The $1.5 Trillion AI Buildout Has a Second Infrastructure Problem

Everyone in AI infrastructure knows about chips. This is the problem with everyone knowing about chips.

The Nvidia shortage has been analyzed in earnings calls, 13F filings, and more newsletters than anyone needed to read. Smart money saw it early. Institutions caught up. Retail followed. At this point, knowing it gives you no edge. The price of knowing it is already in every valuation you are looking at.

There is a second problem. It does not appear in chip demand forecasts because those forecasts stop at the chip. It shows up when you try to build a 100,000-GPU cluster and discover that making all those chips work together is as hard as getting the chips in the first place. Moving data between 100,000 processors at the speed the system needs, without creating bottlenecks, requires its own specialized infrastructure.

At this scale, the connections between GPUs are not a commodity. The cables that carry data between processors, the chips that keep the signal clean over distance, the switches managing traffic across thousands of paths at once are purpose-built. The hyperscalers had the solution designed in before the cluster was funded.

Astera Labs makes this layer. Their retimers are inside every major US hyperscaler and every GPU platform on the market. Their memory controllers let AI systems share memory across servers rather than being stuck inside a single machine. Their switches handle the traffic. The company said it plainly before most investors understood it: AI has outgrown the server. The rack is now the unit of compute.

FY2025 revenue: $852.5 million, up 115% year over year. EPS went from negative $0.64 to positive $1.22. The business changed shape.

Gavin Baker is a well-known semiconductor analyst who manages $7 billion. His most recent 13F shows a leveraged Nvidia position sitting next to an ALAB addition. Baker does not buy into adjacent names without a reason. The logic of that pair is not hard to follow.

The risks are real. A small number of hyperscalers make up most of the revenue, and Broadcom and Marvell both operate in similar segments with the scale to compete. Concentrated bets have concentrated risks.

$ALAB trades on NASDAQ.