Social media giant Meta has suffered "catastrophic" missteps in its AI infrastructure strategy, according to a new report from semiconductor research firm SemiAnalysis. Meta is asking AMD to build a custom MI450X chip with drastically reduced specifications, halving the compute units and slashing high-bandwidth memory (HBM) capacity by nearly two-thirds. The custom version will see its compute silicon area cut in half, with the HBM stack reduced from 12-Hi to 8-Hi. Meta's infrastructure team justified the configuration for recommendation system workloads, but the decision was finalized before TBD Lab, Meta's core research hub for large language models, was established. TBD Lab has zero interest in this downgraded chip, and SemiAnalysis stated, "TBD will heavily favor Rubin." SemiAnalysis called on AMD to bypass Meta's infrastructure team and work directly with TBD Lab, blasting that "Meta's infrastructure team needs a cultural reset." The report also cited Meta's acquisition of Rivos, completed in 2024 for over $2.5 billion, as another example of infrastructure decision-making failure, noting that few within Meta's internal chip division understood the strategic logic of the acquisition.