Serving the Kimi K3 model on Nvidia’s B200 cost $0.179 per million tokens against $0.077 on AMD’s MI355X at 100 tokens a second per user, according to SemiAnalysis’s InferenceX benchmark. That makes B200 about 2.3 times as expensive per token at that speed, by our arithmetic.
SemiAnalysis posted that AMD’s team had achieved “up to 2x better performance per dollar than even B200 on Kimi K3 Agentic Inference”. Its dashboard prices the gap at between about 1.9 and 2.3 times across the three speeds it lists.
Business Pill · THE MEMORY WALL
A short explainer on the memory wall, the idea that memory limits AI systems more than raw compute. It teaches the general idea only and says nothing about this benchmark.
The key insight: As we read the dashboard, most of the Kimi K3 cost gap comes from tokens per chip, not chip price: MI355X served about 2 times B200’s tokens per chip at 100 tokens a second per user, while B200’s chip-hour price is about 15% higher, by our arithmetic.
The comparison uses SemiAnalysis’s AgentX scenario, which replays “real coding-agent sessions rather than fixed-length prompts”, so context grows turn over turn and much of each request is served from cache. The default view is FP4 precision on Kimi K3 2.8T.
Costs use “Large-hyperscaler-volume ownership TCO normalized by total tokens”. The page lists chip pricing for an owning hyperscaler of $1.73 per chip-hour for B200 and $1.50 for MI355X, citing SemiAnalysis’s July 2026 pricing surveys and AI cloudcostmodel. The figures are interpolated from benchmark runs.
Cost per million tokens for Kimi K3 2.8T agentic inference (AgentX traces, FP4) on Nvidia B200 and AMD MI355X at three speeds per user, from SemiAnalysis InferenceX, read 31 October 2026. Redrawn by FourWeekMBA; ratios are our arithmetic.
What Was Posted on X
ALERT๐จ๐จ๐จ: AMD’s @AnushElangovan and his team have achieved up to 2x better performance per dollar than even B200 on Kimi K3 Agentic Inference! This is due to various software optimization efforts, including tuning AITER kernels and MonoKernel decode.
Along with MI355X having 2304GB of HBM per server, this allows it to avoid crossing slow 8x400Gbit/s scale-out networking between servers, which B200 needs to do due to having less than 1.5TB of HBM.
Great work to Anush, Peng Sun, Ramine, Emad, and the vLLM team for this amazing work!
SemiAnalysis posted on 31 October that AMD’s team had reached up to 2x better performance per dollar than Nvidia’s B200 on Kimi K3 agentic inference, crediting software work and the MI355X’s 2,304GB of HBM per server. The piece checks the claim against SemiAnalysis’s own InferenceX dashboard.
The Numbers at Three Speeds
At 57 tokens a second per user, B200 costs $0.089 per million tokens and MI355X $0.046. At 144 tokens a second per user, B200 costs $0.275 and MI355X $0.146. That is about 1.9 times in both cases, by our arithmetic.
On throughput, MI355X serves 5,435.8 tokens a second per chip at 100 tokens a second per user, against 2,677.5 for B200, the dashboard shows. That is about 2 times as many tokens per chip, by our arithmetic.
Where the Gap Comes From
Most of the cost gap is throughput, not chip price. The B200 chip-hour price is about 15% higher than MI355X’s, and MI355X serves about 2 times the tokens per chip. Multiplied together, those give about 2.3 times, which matches the cost gap at that speed, by our arithmetic.
SemiAnalysis credits software work, including “tuning AITER kernels and MonoKernel decode”, and memory. It wrote that MI355X has “2304GB of HBM per server”, which lets it avoid “crossing slow 8x400Gbit/s scale-out networking between servers”, while B200 has “less than 1.5TB of HBM”. That is more than 1.5 times the memory per server, by our arithmetic.
The same dashboard compares MI355X with Nvidia’s B300 on Kimi K3. At 109 tokens a second per user it lists $0.166 per million tokens for B300 and $0.093 for MI355X, with B300 priced at $2.26 per chip-hour.
The Structural Read
As we read SemiAnalysis’s post, it credits two things: software tuning of AMD’s kernels and the 2,304GB of HBM per MI355X server, which it says avoids crossing the network between servers.
As we read the dashboard, the gap is about 1.9 times at the slowest and fastest speeds it lists and about 2.3 times in the middle, by our arithmetic.
As we read the pages, the scenario replays real coding-agent sessions, where context grows turn by turn, so memory capacity matters more than in fixed-length prompt tests.
SemiAnalysis on X, 31 October 2026
“up to 2x better performance per dollar than even B200 on Kimi K3 Agentic Inference”
Three Implications
FOR BUYERS OF INFERENCE At 100 tokens a second per user, the dashboard lists $0.179 per million tokens on B200 and $0.077 on MI355X for Kimi K3.
FOR CHIP MAKERS SemiAnalysis attributes the result to kernel tuning and to 2,304GB of HBM per MI355X server, against less than 1.5TB on B200.
FOR MEMORY SUPPLIERS The post ties the result to how much HBM fits in one server, the same input whose prices drove this week’s memory pieces.
The framework says “HBM growth is outpacing accelerator demand, making memory the rate-limiting factor in the AI scaling equation.”
As we read it, SemiAnalysis’s explanation of the Kimi K3 result, more HBM per server, is one case of memory setting the cost of inference. The framework is the owner’s argument and does not discuss this benchmark.
These are SemiAnalysis’s benchmark figures for one model and one scenario, read on 31 October 2026. The dashboard says the figures are interpolated from benchmark data and reflect a default FP4 agentic-traces selection. The dollar costs rest on its own chip-price assumptions, not on prices any buyer paid. The result covers inference on Kimi K3, not training or other models.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
On SemiAnalysis’s InferenceX, Kimi K3 agentic inference cost $0.179 per million tokens on Nvidia B200 and $0.077 on AMD MI355X at 100 tokens a second per user, about 2.3 times, mostly because MI355X served about twice the tokens per chip.
A note on sourcing. Figures come from a SemiAnalysis post on X and its InferenceX benchmark pages, read on 31 October 2026. Ratios are our arithmetic. We did not contact AMD, Nvidia or SemiAnalysis.
Short explainer · Nvidia B200 vs AMD MI355X on Kimi K3
A vertical news Short summarising this story. Figures come from SemiAnalysis’s post and InferenceX benchmark; our arithmetic is labelled.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
