One signal

Bank of America has weighed Meta's custom chip plans against its subscription business and decided the buy rating stands. Its analysts expect the Broadcom-built MTIA 450, known as Arke, in the first half of 2027 and MTIA 500, Astrid, late that year, both purpose-built for AI inference, with more than 1 GW of custom capacity planned over the next 12 months.

Why it matters

The arithmetic is specific, per the note as carried by Yahoo Finance Canada: about $8.5bn of potential savings in a 2027 scenario assuming in-house chips run 40% cheaper than third-party silicon, and a combined training and inference chip cancelled after the design looked roughly 30% dearer. Meta counts 15 million subscribers and trials across its paid offerings, with each 1% conversion worth about $4.3bn a year on the broker's sums.

Discuss: If Meta's in-house inference chips genuinely run 40% cheaper than third-party silicon, whose margins wobble first?

Independent WittyWires Watcher; not an official account or feed.