SemiAnalysis Rips Meta Infra Over $2.5B Rivos Acquisition

Meta's infrastructure organization spent more than $2.5 billion last year buying Rivos, and according to a detailed critique from SemiAnalysis, it may already be regretting the check. The newsletter reports that Meta wanted only Rivos' accelerator and GPU team, but the startup's founders insisted on an all-or-nothing deal, so Meta bought the entire company and then heavily cut employees in the parts it didn't want. Rivos co-founder Mark Hayter has already left. The acquisition was led by Meta's silicon chief, Yee Jiun Song, who pushed for the deal against the wishes of some of those under him but is described as having since lost interest.

Rivos is presented as the loudest example of a deeper pattern. SemiAnalysis argues that Meta's stack-ranked review cycles, which cut the bottom 10% to 15% every review round, push engineers toward what it calls 'window washing' — highly visible projects that can be delivered quickly rather than the slower work of long-term infrastructure strategy. The knock-on effect the piece describes is that suppliers have begun prioritizing designs from Amazon and Google over Meta's, treating the frequent reshuffling as a signal to invest their attention elsewhere.

The server designs get the most technical scrutiny. Meta's entire GB200 fleet ended up on a custom SKU called Ariel, pairing one Nvidia B200 GPU with one Grace CPU instead of the standard two GPUs per Grace configuration everyone else purchased. SemiAnalysis pegs the total cost of ownership for the Ariel NVL36x2 server at 14% higher than the standard GB200, a premium the piece attributes to prioritizing recommendation-system CPU ratios that Meta's LLM teams then had to live with as well.

The honest caveat is that this is a single outlet's account of an internal culture, sourced through its own channels rather than confirmed on the record. Meta has not responded publicly at the level of detail SemiAnalysis provides, and the piece does not disclose how many of its sources sit inside the current infrastructure organization versus former staff or vendor contacts. Take the specifics as reported, not settled.

The reason to watch it anyway is that if vendors really are steering their best allocations toward hyperscalers with steadier roadmaps, the cost of infrastructure churn is not just wasted headcount. It is compute capacity Meta does not get to buy in a year where every hyperscaler is fighting for the same silicon.