Kimi K3 2.8T — B300 vs H200 Performance per Dollar
Cost per million tokens of B300 (NVIDIA Blackwell) versus H200 (NVIDIA Hopper) on Kimi K3 2.8T. Owning-hyperscaler TCO normalized by output tokens — performance per dollar across LLM workloads. Pick the more cost-efficient SKU at every target interactivity level. Use the chart controls below to switch sequences, precisions, and metrics — same interactions as the main inference chart.
AgentX replays real coding-agent sessions rather than fixed-length prompts, so context grows turn over turn and most of each request is served from cache instead of being recomputed. That turns the comparison into a systems question: KV transfer between nodes, prefix-aware routing, and cache capacity all move the curve alongside raw chip throughput. Fixed-sequence workloads stay the clean baseline for kernel and silicon performance, so the two scenarios answer different questions about the same hardware. Learn more about AgentX →
B300: $0.11 per million tokens. H200: $4.85. Both at 52 tok/s/user on Kimi K3 2.8T, with B300 4493% cheaper.
Around the middle of the 2–202 tok/s/user interactivity band — at 102 tok/s/user — B300 runs $0.16 per million tokens on Kimi K3 2.8T while H200 runs $4.85. B300 is the cheaper choice by 2989%.
On Kimi K3 2.8T at 152 tok/s/user, the per-million math comes out to $0.30 for B300 and $4.85 for H200; B300 delivers 1501% more output per dollar. (Numbers reflect the default agentic-traces · fp4 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
Chip pricing (owning hyperscaler): B300 $2.26/chip/hr · H200 $1.22/chip/hr. Source: SemiAnalysis Market July 2026 Pricing Surveys & AI Cloud TCO Model.
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Dollar per Million Tokens | B300:$0.106H200:$4.850 | B300:$0.157H200:$4.850 | B300:$0.303H200:$4.850 |
| Concurrency | B300:~10H200:~1 | B300:~6H200:~1 | B300:~2H200:~1 |
Inference Performance
Agentic inference metrics from the AgentX scenario and fixed-sequence inference metrics across models, hardware configurations, and serving parameters.