# Odd Lots: Sixty Percent Said AI Financing Would Be a Problem. He Thinks the Answer Should Be a Hundred

Source: Realpha Blog (blog.getrealpha.com)
Original article and charts: https://blog.getrealpha.com/en/blog/oddlots-2026-10-29-what-wall-street-really-thinks-about-the-ai-trade-/

> Notes on the 29 October 2026 episode of Odd Lots, where hosts Tracy Alloway and Joe Weisenthal talk with Alexander Altmann, Global Head of Equities Tactical Strategies at Barclays, about the state of the AI trade: a Barclays investor survey in which about 60% said AI financing would be a problem, which Altmann argues should by definition be 100%; Microsoft's latest Azure growth of 48% against a 44.8% estimate, and Alphabet down 7.9% the same day while the index barely moved. Educational listening notes only — no investment advice, no price targets, and the judgements here may be wrong.

Published: 2026-10-29
Locale: en
Tags: AI financing, podcast-notes, data centres, volatility, market sentiment


![A data centre under construction at night, a single thin cable running from a small field transformer across wet concrete under work lamps in the foreground, unfinished server halls receding row after row behind it, the string lights thinning into haze](/covers/oddlots-2026-10-29-what-wall-street-really-thinks-about-the-ai-trade--cover.png)

> Annual income twenty pounds, annual expenditure nineteen nineteen six, result happiness. Annual income twenty pounds, annual expenditure twenty pounds ought and six, result misery.
>
> —— Charles Dickens, "David Copperfield" (1850)

In the 29 October 2026 episode of Odd Lots, Tracy Alloway and Joe Weisenthal ask Alexander Altmann, Global Head of Equities Tactical Strategies at Barclays, what Wall Street actually thinks about AI right now. He says Barclays has just run an investor survey; one question asked whether AI financing will be a problem for the foreseeable future, and of the responses so far, about 60% said yes — which, given what has happened over the past few weeks, he argues should by definition be 100%. This piece is about that gap. One boundary first: the financing constraint he describes is landing on the outer ring of companies with no operating cash flow, while the core hyperscalers show no sign of being constrained — Microsoft's latest Azure growth came in at 48% against a 44.8% estimate.

## Sixty, and the hundred he thinks it should be

Altmann starts with the consensus: this train is in motion, and nobody is getting off. Sample investors in the US, Europe and Asia, he says, put a gun to their head and ask whether they want to be bullish or bearish AI, and the answer comes back overwhelmingly bullish. He reaches for that bubble chart — the first dip down after the line rolls over is the denial one — and while he won't say the bubble has popped, laying the two narratives side by side puts us in the denial stretch.

Sixty percent saying financing will be a problem means close to half think there is no financing issue. His line: by definition, what has happened over the past couple of weeks means the answer should be a hundred.

![Two tracks of equal length sit one above the other: the top one is filled only to about six-tenths and leaves a long blank stretch on the right, while the bottom one is filled end to end, and that blank stretch is the gap between what the survey answered and what he thinks the answer should have been.](/figures/oddlots-60-vs-100-en.svg)

On how cheap consensus has become, he drops a nice piece of sarcasm. Someone told him Korean memory stocks are on three times earnings, so they're cheap. His reply: you can look up three times earnings on a Bloomberg terminal, and that requires no work at all. The three is the market pricing the cycle — semiconductors are famously cyclical, and the old mantra is that you don't pay up for cyclical earnings at the peak.

## The tip of the spear

To see where financing is stuck, he splits the ecosystem into core and periphery.

The periphery is what he calls the neoclouds: companies whose business is to build compute and sell it to predetermined customers, some of them genuinely predetermined, some of them speculative. They have no operating cash flow today, or so little it rounds to nothing, which puts them at the riskiest end of the AI financing spectrum. The core is the hyperscalers, the large cloud companies building their own data centres, whose legacy businesses are cash volcanoes.

![A thick ring wraps a solid inner circle: the outer ring is the neocloud builders with no operating cash flow, the inner circle is the hyperscalers with cash volcanoes, and the funding strain lands on the outermost edge.](/figures/oddlots-core-vs-periphery-en.svg)

The transmission is worth writing down. A neocloud called Cipher — Joe said on air he had never heard of it — tried to raise money with a convertible and nuked its own share price doing so, which looks like a company that can't tap the debt markets directly. That week CoreWeave said nothing at all, and its bonds sold off anyway, because a handful of quick investors put two and two together.

Then comes the detail that matters most: look at the data centre bonds in isolation, like the one tied to Meta, and they had already been widening for months, with a creep showing up in investment grade hyperscaler paper too. That week's selling only pushed something already months old into plain view. He reaches for a parallel and says up front he isn't trying to be sensationalist: the global financial crisis didn't start with the big banks, it started with subprime vehicles twelve to twenty-four months earlier.

