Is This an AI Bubble? Depends Which Layer You’re Standing On

Ask ten people whether AI is a bubble and you’ll get ten different answers — and most of them will be arguing about completely different things without realizing it. That’s because “AI” isn’t one market. It’s a stack of layers, each with its own economics, its own risk profile, and its own relationship to hype.

If you want an honest answer to the bubble question, you have to stop asking about “AI” as a monolith and start asking it layer by layer.

Layer 1: The Model Labs

This is the layer everyone talks about — OpenAI, Anthropic, Google DeepMind, and a handful of others racing to build ever-larger, ever-more-capable models. Valuations here are enormous, funding rounds are eye-watering, and the public narrative treats every new model release like a moon landing.

Bubble signs: Sky-high valuations built on projected future revenue rather than current profitability. Fierce competition compressing margins. A real possibility that today’s frontier model is next year’s commodity.

Not-bubble signs: Genuine, measurable capability gains year over year. Real enterprise revenue, not just hype-driven signups. Products people actually pay for and rely on daily.

Verdict: This is the frothiest layer — the one most likely to see a correction if growth slows or a leading lab stumbles. But “frothy” doesn’t mean “worthless.” It means expectations are running ahead of near-term certainty.

Layer 2: Compute & Infrastructure

Underneath the labs sits an enormous buildout of data centers, chips, and power infrastructure. This is where the real money is moving — GPU manufacturers, cloud providers, and increasingly, energy companies.

Bubble signs: Massive capital expenditure on the assumption that demand for compute will keep growing indefinitely. If model efficiency improves faster than expected, some of that buildout could sit underutilized.

Not-bubble signs: Compute demand has, if anything, outpaced supply for years running. Unlike a pure software bet, this infrastructure has decades of alternative uses even if AI demand plateaus. Chips depreciate, but data centers and power infrastructure don’t disappear.

Verdict: This is the layer with the strongest case for durable, structural demand — the “picks and shovels” logic. Even in a slower-growth AI world, someone needs to own the compute.

Layer 3: Enterprise Tooling & Applications

This is the layer of startups and established companies building products on top of foundation models — coding assistants, customer service agents, automation platforms, and the like.

Bubble signs: A huge number of thin wrappers around the same handful of underlying models, competing almost entirely on UX and distribution rather than technical differentiation. Many will not survive consolidation.

Not-bubble signs: The winners in this layer are already showing real retention and real willingness to pay — because they solve concrete workflow problems, not because they’re novel.

Verdict: Expect a shakeout, not a collapse. The category is real; most of the current players in it are not.

Layer 4: Public Market Enthusiasm

Then there’s the layer furthest from the actual technology — public market sentiment, where “AI” as a label attached to a stock ticker can move a share price regardless of what the company actually does with the technology.

Bubble signs: This is where bubble dynamics are clearest — valuation multiples detached from fundamentals, momentum trading, and speculative retail enthusiasm.

Not-bubble signs: Almost none. This layer behaves the most like a classic speculative bubble, because it’s the layer furthest from underlying utility.

Verdict: If there’s a “pop,” it happens here first and loudest — and it may say very little about the health of the layers underneath it.

So — Bubble or Not?

The honest answer is: yes, and no, depending on where you’re standing.

• If you’re asking about public market AI sentiment, there are real bubble dynamics at play.
• If you’re asking about model lab valuations, there’s meaningful risk of a correction as competition compresses margins.
• If you’re asking about compute infrastructure, the case for durable demand is much stronger.
• If you’re asking about enterprise applications, expect consolidation — the useful ones will survive, the derivative ones won’t.

The mistake is treating “AI” as a single asset class that will rise or fall together. It won’t. Some layers are speculative froth waiting for a correction. Others are quietly building the infrastructure the next decade will run on — bubble or no bubble.

The real skill right now isn’t predicting whether the AI bubble pops. It’s figuring out which layer you’re actually betting on.

---

This is a general overview of market dynamics and not financial advice. As always, do your own research — and remember that different layers of the same trend can carry very different risk profiles.