# The Compound, 27 Oct 2026: When Money Gets Expensive, the Market Asks What You Stand On

Source: Realpha Blog (blog.getrealpha.com)
Original article and charts: https://blog.getrealpha.com/en/blog/compound-2026-10-27-dirty-dozen-charts-anthropic-s-s-1-amazon-earnings/

> Listening notes on The Compound and Friends, 27 October 2026: the 693 million dollar first-half operating loss in Anthropic's S-1 and the 79% of revenue coming from outside its top five customers, the supply overhang left by Meta's 95 billion dollar equity raise, the ratio Amazon's share price actually trades on, and how the cost of capital ate the AI basket's entire first-half gain. Educational notes only — no investment advice, no calls, no price targets, and the judgements here may be wrong.

Published: 2026-10-28
Locale: en
Tags: The Compound, podcast-notes, Anthropic, cost of capital, Amazon


![An unfinished data centre site at dusk, bare steel framing in the foreground, the view running deep down a long service corridor to idle unlit cranes at the far end](/covers/compound-2026-10-27-dirty-dozen-charts-anthropic-s-s-1-amazon-earnings-cover.png)

> Do not worry about having no position; worry about what qualifies you to stand in one. Do not worry that no one knows you; seek to be worth knowing.
>
> —— *The Analects*, "Liren" (Spring and Autumn period; my own translation)

On the 27 October 2026 live show of The Compound and Friends, Josh Brown and Michael Batnick spent close to half the episode on Anthropic's freshly filed S-1: an operating loss of 693 million dollars for the six months to June 2026, down from 2.1 billion a year earlier, 79% of revenue coming from outside the top five customers, and no single customer above 10% of revenue. Their read was that the short-term doubts got answered by the filing while the two long-term ones — whether these models have a moat, and which side wins between collapsing token costs and exploding usage — did not move an inch. In the same episode Josh described Amazon's share price as a ratio: what the market prices is whether AWS growth is outrunning capital spending, and in February of this year AWS grew 24% and beat estimates, then fell 10% the next day because management guided 2026 capex to 200 billion dollars. The answers arrive with the listing around 16 November and with Amazon's results on Thursday.

## What this episode was about

The Compound and Friends is Josh Brown and Michael Batnick's Tuesday 5pm Eastern live stream, where the two of them walk through the market stories they thought mattered that week. A two-day Fed meeting was running that day with a decision due the next morning, and neither of them mentioned it once — Josh's line was that the S-1 had sucked all the oxygen out of the room.

## Anthropic's filing: the short-term doubts got answered

Michael read the numbers first: more than 11,900 customers spent 100,000 dollars or more over the last twelve months, more than 1,900 spent a million or more, and around 250 spent ten million or more. The story going around had been that the whole business rests on a few big buyers. The filing says no single customer is above 10% of revenue, and 79% of revenue comes from outside the top five. Michael said he is no mega-bull on this company and even he has to call those numbers what they are.

![Three horizontal bars shorten sharply from top to bottom, showing customer counts falling from 11,900 to 1,900 to 250 across the $100K, $1M and $10M-plus annual spend tiers, with a note on the right that no single customer accounts for more than 10% of revenue.](/figures/anthropic-customer-tiers-en.svg)

The 80 billion dollar loss figure also got taken apart: it is an accounting treatment, nobody is handing 80 billion to anyone. Look at the operating loss instead — 693 million for the half, against 2.1 billion a year earlier. Josh said he had just spent an hour recording with Scott Galloway, walked in planning to be sceptical and came out flipped. On the funding side, the company has raised 141 billion in total, holds 112 billion in cash plus 19 billion in short-term investments, and first-half revenue went from 1 billion a year ago to 16 billion.

![On the left, two bars jump sixteen-fold higher to show first-half revenue; on the right, two bars shrink to show the operating loss narrowing, so one side rises while the other falls.](/figures/revenue-up-loss-down-en.svg)

## Penetration under a tenth of a percent, and the question nobody can answer

Gavin Baker pulled one line out of the filing: for Fortune 500 customers paying 100,000 dollars a year or more, that spend averaged under 0.1% of their total expenses in the September 2026 quarter. Michael's annotation was one sentence — that is nothing.

![One large square stands for total large-enterprise spending, and inside it an almost invisible small square shows this cost is less than a tenth of a percent of that, with a magnified view of the tiny block on the right.](/figures/penetration-near-zero-en.svg)

Josh said the argument now moves to where the growth rate settles. There are no earnings to hang a multiple on, so the price has to hang on revenue, and what multiple you give revenue depends on whether you believe growth converges to 50% a year or to something else. A jump from 1 billion to 16 billion in a half will not happen twice; the question is where the deceleration stops.

Michael laid out the bear case cleanly: the bears never said this company cannot grow sales. The bears ask two things — whether these language models have a moat that survives competitors cutting price, and whether the crash in cost per token crossed with the surge in usage nets out well for shareholders. Cheaper is good news for users; what it is for the income statement, the filing does not say.

