feat(blog): B200 NVFP4 vs H200 INT4 on Kimi K2.5/K2.6 — up to 2.95x better perf/$

[functionstackx]

Summary

• New blog post: B200 NVFP4 vs H200 INT4 on Kimi K2.5/K2.6 — up to 2.95x Better Performance per Dollar (/blog/b200-nvfp4-vs-h200-int4-kimi-k2-vllm-perf-per-dollar). On the 8K/1K workload with vllm/vllm-openai:v0.21.0, B200 NVFP4 is 2.71x–2.95x cheaper per million tokens than H200 INT4 across the entire 30–90 tok/s/user serving band, peaking at 2.95x at 32 tok/s/user ($0.140/M vs $0.413/M — a 66% reduction). On the same B200 silicon, NVFP4 vs INT4 is worth another 2.50x–2.74x at iso-interactivity.
• Three-factor silicon-to-perf decomposition anchors the cost gap to specs the reader can audit on /gpu-specs:
  1. 1.67x HBM BW (8 vs 4.8 TB/s) — the decode-bound throughput floor
  2. 1.28x HBM capacity (180 vs 141 GB) — fits K2 in TP=4 on B200 vs TP=8 on H200, halving collective traffic per decode step (Amdahl's law on the serial-collective bottleneck)
  3. NVFP4 precision unlock (9,000 TFLOP/s FP4 cores on B200; Hopper SM90 has zero FP4 tensor cores)
  4. ÷ 1.38x B200 TCO penalty ($1.95 vs $1.41/GPU/hr) = measured 2.95x cost gap
• Kimi K2.5/K2.6 framing anchored on production deployment: open-weights backbone behind xAI's Cursor Composer 2 + Composer 2.5 (1M+ daily active users from the Cursor IDE), leads SWE-Bench Pro at 58.6% over GPT-5.4 (57.7) / Opus 4.6 (53.4) / Gemini 3.1 Pro (54.2), 80.2% on SWE-Bench Verified, 3.3% failure rate on Cline's diff-editing production data (matches Claude 4 Sonnet). K2.5 + K2.6 share the same pre-trained backbone (post-training refinements only) so every serving result applies one-to-one.
• Architecture diagram from Moonshot's model card: 1.0T total / 32B active, 1 dense + 60 MoE blocks, MLA attention, top-8-of-385 expert routing, 256K context (262,144 tokens), YaRN RoPE, vocab 163,840.
• 9 files added: MDX + 6 image variants (benchmark chart, MI355X-vs-B200 specs radar, Kimi K2 architecture diagram, all in light/dark pairs).
• Iso-iv table built with the bundled iso_interactivity.py helper. Three per-config tables (H200 INT4 TP=8, B200 INT4 TP=8, B200 NVFP4 TP=4 + TP=8). FAQ JSON-LD covers the five questions readers actually ask: cost ratio, silicon-vs-precision decomposition, NVFP4-vs-INT4 on same silicon, why K2.5/K2.6 matters, what's not covered.

SKILL.md update

Also bundling a small SKILL.md house-style addition that came out of writing this post: a ban on the "X, not Y" antithesis construction ("the gap is silicon × precision, not framework", "this is a real lever, not a paper one", "every gain came from the kernels, not the silicon", etc.). Reads as performatively contrarian AI flexing and was getting reflexively cut in editorial review — codifying so future drafts skip it. Three before/after examples included.

Test plan

• pnpm dev and visit /blog/b200-nvfp4-vs-h200-int4-kimi-k2-vllm-perf-per-dollar — verify all 3 figures render in light + dark modes
• Post appears in /blog listing with correct title, subtitle, publish date (2026-05-26)
• OG image renders correctly
• DashboardCTA at top + bottom + live-chart link all land on the preset 3-way comparison view on inferencex.semianalysis.com
• Sitemap / RSS feed / llms.txt include the new post
• All HF model card + tomtunguz + cline + NVIDIA datasheet + SemiAnalysis TCO + vLLM repo links resolve

🤖 Generated with Claude Code

---

Note

Low Risk
Documentation and editorial guidance only; no application runtime or security-sensitive code paths change in the diff.

Overview
Adds a new InferenceX benchmark post at /blog/b200-nvfp4-vs-h200-int4-kimi-k2-vllm-perf-per-dollar comparing vLLM on Kimi K2.5/K2.6 at 8K/1K across B200 NVFP4, B200 INT4, and H200 INT4 (InferenceX run 2026-05-19). The writeup anchors 2.71x–2.95x lower $/M tokens for B200 NVFP4 vs H200 INT4 in the 30–90 tok/s/user band, plus 2.45x–2.74x NVFP4 vs INT4 on the same B200 hardware, with per-concurrency tables, an iso-interactivity cost table, /gpu-specs radar and architecture figures, preset DashboardCTA links, and FAQ JSON-LD.

Updates write-inferencex-blog SKILL.md with an editorial rule to avoid the “X, not Y” antithesis pattern, with before/after examples for future drafts.

^{Reviewed by Cursor Bugbot for commit 648e4e0. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
inferencemax-app	Ready	Preview, Comment	May 26, 2026 8:15am

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit cc2330c. Configure here.}