![Two credit-spread lines have been climbing for months, and only inside a narrow band on the right marked "that week" do the sell-off event dots appear, showing that price moved early while attention arrived late.](/figures/oddlots-spread-widening-en.svg)

## A relationship that broke

His team did a piece of work in the summer that had nothing to do with AI. They took net issuance across the whole market — investment grade, high yield, treasuries, equity — back to about 1995, expressed each year's net issuance as a percentage of that market's market cap in that year, and plotted it over time. The result was what you'd expect: rates down, issuance up; rates up, issuance down.

Then 2026 shattered it. On that chart, the relationship between the two-year yield and net issuance falls apart: at this level of rates the market is selling more debt than anything in thirty years of their data. He contrasts it with dot-com, when equity issuance was heavy and rates were high — but the US government was running a surplus.

![On the scatter plot three decades of dots line up in a band sloping from upper left to lower right, high rates meaning low issuance, yet the 2026 dot sits in the upper right far away from that band.](/figures/oddlots-broken-relationship-en.svg)

So his real point is competition for capital. AI is not issuing debt in a vacuum; other verticals need financing too, and the pool is finite. Barclays forecasts net investment grade plus high yield issuance of $1.1 trillion this year. We throw trillions around as though everybody has one, he says, and the limit is real. That closes the loop with the opening: the market used to hold a consensus that AI was immune to high interest rates, and he thinks that is plainly untrue.

## The revenue side doesn't close the gap

Their note from late August, "Placing your AI chips more carefully", comes at the same question from the other end: where does the money come from.

The backdrop then was the roughly $30 trillion total addressable market floated for SpaceX, about the annual GDP of the United States. He cites a colleague's number: since 1990, the cumulative revenue of the two most successful tech companies in modern history, Microsoft and Apple, adds up to less than $10 trillion inflation-adjusted. Someone is talking about a market three times that. His own arithmetic: S&P top line of about $17 trillion, expenses of roughly $11 to $12 trillion, net profit somewhere between $5 and $5.5 trillion. That's the size of the pool.

![Five horizontal bars of different lengths sit side by side: the longest is the $30 trillion addressable market, then S&P-wide revenue, costs and net income each get shorter, and at the bottom is Microsoft plus Apple's cumulative revenue over thirty-five years.](/figures/oddlots-pool-size-en.svg)

Meanwhile, on the spending data people watch, token expenditure has begun to level off, while open-weight share has gone from about 20% to as much as 80% and token cost has fallen to a hundredth — possibly, Joe says, one of the fastest-deflating assets in history. And GPU rental rates are going up. Both hold at once, which is why he thinks this story produces winners and losers.

![A scissor-like split: the token-cost line falls all the way to the floor and the spending line goes flat, while the GPU rental line tilts up.](/figures/oddlots-price-vs-rent-en.svg)

His read on revenue therefore lands on the consumer layer: enterprise revenue won't plug the gap, and what would make him constructive again is somebody finding the killer consumer application. There's an aside in the episode that lands exactly there. Joe's wife was travelling, he was solo parenting, and the schedule was scattered across a pile of emails — his daughter's play rehearsals run one to six on some Sundays, but somewhere in one email it says certain cast members can leave at three. So he asked ChatGPT what time he was supposed to pick her up. Altmann's immediate response: what's that worth to you a month? It matters, he says, because it determines the shape of the revenue. His own conclusion is that the best technological changes humans get often come from changing the application of existing tech, and that a 100-IQ agentic workflow will be close to ubiquitous within twelve to eighteen months.

## Good news comes out and the stock falls

Mid-interview, Alphabet and Microsoft reported: Alphabet down 7.9%, Microsoft up 6.4%, two giants moving violently in opposite directions, and for anyone holding the index it washes out.

![A long bar reaches out on each side of a center line, one down and one up and nearly the same length, while the line beside them standing for the index is flat.](/figures/oddlots-index-cancels-out-en.svg)

Altmann's explanation has two layers. The structural one is how money is deployed: multi-manager AUM was around $300 billion in 2020 and is over a trillion today, with leverage on top. AUM more than tripled while the number of tradeable securities stayed flat or shrank, which creates crowding. The way funds handle crowding is to neutralise style factors — quality, value, size, growth — and they spend an inordinate amount of time and computing power doing it, except everyone is doing the same thing, so what gets manufactured is a different shape of crowding. The volatility goes into factor rotation instead of showing up at the index level, which is why realised vol and the VIX look benign while single stocks are brawling. He describes life for funds running long-short books as a knife fight in a phone booth.

