Barry Ritholtz posted this
up-and-right chart, which is a classic of exponential extrapolation. Given the difficulty, delays, and cost of bringing the current generation of data center projects on-line, highlighted by
Oracle's "force majeure" on a 2.5GW project which is clearly not going to be complete next year, the idea that 4-5GW projects will be up and running in 2029 seems optimistic.
Below the fold I apply some skepticism, because this is starting to look like the classic
South Park episode
Gnomes:
The episode is also known for the nonsensical business plan that the gnomes of the title devise, whose three steps consist of:
- Collect underpants
- ?
- Profit
which later became a common meme used to mock poorly-thought-out business and political strategies.
Ritholtz' graph came from a report from David Crawford
et al of Bain entitled
New Innovation Is Required to Fund AI’s $6 Trillion Buildout which projects that:
By 2031, annual spending on AI infrastructure could reach $1.5 trillion, including new data center infrastructure and compute capacity as well as ongoing upgrades to the installed base of GPUs, memory, and networking equipment.
If we assume that capital expenditures amount to about 25% of industry revenue (an ambitious but reasonable percentage based on trends among cloud providers), sustaining this level of investment would require an AI market approaching $6 trillion annually.
...
Together, the consumer and enterprise AI market could total between $1.2 trillion and $1.8 trillion, leaving about $4.2 trillion of new revenue to reach the $6 trillion market that we estimate will be necessary to fund the buildout (see Figure 2).
That revenue must come from new sources of economic value.
Note that when they say "about" $4.2T they really mean "at least" $4.2T — it could be $4.8T. Their ideas on where "about" $4.2T of new revenue to pay for data centers would come from are laughable. For example, they suggest that autonomous vehicles would supply 10% of it but don't notice that the AI for the vehicles wouldn't be in a data center, it would be in the vehicle.
They
conclude:
The economics required to generate ROI from AI infrastructure are demanding trillions in new revenue, not just cost savings.
The industry needs a wave of application innovation comparable with what mobile and cloud unlocked, not just productivity gains on existing workflows.
But we're not even getting the "productivity gains on existing workflows". Gary Marcus'
It remains “very difficult [to find] actual, materially measurable financial impacts” from spending on AI, per new study reported that:
In an emailed newsletter, Laura Bratton at The Information has reported on a new study
The study was from the Wharton School:
Jeremy Korst, one of the Wharton study's authors, said it was "very difficult" to find "actual, materially measurable financial impacts" from AI spending.
Marcus
commented:
Anybody remember back in early 2023, when people imagined that ChatGPT was going to 10x every employee’s productivity and radically boost the GDP’s of entire countries?
So far, there’s zero evidence of anything like that happening at scale.
Kyle Orland reports that
AI coding agents generate more code, but not more software:
A recent study of actual coding practices across hundreds of firms finds that human code review forms a significant “bottleneck” for the overall efficiency of AI coding tools, resulting in “little evidence that firms increase software output or reduce employment” by using them. Any efficiency increased during the actual coding phase, the study authors find, is “absorbed by downstream constraints in the production process”; as “the code review process significantly increases in length, pull requests are more likely to require revisions, and reviewers leave more comments.”
...
For now, though, letting AI write your code looks like a double-edged sword, with increases in coding speed counteracted by similar increases in human code review time and effort. It’s the kind of result that makes us wonder if the considerable time and expense to get AI coding agents working is really worth it for most companies.
In
Character.ai's Bankruptcy Documents Show the Economics of AI Don't Work Jason Koebler provides evidence that not even enshittification can get the gnomes step 2 to work for AI applications companies:
The company spent $56.1 million on compute costs between January and August of 2026 alone and as of August was losing “approximately $8-9 million per month,” which it said was a massive improvement over the previous year. The company spent $19.7 million on employee salaries and benefits during that same eight-month time period and that a lot of employees quit because they “began questioning Character’s long-term viability,” according to the company.
Character said that its finances had improved to only be losing $8-9 million per month by taking “significant actions to reduce operational costs and improve revenues” by spending less on compute by simplifying its AI model, laying off staff, and injecting ads into its chats.
The original neocloud is showing that the gnomes step 2 isn't working for data centers either, as Rachel Graf and Katherine Schwartz reported in
CoreWeave’s Bonds Turn Distressed, Drawing ‘Yikes’ From Musk:
The AI cloud-computing provider’s unsecured debt due in 2032 traded below 80 cents on the dollar on Wednesday, according to Trace pricing data. That’s widely considered the threshold for when a borrower is financially strained.
At that price, the bonds yield more than 15% — a figure noted by Palihapitiya that led Musk to respond “yikes” on his social media platform.
In
CoreWeave’s bonds scream financial distress, Toby Nangle looks at the cost of insuring against Coreweave defaulting:
someone wanting to buy $10mn of protection against a CoreWeave default needs to pay $1.95mn on day one, then an additional $125k each quarter.
That is rather expensive. Oracle's debt for the "force majeure" data center is
trading at 84 cents; this isn't just about Coreweave.
Fundamentally, finding a working step 2 depends upon finding applications for LLMs that increase productivity significantly more than their cost. This would generate profits for the application vendors above the costs they have to pay the data centers, which might increase revenue more than enough to cover the power, depreciation and debt service. It looks like step 2 remains elusive.
Update: 1stNovember, 2026
In October 20
th's
Slowing Token Spend Shows Compute Demand Isn’t Unlimited, Torsten Slok summarizes the closing argument of this post in two paragraphs with two graphs. First:
The core economic flaw in the AI bull case is that the marginal cost of producing compute exceeds the marginal benefit it delivers to users, raising questions about claims of unlimited demand for data centers. How much compute and how many data centers are needed depends on the returns firms are able to generate with that compute. So far, the AI payoff for firms has been slow to materialize, implying a growing reluctance to pay for AI services. That is consistent with the ongoing slowdown in token spending among corporate customers.
And
second:
Ultimately, the number-one variable governing the viability of data center construction models is not physical capacity or power; it is the actual price users are willing to pay for the compute being built, and companies will only pay a high price for that compute if they are able to turn the compute into higher top-line growth, bigger profit margins or substantial cost savings, little of which is showing up at scale yet.
Four days later Slok followed up with evidence about the "little of which" in
No Signs of AI Payoff Outside the Tech Sector. He notes that:
Despite massive investments and widespread hype around artificial intelligence, the technology has delivered zero financial payoff to corporate America outside of tech. Revenue per worker is rising for the Mag 7 but is flat for the S&P 493,
And
that:
Profit margins are rising for the Mag 7 but are flat for the S&P 493, see the second chart below. ... With no impact on the bottom line, mainstream corporate America is pulling back and starting to view token spending not as a source of growth but as a cost center that weighs on margins.
The first-half flood of revenue that Anthropic's S1 boasted of is mostly coming
straight from VCs:
VC-backed companies are spending more and more on tokens, but mainstream corporate America does not need the latest model and is unwilling to pay for it. As long as agentic AI and token spending are confined to a small part of the economy, the macroeconomic effects of AI will be very limited,
Incidentally, the first graph in this update explains one of the two reasons why Anthropic has to go public ASAP, before investors notice that the revenue growth has collapsed. The other is that they have to stop renting compute from their competitors and start massive capex spending to build their own data centers, switching from an asset-light to an asset-heavy business model.
No comments:
Post a Comment