What If the “Scaling Cliff” Pops the AI Bubble?

“History as well as life itself is complicated — neither life nor history is an enterprise for those who seek simplicity and consistency.”

-Jared Diamond, Collapse

September 10, 2025 — Artificial intelligence is this decade’s tech success story. And that sector’s stocks — led by the almost supernaturally powerful chip maker Nvidia — are primarily responsible for the S&P 500 and Nasdaq being at record highs.

In just the past five years, nearly a trillion dollars have been thrown at AI data centers, chip plants, and model training. And the spending curve continues to steepen, as pretty much every tech firm and most governments enter the AI arms race.

Early AIs improved in line with the amount of computing power and new data they were fed. This led to the assumption that AI investment had a predictable rate of return (which investors absolutely love).

But with the most recent iterations of name-brand AI, that relationship has broken down. They’re not improving in line with the money being spent on them, leading a growing number of analysts to voice doubt about whether the return on this investment can be predicted going forward. This is known as the “scaling cliff.”

As Chat GPT explains the problem:

The entire LLM arms race assumes smooth scaling. If we’re close to a cliff:

Simply making models bigger stops being productive.
Labs must pivot to data curation, architecture changes, or reasoning-focused designs.
Many researchers suspect we’re nearing this cliff.

In short:

The AI scaling cliff is the point where bigger no longer means better — when scaling laws break because of data, optimization, or cost bottlenecks. It marks the boundary between “brute force scaling” and needing new approaches to intelligence.

Here are two video deep dives into the scaling cliff concept:

Could the AI bubble burst?

Bubbles, while they’re inflating, take on the aura of inevitability. In the 1990s, the Internet was going to rule the world, and the leading dot-coms would, as a result, grow exponentially forever. In the real estate bubble of the 2000s, home prices would always rise, so no price was too high for a nice house.

Those bubbles popped, catastrophically. That’s the nature of bubbles, and it would be a denial of history to expect the frantic money pouring into AI to return consistent profits. And to expect the broader markets elevated by this bubble to keep rising when the bubble pops.

By every historical valuation measure, US stocks (other than the commodities miners) are well into bubble territory. So it’s wise to build crash protection into today’s portfolios. Long-dated put options on the S&P or Nasdaq are just basic common-sense insurance at this point.

John Rubino
John Rubino’s Substack & Grey Swan Investment Fraternity

P.S. from Addison: We love AI. Specifically, the LLMs ChatGPT, Claude and Perplexity.

Over the weekend, we drafted a 67-page outline and publisher’s treatment of a future book using our own “AI Clone” (as our buddy Chris Daigle would call it).

At the very least, LLMs can collect, organize and describe data that took hours, days and weeks only two years ago when we updated Demise of the Dollar, Financial Reckoning Day and Empire of Debt in 2003-04 for their post-pandemic third editions. That was time spent in purposeful drudgery I would have preferred to be using to actually think.

We can see how LLMs and other advanced computational platforms will free a myriad of occupations from equal drudgery.

That said, AI doesn’t think for you.

Nor is it any more immune from market forces than routers in the Cisco bust of the 2000s tech wreck or radio transmitters in the great RCA boom and bust of the 1920s.

Grey Swan events – those which you cannot time, but can identify through current trends and historical examples – will pock the innovation cycle as much during the Age of Intelligence as any other age.

Thanks to John Rubino for sharing the growing challenge of AI scalability today.

Grey Swan Live! this week: Mark Jeftovic joins us tomorrow at 2 p.m. ET for “Shadow Fed & the American Dream” — how a September rate cut could hit the dollar’s purchasing power, where the money-market flood might go next, and why “control of money” is migrating from central banks to code, corporates, and courts.

If you’d like, you can drop your most pressing questions right here: Feedback@GreySwanFraternity.com. We’ll be sure to work them in during the conversation.

“History as well as life itself is complicated — neither life nor history is an enterprise for those who seek simplicity and consistency.”

-Jared Diamond, Collapse

As Chat GPT explains the problem:

The entire LLM arms race assumes smooth scaling. If we’re close to a cliff:

Simply making models bigger stops being productive.
Labs must pivot to data curation, architecture changes, or reasoning-focused designs.
Many researchers suspect we’re nearing this cliff.

In short:

Here are two video deep dives into the scaling cliff concept:

Could the AI bubble burst?

John Rubino
John Rubino’s Substack & Grey Swan Investment Fraternity

P.S. from Addison: We love AI. Specifically, the LLMs ChatGPT, Claude and Perplexity.

Over the weekend, we drafted a 67-page outline and publisher’s treatment of a future book using our own “AI Clone” (as our buddy Chris Daigle would call it).

We can see how LLMs and other advanced computational platforms will free a myriad of occupations from equal drudgery.

That said, AI doesn’t think for you.

Nor is it any more immune from market forces than routers in the Cisco bust of the 2000s tech wreck or radio transmitters in the great RCA boom and bust of the 1920s.

Thanks to John Rubino for sharing the growing challenge of AI scalability today.

If you’d like, you can drop your most pressing questions right here: Feedback@GreySwanFraternity.com. We’ll be sure to work them in during the conversation.

But Green is describing more than just a new calculation. He’s talking about a new form of misery.’ It’s a poverty where you may still have most of the accoutrements of middle-class life. But your relationship with the financial elite has changed: you are indentured to the credit industry — for life.

Historically, when the Fed has cut into inflation above 3%, one of two outcomes tends to follow:

A brief reprieve, followed by a larger inflation wave (see: 1970s).

A crisis born from cheap money rather than expensive money (see: housing in the 2000s).

We are heading into another round of cuts with:

• A still-bloated balance sheet

• A new digital plumbing that auto-funds the Treasury

• Hard-asset markets flashing warning lights

Paul Tudor Jones summed it up in one dry quip: interest expense is now one of Washington’s largest bills; commodities are “ridiculously under-owned”; and “all roads lead to inflation.”

The Fed’s flip from QT to easing doesn’t end this inflation episode. It likely begins its next season.

The question investors should ask themselves isn’t whether this trend is sustainable – it isn’t.

Instead, they should ask if the $2 trillion increase in Google’s market cap has sucked capital away from other promising parts of the market – and if so, where investors can expect a rally when Google reverses.

Long have we dwelt on the corrupting influence of funny money on capital asset prices and on the economy. Everything gets distorted, perverse…and false. We get high prices. We get low prices. What we don’t get are honest prices.

Yesterday, we looked at the ‘small time crooks’ — ripping off the public for a million or two.

Today, we move to the big fry.

You’ll recall that the money in question was never earned by anyone. No one has a genuine claim to it. And what kind of apple falls from this funny money tree? Just what you’d expect…a funny one…with the worms already in it.