Labor Day and still hasn't started marketing its shares. We'll come
back to them in just a moment. There have been just three IPOs since Labor Day and five of the
ten largest listings this year are now trading below their offer prices. And for what it's worth,
the last time the Fed was raising rates into a boom like this was 1999. So just how crazy is
$2 trillion in valuation? Well, before I get into that, let me tell you about this week's
video sponsor, Plod. You know how the best ideas can often hit you at the worst times,
driving, walking, at the gym or in between meetings. That's the exact problem Plod was
built to solve. Plod isn't just a voice recorder or a note-taking device, it's more like a searchable
memory system. It helps turn meetings, calls, conversations and ideas into a searchable archive
that you can always come back to later. The Notepin S is great for on-the-go capture,
quick thoughts, in-person conversations or moments when an idea comes up and you want to save it with
a simple press of the button. Then when I get back home, Plod turns everything I've said into
an organized summary and to-do list. It's wearable. I've had mine clipped to my jacket all
week. This one is the Note Pro I talked about a while ago. It's better for more deliberate
recordings, meetings, interviews, phone calls and professional conversations you'll want to
return to later. And with Ask Plod, you can easily search across everything you've captured
instead of digging through old voice memos. Everything syncs to the desktop app too,
so it's all there when I'm back at my computer. Plod comes with a 30-day free return policy.
You won't know how much it helps you until you actually try it. Check it out using the QR code
or the link in the description. And don't forget to use code PBOIL to unlock an exclusive discount
at checkout. Because so many people talk about AI labs in terms of revenue rather than profits
I'm reminded of a famous quote from Scott McNeely, the co-founder of Sun Microsystems.
In 2002, after the dot-com crash, he explained to Businessweek just how crazy it had been
to pay ten times revenues for his company's stock at the peak. And Sun was a good company.
Its servers ran much of the early internet. The problem was just the price. At ten times
revenues, he said, to give you a ten-year payback, I have to pay you 100% of the revenues
for ten straight years in dividends. That assumes I can get that by my shareholders.
That assumes I have zero cost of goods sold, which is very hard for a computer company.
That assumes zero expenses, which are really hard with 39,000 employees. That assumes I pay
no taxes, which is very hard. And that assumes you pay no taxes on your dividends,
which is kind of illegal. And that assumes with zero R&D for the next ten years,
I can maintain the current revenue run rate. Now, having done that, would any of you like
to buy my stock at $64 a share? Do you realize how ridiculous those basic assumptions are?
You don't need any transparency. You don't need any footnotes. What were you thinking?
It's maybe the only time that I can think of that a CEO explained to investors in detail
why they shouldn't have bought his stock. Unfortunately, though, in 2002, it was a bit
too late. As crazy as those assumptions were, McNeely was still being a bit generous as he
left out the time value of money. At the peak of the bubble in early 2000, the 10-year paid
around 6.5%. At that rate, a dollar you receive in 10 years is worth a lot less than a dollar
would have been worth that day. So even with no costs, no taxes and no R&D, an investor paying
10 times revenue wouldn't get their money back in 10 years. It would take about 17 years.
Anthropic, if it was priced at $2 trillion, would be trading at around 31 times its revenue.
So let's be even more generous than McNeely. Assume Anthropic has no costs, no staff,
no taxes and no electricity bill and pays out every dollar it brings in to shareholders forever.
And instead of adding a premium for risk like you normally would, let's discount it at the
treasury rate. Under those assumptions, you never get your money back, not in 10 years,
not in 100. All of that revenue forever is worth about $1.27 trillion today, about three quarters
of a trillion short. So almost all of the $2 trillion is a bet on growth, and the higher
interest rates go, the higher that growth has to be. So if the normal math won't get you to $2
trillion, then you need new math. Enter the TAM, which stands for Total Addressable Market. It's
the amount of money a company might get if it captured an entire industry. While the metric
has been around for a long time, the idea really caught on in the late 1990s when the internet
analyst Henry Bludgett used it to argue that Amazon, which was trading at $240 a share at the
time, was worth $400 a share. Amazon then got there within weeks, which made Bludgett famous.
He was later banned from the securities industry for life, which made him even more famous.
The FT's Lex column has a nice tour of how these numbers have been growing for the AI industry.
