On 17 September SemiAnalysis posted that "AMD MI355X is quickly closing the perf/TCO gap in agentic inference, when compared to GB300 apples-to-apples", over a chart running MoRI SGLang against Dynamo SGLang on Nvidia GB300.
Its comparator carries a bold caveat: disaggregated configurations, MoRI SGLang among them, cost per decode or per prefill chip rather than per total chip count, "which makes direct token cost comparison with aggregated configs not an apples-to-apples comparison."
That page has GB300 cheaper at every matched point. Its default 8K/1K view, a different workload from the chart's agentic one, marked deprecated, prices GB300 at $0.055 per million tokens against MI355X's $0.129 at 43 tokens per second per user, and $0.098 against $0.525 at 103. $0.129 ÷ $0.055 is 2.3x, widening to 5.4x. MI355X already gets $1.50 a chip-hour against $2.31, 35% less, and $1.50 ÷ 2.3 leaves $0.64 to draw level at 43.
"Closing" is a slope, and no earlier agentic curve was pulled here, so it can be true and leave that level. Software gets the credit: AMD's SGLang work, the MoRI library, and UMBP replacing HiCache as the offload backend. No new silicon.
AMD MI355X is quickly closing the perf/TCO gap — 17 September 2026
The caveat, the prices and the $2.31 and $1.50 chip-hour inputs are SemiAnalysis's, from its InferenceX comparator on DeepSeek V4, read 18 September at its large-hyperscaler owning tier. All of it is one firm's model output, not a disclosure from either company, and moves nothing in what we carry for either. The ratios and the $0.64 are ours.
