feat(blog): GB300 vs GB200 NVL72 on DSv4-Pro — up to 2.83x throughput/GPU

[functionstackx]

Summary

• New blog post: GB300 NVL72 vs GB200 NVL72 on DeepSeek-V4-Pro 1.6T, Dynamo+vLLM FP4 8K/1K, disaggregated on both racks.
• Headline: 2.83x throughput per GPU peak at 27 tok/s/user; 2.31x perf/$ at the same operating point after GB300's 20% TCO premium ($2.65 vs $2.21/GPU/hr).
• Framing: identical NVLink fabric, identical 8 TB/s HBM BW, identical 72-GPU world size — the only meaningful silicon delta is GB300's 1.5x HBM capacity (288 vs 192 GB/GPU) and 1.5x FP4. That HBM headroom is what unlocks a wider prefill+decode recipe (conc=3072, 28-GPU prefill, 32-GPU decode EP=16, 6,812 tok/s/GPU at 25.9 tok/s/user) that GB200 has no equivalent for in the 22–32 tok/s/user band.
• Honest gap-inversion: at peak throughput (13–18 tok/s/user) the perf/$ ratio collapses to ~1.0x because both racks run narrow EP=8 recipes; the headline lift is structural to the middle of the curve, not the entire frontier.
• Numbers verified against latest_benchmarks via InferenceX MCP (every row in the per-conc tables maps 1:1 to a CSV row from the 2026-05-22 run, GHA 26306422380).
• Iso-interactivity table interpolated using the bundled .claude/skills/write-inferencex-blog/iso_interactivity.py helper so the numbers match the live dashboard chart.
• Model architecture cited from the HF model card config.json: 1.6T total / 49B active, 384 routed experts + 1 shared, 6 active per token, 61 layers.

Test plan

• Visit /blog/gb300-nvl72-vs-gb200-nvl72-dsv4-pro-vllm-fp4 on the Vercel preview and verify both benchmark-{light,dark}.png and specs-radar-{light,dark}.png render
• Verify the DashboardCTA links land on the right pre-filtered comparison
• Verify the post shows up in /blog, /feed.xml, /llms.txt, and /sitemap.xml
• OG image generates for the post URL

🤖 Generated with Claude Code

---

Note

Low Risk
Editorial MDX and a documentation link change only; no runtime, auth, or data-path changes. Review should focus on numeric claims and external links, not production risk.

Overview
Adds a new InferenceX blog post comparing GB300 NVL72 vs GB200 NVL72 on DeepSeek-V4-Pro (Dynamo+vLLM, FP4 8K/1K, disaggregated), with headline 2.83× throughput/GPU and 2.31× perf/$ at 27 tok/s/user, framed as HBM headroom unlocking wider prefill/decode recipes in the 22–32 tok/s/user band. The post includes dashboard CTAs, benchmark/specs figure paths, per-concurrency tables, an iso-interactivity table, acknowledgments, and FAQ JSON-LD.

Also updates the write-inferencex-blog skill so the canonical SemiAnalysis AI Cloud TCO Model link points to https://semianalysis.com/ai-cloud-tco-model/ instead of the newsletter URL.

^{Reviewed by Cursor Bugbot for commit f3a04a8. 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 27, 2026 1:29am

[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 1448436. Configure here.}