In May, SpaceX, which I had always thought of as a rocket company, but which also makes the
chatbot and owns Twitter, told investors that the market for enterprise apps, including its AI
products, was $22.7 trillion, which is about three quarters of the entire US economy.
Then, as Lex describes it, TAM inflation kicked in. The Wall Street Journal is reporting that
Anthropx filing may cite a $30 trillion market, and last week Morgan Stanley estimated that generative
AI could tap $60 trillion, so about half of everything produced on the planet in a year.
Lex notes that Morgan Stanley is likely to be an underwriter on Anthropx IPO, which is of course
just a coincidence. Bankers are famously conservative when estimating the value of
things they can charge a percentage of. In just four months, the addressable market grew
by $37 trillion, which is faster than Anthropx revenue
and quite a bit faster than the economy it is supposed to be carved out of.
You might be surprised to hear that companies don't always capture much of the TAM
that their valuations are based on. Uber touted a $12.3 trillion TAM when it listed in 2019.
Its annual revenue today is under $60 billion, so they still have 99.5% of their market ahead of
them. WeWork had a $3 trillion addressable market and then went bankrupt. That market, I guess,
went unaddressed. Then there are the really big numbers. Anthropx's internal research arm
has modelled an extreme scenario in which AI adds over $10 trillion to US GDP by 2030,
which Lex's back-of-the-envelope maths turns into about $100 trillion of equity value today.
Beyond that is recursive self-improvement, AI that designs better AI, which then goes on to design
even better AI, where so much economic value is added that money and asset prices stop meaning
anything at all, which is one way to hit your price target. Elon Musk has predicted that AI
and robots will make work optional and that money will stop being relevant. For someone who expects
money to become meaningless, he sure has been collecting a remarkable amount of it. I should
point out that in such a scenario, I'll probably be a lot wealthier than Elon Musk, as while he
might be worth a trillion dollars, I own this $100 trillion Zimbabwean node, which has also
become irrelevant. So how close are we to recursive self-improvement? Last year, Google
DeepMind's Alpha Evolve found a better way of multiplying certain matrices than the one
mathematicians have been using for 56 years, which is impressive. But as House of LAI, a YouTube
channel run by a computer scientist with a PhD, points out in her recent review of the research,
the impressive results all had a scoreboard, a test with a right answer. Anthropic's own automated
researcher closed about 97% of the performance gap on its task, partially by finding ways to
game the experiment. And when its best idea was tried at production scale, the improvement was
half a point, which was within the noise floor. To be fair to Anthropic, they published that
themselves, which is more than most companies do with a disappointing number. If you'd like to
learn more about where the research stands, I'd strongly recommend her video, which I've linked
to in the description. It makes a nice change from numbers just being pulled out of a hat.
Back on earth, 18 months ago, Anthropic expected its 2027 revenue to be $12 billion,
which in a story full of optimistic forecasts, is the one that turned out to be quite pessimistic.
And in August, Aswath Damodaran, the Dean of Valuation, wondered on the ProfGMarket's podcast
whether Anthropic's target might slip from $1 trillion to $800 billion as some of the shine
came off. Seven weeks later, the talk is off $2 trillion. Forecasting in this area is hard
in both directions. But anyhow, so far we've been talking about markets that don't exist yet,
but the AI boom has also found a way to value buildings that don't exist yet.
SB Energy, SoftBank's U.S. data center developer, which I mentioned earlier,
is hoping to go public at a valuation of around $50 billion. Now, $50 billion is
quite an achievement for any company, but it's especially impressive when you consider that SB
Energy hasn't switched on a single data center yet. As the FT's Lex column points out, they've
never built one themselves, although they did recently buy a consultancy that has built 15.
Lex also worked out that at $50 billion, SB Energy would be valued at around 400 times its
annual earnings before interest, taxes, depreciation and amortization. Its own
prospectus says that it will need $174 billion to build what it's already promised to build,
its new debt priced on Thursday yielding 9.75%, the largest junk bond offering on record,
and its lenders won't let it pay a dividend until 2029. It's now pushed back its IPO
while it waits for investors to feel better about data centers.
