If there is an AI bubble, where is it? - Greg Phillips

Greg Phillips

SubscribeSign in

If there is an AI bubble, where is it?<br>What if AI succeeds too efficiently?

Greg Phillips<br>May 28, 2026

Share

You can believe AI is the most important economic technology in decades and still believe there is an AI bubble. That sounds contradictory only if you think bubbles are about artificially inflated value. They usually aren’t. Railroads, the internet, and fiber all fuel massive value in our economy today. The question usually is not whether the technology matters. The question is where the scaling bottlenecks lie, who captures the profits, and how durable those profits are.<br>The market may be broadly right about AI. In fact, markets usually are broadly right. There is already real and growing evidence that AI is making workers more productive.1 The belief that AI will substitute for some meaningful share of human labor has moved from fringe view to mainstream assumption. Compensation for human labor is a bit over half of US domestic income today.2 It is also very likely that AI will yield significant value beyond merely replacing human labor. So there’s no question that the market opportunity is enormous.<br>Thanks for reading! Subscribe for free to receive new posts and support my work.

Subscribe

Right now, the market’s most aggressive bet is not simply that AI upends the economy as we know it. It is that compute remains the dominant bottleneck long enough for today’s suppliers to capture extraordinary economics. There is already a gigantic physical buildout of compute underway, and it explains why so much equity value has accrued to GPUs, memory, networking, power, cloud infrastructure, and the surrounding capex chain.3<br>But what if AI succeeds too efficiently? The most dangerous scenario for AI infrastructure businesses may not be that models stop improving. It may be that models get good enough, cheap enough, and open enough faster than expected.4 If the declining costs of fixed-capability intelligence hold, the rest of the economy will win while the scarce-infrastructure trade loses.<br>This is the part that feels under-discussed. AI could create many trillions of dollars of productivity value and new innovation and yet still not justify today’s infrastructure valuations. The users may capture most of the surplus. Boring companies may get leaner. Small teams may do more. Services businesses may deliver more to their clients for the same cost. Software companies may support more revenue per employee. Consumers may get cheaper expertise and cheaper goods. But in the face of all this, the GPU scarcity premium could vanish. The bubble, if there is one, may not be in AI adoption. It may be in the assumption that the current bottleneck remains the long-term bottleneck.<br>So the speculative bubble question is not “AI bubble or no AI bubble?” It is more precise: is the market pricing the right sector? The long trade may be the broad U.S. economy, productivity beneficiaries, and companies that can use AI to expand margins. The short, or at least the underweight, may be the narrow supplier stack priced as if today’s scarcity lasts forever. The paradox is that the cleanest way for today’s AI infrastructure trade to break is for AI to work better than expected.<br>1A large customer-support study found that access to a generative AI assistant increased productivity by roughly 14%, with the biggest gains among less-experienced workers. We see a related labor-market pattern in work from the Stanford Digital Economy Lab which reports that early-career workers ages 22-25 in the most AI-exposed occupations experienced a 16% relative employment decline after controlling for firm-level shocks.

2Compensation of employees was 51.9% of gross domestic income in 2024 per the US Bureau of Economic Analysis.

3Goldman Sachs estimates roughly $7.6 trillion of AI-related capital investment between 2026 and 2031 across compute, data centers, and power. Goldman also recently estimated U.S. data center power demand rising from 31 GW in 2025 to 66 GW in 2027.

4Stanford’s 2025 AI Index found that the cost of querying a model with GPT-3.5-level performance fell from $20 per million tokens in late 2022 to $0.07 by October 2024, more than a 280-fold decline. Epoch AI similarly found that inference prices for fixed capability milestones have been falling at rates ranging from 9x to 900x per year, depending on the task.

Share

Discussion about this post<br>CommentsRestacks

Ready for more?

Subscribe

© 2026 gphil · Privacy ∙ Terms ∙ Collection notice<br>Start your SubstackGet the app<br>Substack is the home for great culture

This site requires JavaScript to run correctly. Please turn on JavaScript or unblock scripts