Chinese-market NVIDIA accelerators (H800 / H20 / export status) — web sweep 2026-07-09

Agent: h20-researcher (Claude web-research subagent, Exa+Brave). Cross-check source for the calculator's Chinese-accelerator entries; pairs with the GPT Pro Chinese-accelerator dive and the Ascend sweep.

1. H800 vs H100

H800 SXM (the variant DeepSeek used): compute DISCLOSED-identical to H100 SXM — same GH100 die, 80GB HBM3, 3.35 TB/s, dense FP8 ≈1,979 TFLOPS. The one deliberate cut: NVLink capped at 400 GB/s vs 900 GB/s (REPORTED, concordant). Training/multi-GPU penalty, not per-chip inference penalty.

• https://www.philisun.com/technical-analysis/nvidia-h800-vs-h100-a-technical-deep-dive-into-politically-defined-ai-chips/
• https://electronics.alibaba.com/buyingguides/nvidia-h800-vs-h100-ai-gpu-choice-guide

H800 PCIe (separate weaker SKU): 80GB, 2.0 TB/s, dense FP8 1,513 TFLOPS, 350W. DISCLOSED (vendor sheet): https://flopper.io/gpu/nvidia-h800-pcie-80gb/spec-sheet.pdf

China rental: no clean current H800 $/hr. DeepSeek's own $2/hr assumption is DISCLOSED (V3 tech report, https://arxiv.org/pdf/2412.19437). Trend proxy: TrendForce Apr 2026 — H100 one-year leases in China up 20–30% in ~6 months (RMB 50–60k → 80–90k rack-inclusive) on >1,000× token-volume growth (100B→140T tokens/day, early-2024→Mar-2026, National Data Administration). https://www.trendforce.com/news/2026/04/17/news-nvidia-h100-rentals-in-china-reportedly-up-20-30-as-token-demand-surges-h-series-also-rises/

2. H20

DISCLOSED specs (concordant: inferencebench.io, willitrunai): 96 GB HBM3, 4.0 TB/s, dense FP8/INT8 296 TFLOPS, BF16 148 TFLOPS, TDP 400W (one outlier source says 500W), NVLink retained ~900GB/s.

• https://inferencebench.io/gpus/nvidia-h20/
• https://willitrunai.com/gpus/h20-96gb

Inference-friendly because decode is bandwidth-gated: 4.0 TB/s > H100's 3.35 TB/s. futunn/PETA 2024 analysis: H20 ~20% higher peak tok/s than H100 at low batch (REPORTED, dated): https://news.futunn.com/en/post/34012455/compared-to-the-h100-how-are-nvidia-s-h20-l20

Price: launch $12–15k (REPORTED); mid-2026 ~$10–12k (TechPowerUp citing Reuters, Jun 2026): https://www.techpowerup.com/340664/nvidias-export-compliant-b30a-for-china-priced-at-roughly-twice-the-h20

Rental China: fragmentary. 191idc aggregator: RMB 4.82/hr ($0.67/hr) on-demand floor, subsidized/oversupply listing — ESTIMATE floor, not market-representative: http://www.191idc.com/jiageinfo/402.html. GAP: no official Aliyun/Tencent/Volcano H20 rate card found.

3. Export-control timeline 2025–2026

B30A as of Jul 2026: treat as still blocked (one contradictory TechPowerUp/Reuters pricing piece ~$20–24k, discount vs Huang's May statement).

Legal-to-buy today (Jul 2026): H20 (case-by-case, main legal SKU); H200 (10 cleared firms, deliveries just now plausible); everything else (H100/B200/B300/Rubin) banned — smuggling only.

