SemiAnalysis: AMD chips can earn up to $3/hr per chip serving Kimi K3
Editorial responsibility
- No named human review is recorded for this page.
- Source reporting is collected, normalized, translated or condensed automatically when needed.
- Automatically published source-backed update
SemiAnalysis posted a profit alert claiming that AMD chips running the vLLM project can generate up to $3 per hour per chip in pure profit when serving the open-source Kimi K3 model, even after accounting for the rapid rise in chip rental prices following Claude Code. The post argues Kimi K3 2.8T, with over 2.8 trillion parameters, suits AMD because it can use AMD's 288GB per chip. By contrast, B200 cannot fit Kimi K3 in a single server, forcing tensor or pipeline parallelism over 400Gbit/s InfiniBand/Ethernet, which the post says is 18x slower than bandwidth within an NVLink domain. The claim is an attributed opinion from SemiAnalysis rather than a verified financial result.
Source report
AMD chips running the open-source Kimi K3 model via vLLM can generate up to $3 per hour per chip in pure profit — even after accounting for the rapid increase in chip rental prices following the release of Claude Code.
Why Kimi K3 Favors AMD
Kimi K3 2.8T is a strong fit for AMD hardware:
- 2.8+ trillion parameters can take advantage of AMD's 288GB of memory per chip.
- B200 cannot fit Kimi K3 in a single server, forcing it to rely on tensor parallelism or pipeline parallelism for communication across 400Gbit/s InfiniBand/Ethernet.
- That interconnect is 18x slower than the bandwidth available within an NVLink domain.
Source
SemiAnalysis_Neutral / independent