Comment by vannevar | Hacker News Reader

Comment by vannevar | Hacker News Reader

Parent Full thread

>Anthropic's S-1 filing says that they make $2.08 for every $1 of compute. In other words, once you ignore the cost of building infrastructure, AI is very profitable.

If you ignore the cost of building cars, they are also very profitable.

>Profits Can Grow, By A Lot

They can also decline by a lot, and Anthropic's S-1 suggest that is what is happening, as their losses are widening, not narrowing.

>Google Has No Moat, But Neither Does McDonald's

That's why no one is sinking several trillion dollars into a new fast food chain.

>Alphabet, Meta, and NVIDIA are all AI companies that have seen their market capitalization explode over the past few years, but all of them have stable P/E ratios in their 20s.

The "E" in their PE ratios is circular revenue that is ultimately derived from debt and equity, not real end-user demand from profitable companies. For some of these bubble companies, if you subtract the debt-fueled private stimulus money, their actual PE ratios skyrocket.

> $2.08 for every $1 of compute.

sounds pretty good. if I made 2.08 for every dollar of my spend I would make more money. Plenty of businesses operate on much much thinner margins, and thats not utility-like at all.

I could only find these numbers in articles on something called crypto briefing dot com, not sure how the author got this number from the S-1.

When using traditional business metrics, such as operational net income, we'll see that Anthropic has been in the red for the entirety of its history, at times to the tune of $3B per quarter, except in Q2 2026.

When looking at amended S-1 (with Q3 numbers), page 23, "Consolidated Statements of Operational Data", we can see that for 2026 YTD (Q1 - Q3) net loss was $103B; operating loss was $1,260B.

Anthropic also disclosed up to $895B (multiyear) commitments for infrastructure build-up.

Basically, the company's financial results do not allow for any reasonable valuation (which would be more than 0), leave alone $2T or even $1T. If this is not an indication of a bubble I don't know what is.

If your hyper growth startup is not in the red for its entire history, it fucked up. One can argue their "profitable" Q2 was a mistake, startups at this stage should be far from profitable

There's no dynamic in their financial results which points to profitability with any probability at any point in the future. Yes, the revenue grows. The expenses grow even faster, AND they plan to move to building their own infrastructure instead of renting in the future (more risk as they need to be able to execute that).

While this is true, most of Anthropic's expenses involve locking up compute for the future. But they have no idea how much compute they will actually need in the future, or how much they're going to be able to charge their customers for it. Per token prices are dropping like a rock, and performance optimizations are making model serving more efficient. Combine that with the comparatively slower growth in organic AI demand (compared to debt-driven/investor subsidized demand), and you have the very real possibility of a 2000-style "dark fiber" problem.

That’s 52% gross margin. Not much higher than the overall S&P 500. If you break it down by sector, Real Estate, Health are, Financials (and of course Tech overall) have higher gross margins at the sector level.

That $1 of compute is not taking into account R&D spending, infrastructure, etc.

And that's their only revenue stream

>That $1 of compute is not taking into account R&D spending, infrastructure, etc

People assume this, but it is wrong. Check the S-1 for Q2, the $2.08 in revenue is for every $1.00 spent on "compute and infrastructure expenses" which also include R&D compute. (Chart is on page 137)

"Compute and infrastructure expenses primarily consist of costs attributable to operating our technology infrastructure, which supports the full range of workloads for our models, including research, model development, serving customers, and internal usage of Claude." page 146 in the S-1/A-1

It is not really clear. In the graph on page 137 does the "dollar of compute" includes only compute required for inference to provide the answer? Or does this include training as well?

It would be logical to assume that they show expense of compute required to run inference. Leaving training compute (which is huge expense) out.

It's damning that out of the only two companies in the world that absolutely must dogfood Anthropic's products, one of them, the investment bank behind Anthropic's IPO (Goldman Sachs), didn't find enough of a use case for to justify its continued use by most of the bank. Now LLM use is heavily restricted to specific departments and even in those departments spend must be approved. They've also repeatedly dropped the IPO target by almost half in just 4 weeks. That's WeWork levels of adjustment right there.

It's almost as if they know something about the finances that the rest of us don't...

You assume these companies telegraph their motives, they could just as easily be feeding this narrative so they can buy more of the company at a lower price.

This should be a top level comment

Or it's like no one wanted to believe Ed Zitron, who has been telling everyone everything they need to know about this real-life, actual AI bubble....

Along with the previous 34 AI bubbles.

I mean it's all the same bubble

Being right at the wrong time is often worse than being wrong in the general case. Predicting bubble implosions and bear markets are the classic examples.

Reply on news.ycombinator.com