The AI Trade Is Becoming a Two-Sided Market

For most of the AI boom, investors treated faster model development as an unqualified positive. More capable models required more chips, more memory, more power and more data centers. The same spending also promised new products for software companies. One assumption supported both sides of the trade: faster was better.

That assumption is beginning to split. Anthropic Chief Executive Dario Amodei argued over the weekend that the industry should slow the pace of capability gains to allow safety measures to catch up. Leaders at other frontier labs expressed support for a slower pace, according to the Associated Press and Axios. Chip stocks sold off sharply Monday while several software and cybersecurity shares rallied.

The market’s first reaction may have been too literal. A public debate about pacing does not automatically reduce capital spending, cancel chip orders or lengthen a data-center construction schedule. Companies have competitive reasons to keep building even if the release cadence of frontier models slows. But the reaction exposed a distinction that should have been visible all along.

Compute sellers and compute users do not own the same economics

A semiconductor company benefits when customers buy more capacity. A software company benefits when it can use that capacity to improve a product without allowing inference costs, reliability problems or safety failures to consume the value created. The supplier is paid for the buildout. The user is paid only if the buildout produces something customers will buy.

A slower development cycle could therefore cut both ways. It may reduce the urgency behind some infrastructure orders. At the same time, it could give software companies longer to make existing models dependable, cheaper and easier to integrate. The economic value of an extra quarter spent improving deployment may be higher for a software buyer than the value of another benchmark gain is for a chip supplier.

Tuesday morning’s trading offered a small example of that complexity, not proof of a new regime. AMD, Intel and CrowdStrike rebounded while Sandisk, Alphabet and Nebius remained lower. The semiconductor ETF was roughly flat. Investors were distinguishing among companies instead of applying one AI multiple to all of them.

This matters more to a concentrated portfolio

AEA’s portfolio owns both sides. Sandisk, Micron, AMD and Intel sell parts of the physical stack. Meta, Alphabet, Palo Alto Networks, CrowdStrike and Datadog are closer to the application or platform layer. Nebius, CoreWeave, Vertiv and the power holdings sit between them.

Calling all of those positions “AI exposure” is descriptively true and analytically weak. Their revenues, margins and capital requirements respond to different points in the cycle. Memory pricing can improve while software multiples contract. Security companies can benefit from better automation while cloud operators struggle with financing costs. Power demand can remain strong even if the market pays a lower multiple for chip earnings.

The lesson from Monday is not that the AI trade is over. It is that the trade is becoming harder to summarize. That is healthy. A market that distinguishes between capacity, utilization and monetization is doing more work than one that rewards every company attached to the same theme.

The next test is evidence. For the suppliers: backlog, pricing, utilization and customer concentration. For the users: adoption, revenue per customer, gross margin and measurable labor savings. If those indicators diverge, share prices should diverge too.

A two-sided AI market will be less comfortable than the one-way version. It will also be more useful. It forces each company to earn its own thesis.