Key sources: https://www.bis.gov/press-release/department-commerce-revises-license-review-policy-semiconductors-exported-china · https://www.nbcnews.com/world/asia/us-approves-nvidia-h200-chip-exports-china-conditions-rcna253948 · https://www.reuters.com/business/retail-consumer/us-clears-h200-chip-sales-10-china-firms-nvidia-ceo-looks-breakthrough-2026-05-14/ · https://www.bloomberg.com/news/articles/2026-07-08/china-to-let-ai-firms-buy-nvidia-h200-chips-information-says

4. Fleet reality (mid-2026)

• DeepSeek: V3/R1 trained on H800 (DISCLOSED). V4-class: US officials believe smuggled Blackwell in Inner Mongolia (REPORTED, Reuters Feb 2026). Tried Ascend for training → failed run → back to Nvidia for training, Ascend for inference (REPORTED, CAISI via peterwildeford/Grunewald: https://blog.peterwildeford.com/p/how-banned-ai-chips-end-up-in-china). Jul 2026: developing own inference ASIC (~1 yr in): https://www.businesstimes.com.sg/companies-markets/telcos-media-tech/chinas-deepseek-developing-its-own-ai-chip-sources
• Zhipu: GLM-Image trained fully on Ascend Atlas 800T A2 (Jan 2026, company-DISCLOSED): https://www.theregister.com/software/2026/01/15/chinas-zai-trained-a-model-using-only-huawei-hardware/4198774 ; The Information Jul 7 2026: weighing custom chip.
• Alibaba/ByteDance/Tencent: Aug 2025 ordered to suspend Nvidia purchases (REPORTED); Alibaba PPU "Zhenwu 810E" ~H20-class + 10k-chip self-built DC (Apr 2026, DISCLOSED); all among the 10 firms cleared for H200.
• Moonshot: GAP — no chip-specific sourcing found.
• Smuggling aggregate: Epoch AI ~660k H100-equivalents through end-2025 (90% CI 290k–1.6M) ≈ ⅓ of China's compute: https://epochai.substack.com/p/diversion-and-resale-estimating-compute · Substrate ~20k B300-equivalents in 2026 (CI 2k–100k): https://www.the-substrate.net/p/where-will-china-get-its-compute
• Huawei self-claims: 910C ~800 TFLOPS FP16 (~80% H100); Ascend 950 1.56 PF + 112GB, "3× H20 inference at ¼ cost" (self-reported, skepticism warranted): https://www.tomshardware.com/tech-industry/semiconductors/chinas-huawei-to-enter-south-korean-ai-chip-market-with-new-atlas-superpods-clusters-pack-8-192-ascend-950-accelerators-per-deployment-reportedly-challenges-nvidia-dominance-with-tripled-inference-performance-of-h20-at-one-quarter-the-cost

5. H20 MoE throughput anchors (MFU calibration)

• LMSYS/SGLang + Ant Group (Sep 2025, production, DISCLOSED): DeepSeek-R1 on H20-96G — 16.5k prefill / 5.7k decode tok/s per 8-GPU node; ≈714 tok/s/GPU decode at BS=48 (423 at latency-traded spec-decode config). https://www.lmsys.org/blog/2025-09-26-sglang-ant-group/
• Alibaba InferSim (Apr 2026, validated ±10-15%): measured Qwen3-30B-A3B on H20: prefill 16,594 / decode 2,749 tok/s/GPU. Measured DeepSeek-V3 on H800: prefill 7,839 / decode 2,324 tok/s/GPU. https://pyshine.com/InferSim/
• StepFun Step3 on 8×H20 (AMD-sponsored framing; StepFun-attributed baselines): decode 3,147 tok/s aggregate (~393 tok/s/GPU, high-concurrency short-output operating point). https://moreh.io/customer-case/step3-inference-optimization-on-amd-instinct-mi308x-1-30x-higher-decode-throughput-vs-nvidia-h20-251229/

Recommendation: anchor H20 on InferSim (validated, error-bounded); LMSYS 714 tok/s/GPU as the large-model production lower bound.

Gaps

1. No official hyperscaler H20 rate card. 2. No Moonshot hardware sourcing. 3. B30A status has one contradictory source (discounted). 4. H20 TDP 400W vs 500W discrepancy (400W primary).