Speaker
Ben Inker
Appearances over time
1 episodes
Episodes
1
Podcasts
Quotes & moments
GMO classifies today's AI market as an 'easy bubble' because investors can avoid overvalued US stocks by shifting to international risk assets without abandoning equities entirely.
The slope of GMO's risk-reward line for all assets fell from 0.4 in late 2025 to just 0.1 today, though it remains ~0.4 when US equities are excluded.
Unlike 2000's valuation bubble, today may be an earnings bubble: surging AI data center spending boosts profits before depreciation arrives, making valuations look cheaper than they are.
European corporate earnings rose 100% over 4 years ahead of the 2007-08 crisis and have never fully recovered to those levels on an index basis.
2025 US data center spending is forecast at roughly $700 billion to $1.6 trillion — approximately 2.2% of US GDP — comparable to the fiber optic buildout of the late 1990s.
Railroads transformed the world by collapsing transportation costs, but running a railroad has never generated amazing returns on investment — a lesson for AI infrastructure builders.
Historical data suggests a 1% increase in US stock market supply is associated with a 7.5% worse return over the subsequent year.
If SpaceX, OpenAI, Anthropic and other large private companies go public, they could add 5-6% of US market cap as supply over the next 12 months — more than in living memory.
Major cloud and AI hyperscalers have roughly doubled their debt ratios in the past nine months as they finance massive data center buildouts.
In the entire history of US leveraged buyouts, only one deal involved a company that could be called mega-cap at the time: RJR Nabisco.
GMO used AI to analyze over 700 leveraged buyouts going back to 1981 to understand what kinds of companies private equity actually buys.
For the average US endowment or foundation, roughly half of equity exposure comes from private equity, creating a massive hidden bet on small-cap, lower-quality companies.
GMO estimates fair value for equities at 21x normalized earnings in a low-rate environment but only 16x in a higher-rate environment — a 31% difference driven purely by the cash return assumption.
The idea that US stocks should trade at a persistent valuation premium to the rest of the world had no historical precedent as of 2010.
OpenAI's deal to buy tens of billions in AMD GPUs was structured so that AMD warrants granted to OpenAI were worth roughly half the purchase price, obscuring what was effectively a 50% discount.
Not all bubbles are equally dangerous to navigate. Today's AI bubble is concentrated in US equities, meaning investors can shift to international stocks and still hold a normal risk portfolio. In 2007, every asset was overpriced and there was no escape without going to cash.
In 1999, GMO was getting two opposite complaints from the same portfolio: too much tracking error, and why own US large-caps at all if you hate them? The benchmark-free strategy, launched in 2001, promises every holding must make sense on its own — no position held out of fear it might go up.
In 2000, the risk-reward slope was still positive — you were paid less for risk, but still paid. In 2007, the slope went negative: you were paying for the privilege of taking risk. In 2021, everything had a negative expected real return. Today's slope is 0.1 globally, but 0.4 ex-US.
In 2007, every single risk asset GMO could find was overvalued, and the equal-weighted portfolio looked exactly like the cap-weighted one. There was no diversification escape — you had to move toward the origin, meaning cash, which is career suicide for a portfolio manager.
When Microsoft spends $200 billion on data centers, that spending becomes someone else's revenue immediately — but the depreciation is spread over years. Right now, a huge chunk of that investment hasn't even started depreciating yet. This makes corporate profits look unsustainably good.
Every transformational technology from railroads to fiber optics changed the world but destroyed returns for investors. Competition floods in, overcapacity destroys ROI, and the benefits accrue to users — not builders. AI infrastructure is following the exact same script.
OpenAI bought AMD GPUs using AMD warrants worth half the purchase price — effectively a 50% discount disguised as a profitable deal. Anthropic is leasing $36 billion in Alphabet TPUs with Broadcom as a backstop guarantor. Circular finance is back, but at far greater scale than the dot-com era.
A 1% increase in stock market supply historically associates with a 7.5% worse return over the following year. SpaceX, OpenAI, and Anthropic alone could add 5-6% of US market cap as supply. The real impact hits not at IPO but 12 months later as lockups expire.
GMO's 7-year forecasts assume capitalism works: assets will eventually trade at fair value. They estimate income, growth, and the annual cost or benefit of reverting valuations by 1/7 toward fair value each year. The controversial part: 'fair value' depends on whether we're in a high-rate or low-rate regime.
Non-US equities look significantly cheaper than US stocks, the dollar is overvalued giving currency tailwinds, and value stocks range from 'quite cheap' to 'extraordinarily cheap' globally. The idea that the US should trade at a premium had no historical basis before 2010.
GMO analyzed 700+ LBOs and found they skewed massively toward small, lower-quality, highly leveraged companies. Only one mega-cap LBO has ever occurred in US history. The average endowment with 50% in private equity is unknowingly running a giant concentrated bet on small-cap junk.
If your private equity portfolio is secretly a bet on small-cap junk, the public portfolio fix is to go long S&P 100 (large, high-quality) and short Russell 2000 (small, lower-quality). Better yet: bias toward quality stocks at reasonable valuations, and actively short expensive junk.
Tail risk hedging that promises cash-like returns shouldn't exist — no rational counterparty would accept terrible correlated losses for only a cash return. If you understand why risk premiums exist, you can immediately spot when someone is selling you something that makes no structural sense.
Even if a manager is certain a bubble exists, they must run a portfolio that doesn't look insane if the bubble doesn't burst quickly. Clients will fire you and hire the aggressive growth manager at the top — maximizing their own losses. Career risk is the bubble's greatest ally.
The bull case for AI semis is that AI demand has made the business non-cyclical. History disagrees. Memory manufacturers have all the characteristics of commodity businesses. SK Hynix and Micron may look cheap on trailing PE before the cycle turns — which is exactly when they are the most dangerous to own.
Analysis
What they talk about
- Business 100%
