Is frontier AI investment a bubble?

The money flowing into frontier AI is staggering: US private AI investment reached $109B in 2024 by Stanford's AI Index — the latest fully-counted figure — global corporate AI investment topped $250B, and the leading labs carry valuations in the hundreds of billions while still spending far more on compute than they earn. Naturally, people ask whether this is a bubble.

2025 made the question sharper, not simpler. In January, DeepSeek's openly released R1 matched leading reasoning models at a fraction of the training and inference cost — and briefly erased the largest single-day market value in US history from Nvidia. That cuts both ways: cheaper, spreading capability is bullish for adoption but bearish for the moats that the leaders' valuations assume, and the year's shift toward reasoning and test-time compute moved spending toward inference, where every query carries a cost. Capability became more real and more commoditized at the same time.

So our read is less 'is the capability real' — it clearly is — and more 'do the leaders' valuations survive capability that keeps getting cheaper and more widely available, against brutal inference economics?' The honest way to watch it is to hold the investment number next to the capability and compute numbers rather than the headlines. If durable revenue catches up to the capital, today's valuations may look justified; if capability keeps commoditizing or deployment disappoints, the gap closes the painful way. We track all three so you can judge the alignment yourself, quarter by quarter.

This is our interpretation of the public data, not investment advice.