Now, I can't lie to you, numbers like that have inspired me. I'm currently drafting the paperwork
to list my own company, Boyle Compute on the NASDAQ. We also have no working data centers,
but we do have an excellent PowerPoint deck and I once successfully plugged in a Wi-Fi router,
which is the same sort of thing. I'll be looking for a valuation of around $25 billion
because I'm kind of hoping to attract value investors. Now, some traditional investors
might see a complete lack of functioning buildings as a drawback for an infrastructure company,
but in accounting terms, it's actually a tremendous advantage. A data center that's
still being built sits on the balance sheet as a construction in progress and construction in
progress doesn't get depreciated. Jim Chanos made a similar point about the tech giant's GPUs.
Chips that have been bought but not yet switched on don't get depreciated either.
So, if you think about it, the unbuilt data center may be the ultimate high margin business.
It uses no electricity, it needs no maintenance and nobody ever complains about latency because
the product doesn't yet exist. The only problem is that eventually someone will push you to build it
and building things is where it all goes wrong. Lex points to the Danish academic,
Brent Fleiberg's Iron Law of Megaprojects, that they come in over budget, over time,
over and over. Getting a grid connection can take up to 10 years in some states.
A Gallup poll found that 71% of Americans oppose having a data center built near them.
And S.B. Energy's own filings warn that if a project runs late, the customer can in some
cases buy it outright, potentially at a low price. Its main customer, by the way, is OpenAI.
The bad news is that the day you switch your data center on, the depreciation starts. S.B. Energy
only owns the buildings, the chips inside belong to its tenants, and chips age a lot faster than
buildings do. Paul Kedrowski in a research piece for Mann Group points out that lenders are
financing these projects as if they were long-lived infrastructure like commercial property,
when the chips inside will be out of date in a year or two, which is a bit like taking out a
30-year mortgage on an iPhone. So who's paying for all of this? Well, that's where things get
a little circular. Let's start with S.B. Energy's parent company, SoftBank, which is paying junk
too. It's currently marketing more than $11 billion of new bonds to fund its next payment
into OpenAI, and the FT reported that those bonds went out at a yield of 9.75%.
OpenAI needs the money, as the FT reports that it expects to burn through almost $280 billion
by the end of 2030. Some of that goes towards leasing S.B. Energy's enormous Ohio campus for
20 years, which gives SoftBank's data center company the revenue it needs to justify its
valuation. OpenAI is also an investor in S.B. Energy, and just to keep everyone properly motivated,
it holds warrants that pay out if S.B. Energy's valuation ever reaches $80 billion.
Then there's NVIDIA. S.B. Energy's filings show that NVIDIA bought $1.5 billion of its shares
in August at a 10% discount to the eventual IPO price, and will buy another $1.5 billion at the
IPO. NVIDIA has also agreed to guarantee up to $105 billion for the Ohio campus. The FT reports
that NVIDIA won't record any liability for this until OpenAI's leases begin in 2028, and in
exchange the campus will run exclusively on NVIDIA hardware for 20 years. If you watched my
video on Big Tech's hidden debt, this will look familiar. So, SoftBank borrows at junk rates to
fund OpenAI, which rents the building that NVIDIA guarantees and fills with NVIDIA chips. It's a
very tidy arrangement. Boyle Compute is currently looking for someone to guarantee my Wi-Fi router.
You see the exact same problem across the industry. According to the Wall Street Journal,
SpaceX is renting compute capacity to Anthropic for $1.25 billion a month. On the Prof G Markets
podcast, Ed Elson pointed out that 85% of Amazon's net income in its latest results
came from unrealized gains on its stakes in Anthropic and OpenAI. For Google, it was 87%
from its stakes in SpaceX and Anthropic. Aswad Damodaran's response was that he used to hate
valuing family-run conglomerates in Asia, because to value one company, you needed to first value
four others. Now, to value Microsoft, you have to value OpenAI. So, the industry has essentially
agreed to buy each other's products, guarantee each other's debt and mark up each other's
valuations. Which brings us to the biggest puzzle in the whole chain, NVIDIA. NVIDIA is the company
actually selling the picks and shovels in this gold rush. And business at NVIDIA is good. Bloomberg
reports that its sales have gone from about $27 billion four years ago to an estimated $410
billion this year. And its net income is expected to almost double. And yet, at less than 17 times
the profits it is expected to make over the next year, NVIDIA is trading near its cheapest level
in more than a decade. Investors are now paying about half as much for each dollar of NVIDIA's
profits as they were last year. Jensen Wong, NVIDIA's chief executive, told a Goldman Sachs
conference this month that NVIDIA is the world's first and only growth value stock, telling the
audience that the company is incredibly misunderstood. It's not every day that the boss
of the world's most valuable company pitches his stock to value investors. Although, as the founder
of Boyle Compute, I know how he feels. So, why is the shovel seller cheap? You might think the market
simply punishes reality and rewards promises. But there are four fairly sensible explanations.
