feat(blog): port DeepSeekV4 1.6T Day 0 to Day 43 performance article from Substack

[functionstackx]

Summary

Republishes the SemiAnalysis newsletter article DeepSeekV4 1.6T Day 0 to Day 43 Performance Over Time — Huawei, GB300 NVL72, MI355X, B200 (published 2026-06-09) on the InferenceX blog, at the article owner's request.

• New post: packages/app/content/blog/deepseekv4-16t-day-0-to-day-43-performance.mdx (date/publishDate 2026-06-09, matching the original publication)
• 29 figures downloaded at original resolution with descriptive filenames, resampled to ≤1456px; 5 photographic images converted to JPEG (dir total 11MB incl. videos)
• 4 Substack-hosted chart-animation videos (vLLM/SGLang day-0-onward, MI355X over time, GB300 NVL72 over time) ported as local MP4s rendered via raw <video> tags
• InferenceX dashboard links from figure captions preserved as [Live chart](…) links; the dashboard link converted to <DashboardCTA> blocks

Adaptations from the original

• Substack chrome removed (subscribe widgets, share buttons, "Thanks for reading" block)
• ~19 unambiguous typo fixes (e.g. "Turing to" → "Turning to", "DeekSeep" → "DeepSeek", "north start" → "north star", "scare quantity" → "scarce quantity", "is-interactivity" → "iso-interactivity", "compute<>communication lap" → "compute ↔ communication overlap")
• ## SGLang heading demoted to ### so it sits under "What's next for DeepSeek v4" alongside ### vLLM
• Added a republish note at the top linking the original

⚠️ Known gap

The original is a paid post; anonymous access only exposes the free preview, which ends right after the MegaMoE section. The subscriber-only TCO / cost-per-token analysis (H200 vs GB200 NVL72) at the end of the original is not included — the post closes with a pointer to the subscriber edition instead. If the full tail should be included, paste it in and it can be appended.

Verification

• Rendered locally via pnpm dev: post returns 200 at /blog/deepseekv4-16t-day-0-to-day-43-performance, all 29 <figure>s and 4 <video>s present, assets serve with correct MIME types, post listed on /blog
• pnpm typecheck and pnpm fmt pass; pre-commit hook (scoped lint/format/typecheck) passed
• Unofficial-run overlay support: N/A — content-only blog post, no chart code touched

🤖 Generated with Claude Code

---

Note

Low Risk
Content-only addition of a static MDX post using established blog components; no application logic, auth, or data pipeline changes.

Overview
Adds a new InferenceX blog post at deepseekv4-16t-day-0-to-day-43-performance.mdx, republishing the June 9, 2026 SemiAnalysis newsletter piece on DeepSeek V4 Pro 1.6T inference from Day 0 through ~43 days across NVIDIA (B200/B300/GB300), AMD MI355X, and Huawei Ascend 950DT.

The MDX follows existing blog patterns: frontmatter and tags, a link back to the original Substack article, Figure components and local chart assets, four video loops for InferenceX time-series animations, DashboardCTA blocks and preserved [Live chart] InferenceX URLs, and a JsonLd FAQ block for SEO. Editorial tweaks from the source include typo fixes and demoting the SGLang “what’s next” section under vLLM. The paid-newsletter H200 vs GB200 NVL72 TCO tail is omitted and replaced with a pointer to the subscriber edition.

^{Reviewed by Cursor Bugbot for commit ac5924c. 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	Jun 10, 2026 7:43pm

[adibarra]

@claude make sure this matches the same format that all of the other posts use

[github-actions]

Claude finished @adibarra's task in 5m 50s —— View job

---

Format alignment — done

Compared the new post against all 13 other posts in packages/app/content/blog/. The post already follows the established template (frontmatter with date/publishDate/tags, <Figure> with src/alt/caption, <DashboardCTA> blocks, [Live chart] links, ##/### heading hierarchy). Two stylistic choices match the prior republished-newsletter posts (inferencex-v2-…, inferencemax-open-source-…) rather than the newer InferenceX-authored data-driven posts: single-src <Figure> (no srcLight/srcDark) and curly typographic quotes throughout. Both are correct for this content — the figures here are photos/screenshots/diagrams that don't need theme-aware variants, and curly quotes preserve the original Substack copy.

One real deviation found and fixed: this was the only post of 14 missing the <JsonLd> FAQPage SEO block at the end. Added a 6-question FAQ anchored to numbers verified from the article body:

#	Question	Anchor in post
1	MI355X SGLang 26-day gain	>100x performance by Day 26 (line 61, 197–199)
2	GB300 NVL72 cost/M tokens with MTP	$0.156 at 50 tok/s/user, 8K/1K (line 317)
3	DeepSeek V4 CSA/HCA architecture	m'=128, m=4, 50x KV reduction at 1M (lines 492, 500, 508)
4	B200 tok/s/MW improvement	300k → 500k by June 5, 2.17 kW/GPU (lines 289, 299)
5	MegaMoE speedup	1.92x theoretical in DSv4 Flash config (line 520)
6	Day 0 stack support matrix	TRT-LLM FHC_HIDDEN=4096/hidden_size=7168 bug (lines 183–187)

Verification

• JSON parses cleanly (Type: FAQPage, Questions: 6)
• pnpm typecheck, pnpm lint, pnpm fmt all pass
• Vercel preview will rebuild from the push

Commit ac5924c pushed to blog/deepseekv4-16t-day-0-to-day-43-performance.

• Read existing blog posts to understand the standard format
• Read the new post and compare
• Identify format deviations
• Add JsonLd FAQ block (the only deviation)
• Run typecheck/lint/fmt
• Commit and push
  · Branch