A Growing Risk to the AI Trade

[3 minute read]

AI is a transformational technology and the productivity gains are undeniable.

But a transformational technology is not necessarily a transformational business model.

The economics remain unclear.

Nowhere is that more visible than with the hyperscalers, the companies that operate the massive data centers powering AI: Microsoft, Amazon, Google, Meta, and Oracle.

For individual users, AI labs like OpenAI and Anthropic sell high usage plans for $100 to $200 per month.

Power users on those plans routinely consume thousands of dollars per month in compute.

Every heavy user is a loss leader.

That was supposed to be fine.

Individual plans were priced to gain market share.

Corporates paying per use were the business model.

Then the corporate bills arrived.

Corporates want the productivity gains, but they don’t have the budget for frontier pricing.

So users are going somewhere else.

One year ago, U.S. frontier models processed roughly 70% of tokens on OpenRouter, a marketplace for AI models.

Today that share is about 30%.

Chinese open source models like DeepSeek took it.

These models run on-premise, meaning companies host them on their own servers.

They typically trail frontier model performance by a few months, but for most business tasks they are good enough.

And the cost is an order of magnitude lower, limited to the compute itself.

The hyperscalers are spending like the frontier business model is proven.

Stock buybacks have been almost entirely forgone to fund AI capex.

Free cash flow has collapsed.

Wall Street says the payoff is coming.

Consensus forecasts show hyperscaler free cash flow tripling by 2030.

Those forecasts assume corporates adopt closed source frontier models at full price.

The data above suggests they are doing the opposite.

The Mag 7 are up just 2% this year.

The other 493 stocks in the S&P 500 are up 14%.

The companies spending the most on AI are the worst performers in the index.

That is not what a proven business model looks like.

If the shift to open source accelerates, hyperscaler and Mag 7 underperformance is just getting started.