![Two containers of the same size stand side by side: the right one is filled past the brim and its mouth is a little narrower, while the left one holds only a small layer at the bottom.](/figures/oddlots-crowding-en.svg)

The technical layer is harder. Investors keep asking him whether he has managed to map the AI factor, and the answer is no, because the AI factor and its betas keep changing. Early in the year plenty of people wanted to be short software, and that was part of the AI factor; now software is trading near twelve-month highs and whether shorting it is still an AI expression is unclear. The narrative shifts, so the factor shifts, and hedging it gets harder. The bottom line is that you are either long AI or short AI, and for the past three years short AI hasn't been a lucrative trade. If you can't say which side your own holdings sit on, that's worth sitting with — the people who do this for a living say they can't build the factor either.

## What would make this reading wrong

He grants the bull case in full: build it and they will come. Somebody will pay for these GPUs, demand is rising, and if you have power and cards they're worth a lot of money somewhere or nothing. Rental rates are going up, which is evidence on that side.

The harder counter is cash flow in the core. Microsoft's Azure grew 48% against the 44.8% estimate Bloomberg had collated; Joe read Alphabet's cloud revenue of $30 billion off his screen and said he couldn't believe the growth number. Altmann's framing: if you have a cash volcano from a legacy operating business, then you get to run the largest capex experiment in American history.

The third counter is how the market reads self-restraint. Oracle said at its investor day it would limit additional fundraising to about $20 billion through May, and the stock rose 12% that day — read as a good thing, as long as people raise a reasonable amount the build-out can continue.

So for the financing constraint to reach the core, Altmann says you'd want to see capex reined in by one of the hyperscalers. Asked by a client whether that might show up in this week's results, he said it's probably premature, and that tempering the enthusiasm might have to wait for the new year. He adds a geography: the closer you sit to the epicentre of the technology, the more both companies and investors are bought into the narrative, and there's an echo chamber inside that ecosystem. It was Alphabet, in the prior set of results, saying capex would be significantly higher that got everyone enthused all over again.

## A few other things worth remembering

- He tried to find a historical comparison for this capex and couldn't. On his rough arithmetic — and he flags "if you believe the forward numbers" — the spend is about 12% of GDP, which he likens to two Netherlands. The Louisiana Purchase and the railroads are left far behind; he gets to the pyramids and the Great Wall, which took centuries.
- On S&P concentration he says the index has been Kospi-fied: roughly two-thirds of next year's S&P earnings growth comes from semiconductors, and about 75% once you count the broader AI ecosystem.
- One mutual exclusion: what's good for hyperscalers is bad for memory. The index sits still while violent rotation goes on underneath.
- Joe picked up a phrase in San Francisco, the frontier sink — which industries genuinely need PhD-level intelligence and beyond. Biotech, semiconductor design itself and maybe quant trading come easily; after that it gets hard, and his takeaway is that for a lot of industries there may be a level of intelligence that is simply enough. Altmann's iPhone analogy fits: in 2007 and 2008 people thought peak iPhone was the app that let you pretend to drink a pint of beer.
- Joe relays a Stanford computer science student's view that Wall Street hasn't taken the AI pill; in that student's eyes, if you're even discussing an overabundance of GPUs, you don't get it.

## Worth a look

- The 29 October 2026 episode of Odd Lots, the source for this piece (recorded, as mentioned on air, on 28 October)
- The Barclays note from late August, "Placing your AI chips more carefully", which Altmann refers to in the episode
- Barclays' work on net issuance versus the two-year yield, with data going back to 1995
- Anthropic's S-1 filing, searchable on the SEC's EDGAR system
- Oracle's October 2026 investor day materials
- Microsoft's and Alphabet's latest quarterly results, released after the close on the day of recording

## The one thing to take with you

**When something starts to go wrong, the first thing you hear from is the thinnest line holding it up.**

The concrete version of that is what I'll keep from this episode: in the week it broke, CoreWeave said nothing and the thing that got hit was a company nobody had heard of. The signal comes from the edge because the edge has no buffer — the core has a cash volcano, so the problem gets absorbed and never shows. Which is also how you get 60%: everyone is watching the most visible names, and those names are printing great numbers.

![A chain runs from thick to thin: the leftmost link is heavy and wide, each link to the right gets thinner, and the rightmost one is as fine as a hair and singled out with an arrow pointing at it.](/figures/oddlots-thinnest-link-en.svg)

Here's something I've tried that has nothing to do with investing. Pick the thing you care most about — the school run, the weekly call with your parents, the job you have now, a relationship — and trace downwards through what holds it up until you reach the thinnest strand: a person willing to cover for you, three months of cash, some process only you know how to run. Then write two sentences: what happens the day it snaps, and when I last checked on it.