First, the market is treating NVIDIA like a cyclical company at the top of its cycle.
Its growth margin was 75% last quarter, but analysts expect it to slip below 72%
by the end of the year, partially because NVIDIA has to buy memory chips from other companies and
those companies have been raising their prices. The memory maker Micron shares are up more than
180% this year. On top of that, NVIDIA's biggest customers like Meta and Alphabet are building
their own competing chips in-house. David Russell of TradeStation told Bloomberg that NVIDIA's
margins are more likely to go south than improve. Cyclical stocks often look cheapest when profits
peak because investors back away, not trusting that the peak will last. Second, NVIDIA's revenue
is everyone else's spending, and a lot of that spending depends on labs that are burning cash.
TCW's Eli Horton told Bloomberg that NVIDIA's multiple only really makes sense if AI spending
slows, either because the tech giants pull back or because of a regulatory framework that delays
or shuts things down, which is more or less what Anthropic's CEO has just asked for. The Monday
after his essay came out, the Maine Semiconductor Index fell almost 6%. The Economist and Lex have
both pointed out that a slowdown could be good for the lab's profits. It saves them a fortune
on computing, which is great for them and less great for the companies selling the computing.
Third, uncertainty can be valuable. Analysts know roughly what NVIDIA will sell next year.
An AI lab could go bust or it could capture a slice of that $60 trillion addressable market.
The finance academics Lubos Pastor and Pietro Vernassi argue that uncertainty about a young
company's future profits can raise its value, and Barberos and Huang argue that investors overpay
for stocks that look like lottery tickets. Fourth, the two prices are set by very different people.
NVIDIA is priced every second by millions of investors, including short sellers who can bet
against it. Anthropic's price has been set in private funding rounds, partially by cloud giants
whose own profits go up when its valuation does, and nobody in a private round can bet against it.
The economist Edward Miller argued back in 1977 that when pessimists can't bet against a stock,
the optimists set the price. Which brings us all the way back to Scott McNeely. Anthropic,
at $2 trillion, fails his test for reasonable pricing. NVIDIA, on the other hand, trades at
17 times its actual profits, and so far those profits are growing. It's the only company in
this whole story that passes the McNeely test, and it's the one being marked down.
So, why don't the labs all just agree to take a break? Well, because a slowdown would also shrink
the technological lead that allows them to charge premium prices, and their competitors don't appear
to be taking one. The FT reports that cheaper open-weight models, mostly from Chinese companies
like DeepSeek and Moonshot, are getting close to the performance of the frontier models,
and are taking market share. And on Tuesday, Anthropic and OpenAI both released cheaper
models on the same day, 40 and 50 percent cheaper than the ones they replaced. A price war is good
news for customers, but it's hard to think of a less convenient time to have one than right before
your IPO. Boyle Compute, I'm pleased to say, has been unaffected by any price wars, as we
don't have any customers. The price cuts are part of a larger trend. A new study from Epoch AI
estimates that the cost of getting a given level of performance from an AI model has fallen about
13-fold a year since 2023, which they say may be faster than for any other transformative
technology in history, including electricity, computing and DNA sequencing. They also found
that prices fall fastest right after a new top model comes out, meaning that whatever premium
a lab can charge for its best model doesn't last long. As Epoch points out, the AI boom is pushing
up the price of everything that goes in, chips, power, even electricians, while the price of what
comes out is collapsing. So this is great if you sell the chips, but worrying if you sell the
thinking. Last week, a startup called TypeSafe AI launched a model called Jev, which is aimed
almost exclusively at software developers. Its creators claim that it's up to 440 times cheaper
than the frontier models for simple decision-making tasks. Jev was built on $40 million of seed funding,
which is around 0.002% of $2 trillion. Eric Gleiman, the co-CEO of the corporate payments
company RAMP, told the FT that his company had cut its AI bill by 40% by sending different tasks
to different models, using routers that make it easy to switch providers. As he put it,
you don't need a Ferrari to go pick up your groceries. He even asked whether we've reached
intelligence saturation, a point that I feel I've reached some time ago. Now, tech executives