## Owning your compute versus renting it

Josh raised a distinction he said does not matter now and might matter later: only two massively commercial large language models are offered by a company that owns its compute end to end, Alphabet's Gemini and Meta's Muse. Anthropic and OpenAI rent, with contracts locking up usage and no asset on their own books. And the filing has Anthropic saying it will have to build some of its own compute, in words that call it extremely costly.

![On the left, compute blocks sit inside the company boundary to show ownership; on the right the boundary is empty and the compute blocks sit outside it, linked by a dashed contract line to show rented access.](/figures/own-compute-vs-rent-en.svg)

That runs straight into what both hosts called the biggest open item in the document: 500 to 900 billion dollars of committed compute purchases, spread across several cloud providers. Josh pushed it one step further — if financing for the neoclouds keeps deteriorating while Anthropic sits on well over a hundred billion in cash, buying one outright starts to make sense.

## Expensive money hits the ring around AI first

Blake Millard passed along a stat from Tony Pascarella: a basket of AI leaders against laggards was up nearly 100% year to date at the end of June, and the entirety of that performance has been given back since. Semis, neoclouds, power infrastructure, memory, photonics, all down together.

![One line climbs to nearly double through the first half of the year, then slides all the way back to its starting point in the second half, while five thinner lines from the same group bend down at the same time.](/figures/ai-basket-round-trip-en.svg)

Musk said on an earnings call that over the past few weeks debt markets have almost entirely shut off access for any company without substantial revenue independent of AI data centres; they closed 40 billion of GPU debt because they are investment grade with Starlink and launch revenue behind them. His description of the 200-plus neoclouds was death at the guppy stage, with survivors coming through consolidation.

The CEO of Nebius put it more concretely in a Berlin interview: some capital sources are temporarily closed, so they are funding the build with customer advances. The same day, that company announced a 3.5 billion dollar facility with Citi — a 364-day bridge, a different animal from long-dated borrowing. Asked why the balance sheet is run so tightly, he said a real company is measured by earnings and a healthy business is a business with cash.

Meta's version is a 95 billion dollar equity raise: 25 billion of common stock priced at 625 dollars a share (the market snapped straight back to 625 that day), 12.5 billion of convertible preferred at a 6.5% dividend converting into common by 2029, and up to 60 billion through an at-the-market programme. Josh's read on that last piece is the vivid one: any time good news lifts the stock, they sprinkle shares on your head, so what you own is a stock with 60 billion of supply hanging over it. Ben Thompson's framing is that public market investors are being asked to pay for the buildout the credit market has gone nervous about, and the payment comes as a steadily shrinking share of the company. Michael added the structural piece: this group collectively bought back a trillion dollars of stock a year and has now turned into a net issuer.

Meta's year is the footnote: up 22%, down 30%, up 31%, down 20%, up 26%, down 21%, up 44% off the lows, then giving it all back, still negative year to date.

![A sharply zigzagging price line ends below where it started, with a shaded band overhead standing for up to $60 billion of at-the-market issuance pressing down on it.](/figures/meta-zigzag-supply-overhang-en.svg)

## Rates backed off and the discarded stocks all came home

Michael said the median stock's shortfall against the index, measured over the last decade, is as bad as it has been. They had been showing that chart for months waiting for the index to fall down to the median. The index did not follow; the median climbed back up instead. Duality Research's threshold for a breadth thrust is 55% of the index hitting new 20-day highs at once — this run did not reach it, but the share of stocks at new 20-day highs over the last few sessions was the highest of the year.

Josh's attribution was one word: rates. Two weeks ago everyone was talking about yields at five going to six; rates came off a little and everything came back. The homework he draws out of it is that once you have finished asking what could go wrong, keep asking what could go right. He quoted Bill Miller's answer to how much time he spends worrying about the stock market: none — risk is of course real, and the market spends all day doing nothing but worrying about risk, so he puts his attention on what could improve.

## Amazon: the market is pricing a ratio

Josh put up three years of chart: earnings per share stepping up like a staircase, the share price giving you nothing. He used five earnings reactions compiled by a writer at Seeking Alpha to show the mechanism:

- August 2025: AWS grew 17.5% while Azure grew 39% and Google Cloud 32%, and Amazon spent 31 billion on capex that quarter, more than either of them — the stock fell 7%
- Three months later: AWS accelerated to 20%, its fastest since 2022 — the stock jumped 13%
- The following February: AWS grew 24% and beat, but management guided 2026 capex to 200 billion, a third above the street — down 10% the next day
- April 28%, July 37%; the July report sent the stock up 15%

What gets punished is spending that runs ahead of growth. Going into this quarter, AWS revenue is growing 36.7% and backlog went from 200 billion a year ago to 496 billion. Josh also gave a reason for the long-standing discount: of the three hyperscalers, Amazon carries the lowest price to trailing cash flow at 16.6 times against Microsoft's 20.9 and Alphabet's 21.6, and the discount comes from this being the most physical of the three — supermarkets, the vans prowling his neighbourhood day and night, logistics facilities, millions of robots. His judgement is that owning its own chips plus cheap borrowing turns the capex itself into a moat, so the discount does not hold up.

