AMD Says Report Of Delayed Instinct MI455X AI Accelerator Launch Is Total BS

AMD's Instinct MI400 series is its upcoming family of AI accelerators based on the company's "CDNA 5" architecture. AMD has made many lofty claims about MI400-series parts; particularly, it notes that the "Helios" rack-scale compute platform based on Instinct MI455X AI accelerators will match or beat NVIDIA's upcoming Vera Rubin platform in key areas of AI compute throughput. Those chips were expected to ship in the second half of this year, but analyst firm SemiAnalysis claims that the full mass production launch of MI455X has been delayed to 2027.


That would be significant egg on AMD's face, and the company isn't standing for it. In a to-the-point Xwitter post quote-retweeting an investor discussing the SemiAnalysis report, AMD's VP of Software Development Anush Elangovan said "BS. Right on target for H2 2026." Well, it doesn't get much more direct than that as far as corporate rebuttals go.

It's interesting to consider what "right on target" means, though. After all, SemiAnalysis reportedly says that low-volume production of MI455X will indeed happen in the second half of this year; it's only the mass production that the site claimed would slip into next year. Anush was more verbose in another, earlier post, where he wrote "MI455X is right on target for shipments in 2H 2026 - irrespective of what SemiAnalysis says".


So what does "shipments" mean, here? In semiconductors, "shipments" can mean anything from early qualification silicon to revenue-scale volume deployment. Does it mean customer ramping for scale? Does it mean engineering samples sampled to select high-value customers? Does it mean shipping the chips from AMD to a partner for final assembly? The most generous reading of Elangovan's statements is that he means AMD will be shipping complete and final MI455X silicon to customers.

We're inclined to take Anush at his word when he says "BS," but at the same time, this is the first time AMD will ever have shipped a product like Helios: a rack-scale AI supercomputer with more than a hundred separate processors inside and compute performance that would have required an entire supercomputing cluster just a few years ago. For the market's sake, let's hope AMD can execute with panache.