have always had a clever answer to cheap competition. Scott McNeely once said that
open-source software is free like a puppy is free, meaning that it costs you nothing up front and
then you have to look after it for years. That's the lab's best argument against cheap open models,
although we all know what cheap software running on cheap hardware eventually did to Sun Microsystems,
which ended up being sold to Oracle for a fraction of its peak value. To be fair to the labs,
there is a real bull case too. People are using these tools more and more. The FT charted data
from OpenRoute are a platform that routes requests between AI models, showing that weekly usage has
grown by about 25,000% since the start of last year. Alay Tavinsky, an economics professor at
Yale, found that 22.5% of Anthropic's users were still using its models a year later, compared with
about 13.2% for OpenAI. DeModern thinks Anthropic may be the one AI lab with real end customer
revenue. Analyst Ben Thompson has argued that if the labs build both the model and the software
tools around it, they may be able to lock customers in and keep the profits, although
RAMPS routers suggest customers are working quite hard to avoid being locked in. So, it's entirely
possible that the AI labs eventually figure out a highly profitable business model, but in the
meantime, they'd like to sell you some shares. None of this means that AI is a fad. It may well
change the world. The internet certainly did, but the internet gave us Yahoo, Lycos and AltaVista
long before anyone had heard of Google. And it's possible that AI ends up more like email,
something that we use every day that nobody makes much money selling. That would be great news for
and not such great news for anyone who paid $2 trillion for the company selling it.
An IPO is the moment when a company and its early investors decide that it's a good time to sell.
It's also the moment when this whole circular chain gets a real market price for the first time.
That's a bit of a trap for SoftBank. A big part of its balance sheet rests on its stake in OpenAI,
which was last valued privately at $852 billion. And Bloomberg reported in August
that SoftBank has taken out a $10 billion margin loan secured against its OpenAI shares. A listing
would let SoftBank turn that paper value into cash, but it would also put a market price on
shares that have never traded. If investors decide OpenAI isn't worth $852 billion,
SoftBank's collateral shrinks. Anthropic's IPO will be the first public test of what a pure
AI lab is worth, which will also put a market price on Amazon's and Google's paper profits.
The awkward thing for SoftBank is that OpenAI controls the timing. Sam Altman told Fortune
that given everything happening with safety, this would be an ill-advised moment to go public.
While OpenAI is reportedly in talks to raise money privately at a valuation of $1.2 trillion,
so apparently $1.2 trillion is a price for professionals only.
As what Damodaran pointed out, the dot-com correction happened as a collection of small
things. Trees falling in the forest until you wake up and realize half the forest is gone.
The pulled IPOs, the junk bond yields and the delayed filings may be the small trees.
His message to the AI companies is show me the numbers, not just the words,
and a prospectus is the one document that forces them to do that.
Wall Street, meanwhile, is getting ready to sell a lot of shares. Jim Chanos expects this
year to break the record for stock issuance, and the academic research on that isn't encouraging.
Loughran and Ritter found that companies that issue new shares tend to underperform for years
afterwards, and Baker and Wurgler found that when a lot of new stock is being issued,
the whole market tends to do worse over the following years. Insiders, it turns out,
are quite good at knowing when to sell. The investor Mike Paul has told the FT that
while the lab CEOs are saying we should slow down and take care, the market is saying that the
profit motive is overwhelming. We may look back, he said, and wonder why we didn't take them at
their word. That may be good advice about AI safety, but we should probably take the sellers
at their word about the price too, because if you buy in at $2 trillion, a price that doesn't add up
even if you assume no costs, no taxes and no staff, and it doesn't work out, someone is going
to ask you the same question Scott McNeely asked back in 2002. What were you thinking?
If you found this video interesting, you should watch my video on the IPO index inclusion scandal
next, where I looked at how the index providers changed their rules to fast track these huge
IPOs into your pension. Don't forget to check out our sponsor Plod using the link in the video
description. See you in the next video. Bye.