SemiAnalysis, the semiconductor research shop led by Dylan Patel, published a scathing look at Meta's AI infrastructure organization, arguing years of political dysfunction have already cost the company billions and could now derail its custom chip partnership with AMD. The report traces the pattern through Meta's 2.5 billion dollar acquisition of chip startup Rivos, where founders forced an all or nothing deal, roughly 30 percent of the acquired engineers were later laid off, and the chip project meant to use Rivos technology, Olympus, was cancelled outright. It continues through Meta's GB200 server design, called Ariel, which paired each Nvidia GPU with a full CPU instead of the industry standard two to one ratio. SemiAnalysis calculates that setup cost Meta 14 percent more per server than the standard configuration, across its entire GB200 fleet, for extra CPU and memory capacity its own AI teams barely used. Now SemiAnalysis warns Meta and AMD are about to repeat the mistake with a custom, cut down version of AMD's upcoming MI450 chip that halves its compute and memory capacity. Meta's new AI research unit reportedly has no interest in the crippled version and would rather use Nvidia's Rubin chips instead, which could hurt AMD's sales at one of its largest customers.
Why it matters · If Meta keeps designing away the very silicon it needs for frontier AI models, the billions spent on custom chips could go to waste, and AMD may lose a marquee AI customer just as it tries to challenge Nvidia.
Worth asking · SemiAnalysis says Meta's infrastructure culture, not its AI research, is the real bottleneck, and its favorite fix, cutting AMD's MI450 in half, could backfire either way. Should investors read this as a Meta execution risk, an AMD opportunity if the design gets fixed, or both?