![The top row shows the growth rate rising steadily, while the bottom row shows next-day stock reactions alternating between positive and negative, with both punished quarters marked as capital spending running ahead of growth.](/figures/amazon-growth-over-capex-en.svg)

## How the twelve charts got filtered

Josh was open about the Dirty Dozen process: the universe is the Russell 1000, the "best stocks in the market" list he and Sean Russo maintain currently holds 152 names (it has been under 100 at the lows), he first keeps only charts making higher lows and higher highs, which leaves 29, then applies their own score and takes the top 12.

![A four-stage funnel narrows from wide to thin, going from the Russell 1000 down to 152 names, then 29, and finally just 12.](/figures/dirty-dozen-funnel-en.svg)

The score looks at four things: trend quality (one-year max drawdown, the 50-day to 200-day spread), structure (distance from the 52-week high), trend velocity (RSI between 50 and 65 is best, over 70 gets penalised, because he does not want to pay for yesterday's winner), and Aroon — Sanskrit for dawn, measuring how long ago the last high was made. Michael's reaction to the Sanskrit detour was to tell him to cut it out and put the charts up.

The shape of the twelve: three refiners (the top three energy performers of the year, and they got hit that day on a rolling-over crack spread), Microsoft, two or three life sciences and healthcare names, one cybersecurity name, one logistics name. Josh drew the boundary himself: technicals are the jumping-off point, the next job is the fundamental work on why these names act so well; and one name on the list is one he does not like (listless RSI, consolidating below the old highs too long), but the score put it there so it stays. Michael's verdict on the segment was that the two of them are Mr Miyagi pruning bonsai.

## The numbers keep improving — why won't the stock move?

This was the part I felt most, because I own one of these: results that get better every quarter and a share price that has gone nowhere in three years. The Amazon segment gives the shape of the answer — in February growth beat estimates, which should have been enough, and then the capex guide went up and the stock dropped a tenth the next day. The box the market is marking is growth against spend, and I had been staring at growth alone.

One layer down, the question becomes how I find out which box my own holding sits in. What I do is go back four to eight quarters of earnings reactions and note three numbers each time — what the street expected, what the company delivered, and how the stock moved the next day — then see which of those numbers the reaction moves with. For AWS it is growth divided by capex; for a consumer company it might be gross margin; for a biotech it might be trial progress. Until I find that box, the good news I react to is good news I defined myself, and my guesses about earnings direction keep missing because I am guessing at a number while the market is pricing that number relative to another one.

## My stock is down 40% — is the company broken, or is money just expensive?

The episode hands you a clean comparison. The AI leaders basket nearly doubled in the first half and gave the whole thing back in the second; then rates eased a little and the median stock climbed back as a group. Over that stretch semis, neoclouds, power infrastructure, memory and photonics fell together. When a crowd of names moves at once, the thing moving is the cost of capital layer; companies breaking individually do not break on the same day.

So the first move is to line your holding's drawdown up against five others in the same family: everyone down by a similar amount is beta, only you down sends you back to the filings.

![In the left panel six bars fall together by similar amounts, standing for the cost of capital; in the right panel only one bar falls while the other five barely move, standing for trouble at that one company.](/figures/beta-or-alpha-six-bars-en.svg)

Expensive money does turn into real damage if it lasts, and the episode marks where that line sits. Musk's sentence draws it at whether there is revenue once you strip out the AI data centres; a company that cannot borrow stops building, sells assets cheap, or gets consolidated, and at that point beta has become alpha. So the second question is whether this company survives a year of no cheap money. The Nebius CEO's line was aimed at lenders, and lenders are listening right now.

The mistake I have made is blending the two questions. During the leg when rates were climbing I sold one holding, and looking back, my reason was that the whole group was falling while that company's cash flow for the quarter was still positive.

## Worth a look

- The Compound and Friends live show of 27 October 2026, the source for these notes
- Anthropic's S-1 filed with the SEC — customer counts, revenue concentration, cash position and purchase commitments are all in there to check against what was said on air
- Meta's offering announcements and 8-K, for the terms on the common stock, convertible preferred and at-the-market programme
- Amazon's quarterly filings and earnings decks, for AWS growth, backlog and capex guidance
- FRED's high yield and CCC corporate spreads, for tracking the cost of capital line
- Bond and financing announcements from CoreWeave and Nebius, for how neocloud borrowing terms are changing

## The one thing to take with you

**What people price is your output divided by what it cost you.**

Amazon beat on growth in February and fell a tenth; growth outran spend in July and it rose 15% — same company, same people, five months apart. Josh said on the show that sometimes you have to toil away in the dark before anyone says that person is really doing something; the second half of that sentence is that the moment they speak up is the moment they put your output next to your input.

Here is something I have tried that has nothing to do with investing: pick the thing you have put the most hours into this year — the gym, a language, time with your kid, some side project after work — and write two numbers on paper, the hours a week it takes, and one concrete thing that changed in the last three months (something you can write down, not a feeling). Look at the two numbers side by side. If the ratio looks bad, hold off on adding hours.