brain/
sourcestock-market

Autoresearch: China Blackwell/H100 compute routing — new evidence May 2026

The US relaxed H200 export controls to China in Jan-Mar 2026, but China is now self-blocking purchases to force use of Huawei Ascend. Blackwell remains fully restricted. Smuggling continues at scale.

Source

Autoresearch: China Blackwell/H100 compute routing — new evidence May 2026

Generated by /autoresearch on 2026-05-22. Synthesized across 2 rounds from search snippets (WebFetch blocked systemically — all domains returning 403 for sixth consecutive run; synthesis based on search snippet extracts only, lower confidence than direct-fetch synthesis). No Grokipedia anchor (HTTP 403 on fetch). Treat as raw material — review before promoting. Context: vault/projects/stock-market

Summary

The dominant dynamic since May 19 is a complete inversion of the prior export-control thesis. The US relaxed export controls for H200/AMD MI325X chips in January 2026 and cleared 10 Chinese firms (Alibaba, Tencent, ByteDance) for H200 purchases in May 2026 — but China itself is now blocking purchases to force Chinese tech firms to use domestic Huawei Ascend chips. Nvidia's China revenue risk is now primarily Beijing's choice, not Washington's restriction. Full Blackwell remains banned. A downgraded Blackwell ("B30A" or similar) is planned for June production targeting the China market, but its commercial viability is in question given China's self-blocking stance. Meanwhile, Huawei's Ascend 950PR is now available (Q1 2026), claims 2.8x H20 performance, and DeepSeek V4 became the first frontier-class model trained entirely on Chinese domestic silicon.

Findings

Theme 1: The US relaxed — China self-blocked

BIS revised its license review policy for AI chips to China effective January 15, 2026, shifting from "presumption of denial" to "case-by-case review" for Nvidia H200 and AMD MI325X exports to China and Macau (Morgan Lewis summary via search snippet). Three conditions apply: the export must not reduce US supply available to US customers; the Chinese purchaser must have adopted export compliance procedures; and the chip must pass independent US security testing.

By mid-May 2026, the US Commerce Department cleared approximately 10 Chinese firms — including Alibaba, Tencent, ByteDance, and JD.com — to purchase Nvidia's H200 chip (Yahoo Finance via snippet; CNBC May 14, 2026 as cited in search snippet). Under the deal framework, the US takes a 25% cut of chip sale revenues; chips are routed through US territory.

Despite this US clearance, not a single H200 chip has shipped as of mid-May 2026 (The Next Web headline via snippet). Beijing has blocked domestic firms from purchasing even the approved chips, steering them toward Huawei Ascend instead (Tom's Hardware, search snippet). Trump stated publicly that China "chose not to" approve purchases because "they want to develop their own" (Tom's Hardware snippet). Commerce Secretary Lutnick confirmed Beijing is deliberately steering investment toward domestic chipmakers like Huawei (International Banker / search snippets).

Jensen Huang traveled to China seeking a breakthrough, reportedly picked up by Trump in Alaska en route to Beijing — deal remains unresolved as of mid-May 2026.

Theme 2: Blackwell status and the China-specific chip

Full Blackwell architecture (B200, B300) remains fully restricted to Tier 3 countries including China — no export licenses are being issued (search snippets from BIS, IFP analysis). US authorities have already confirmed Blackwell GPUs are present inside China despite the restriction, indicating ongoing enforcement gaps ([search snippet aggregate]).

Nvidia is developing a downgraded Blackwell-architecture chip for the China market, referred to internally as "B30A" (though Nvidia officially denied this product name) (DataCenter Dynamics via snippet). Planned specs: single-die design (vs. multi-die Blackwell), priced $6,500–$8,000, targeting mass production as early as June 2026 (TBS News via snippet). A second, even cheaper Blackwell-inference chip targeting September production start is also planned (AI News via snippet). The commercial viability of both is in question given China's current block on domestic firm purchases of Nvidia chips.

Nvidia's Q1 FY2027 (calendar Q1 2026) results: $44.1B revenue (+69% YoY), $39.1B data center; however, Nvidia took a $4.5B charge for H20 excess inventory and purchase obligations tied to export restrictions, and guided zero China data center revenue in the following period (Manufacturing Dive / Futurum Group / SEC filing, all via search snippets).

Theme 3: Third-country routing and enforcement

Despite the policy gyrations, smuggling of H100/H200/Blackwell to China via third countries has been extensive. Routing schemes documented through early 2026:

  • Routes: Malaysia, Singapore, Thailand, UAE, and Taiwan-to-Hong Kong direct routes (Wire China, aggregate via snippets)
  • Scale: Between 290,000 and 1.6 million H100-equivalents smuggled to China through end of 2025, with a median estimate of ~660,000 H100-equivalents — representing roughly one-third of China's total AI compute ([Bloomsbury Intelligence and Security Institute estimate, cited in search snippet aggregate])
  • Super Micro case: $2.5B case involving B200/H200 servers routed through Malaysia, defendants charged in late March 2026 (tech-insider.org via snippet)
  • DOJ Operation Gatekeeper: Announced December 2025, disrupted $160M in chip smuggling, multi-defendant network ([search snippet from early 2026 sources])
  • Taiwan detentions: Three persons detained in Taiwan for alleged illegal high-end AI server exports to China via SuperMicro supply chain (The Next Web via snippet)
  • BIS budget: Congress approved 23% increase in BIS FY2026 budget with bipartisan support, several million marked specifically for semiconductor enforcement

Theme 4: Huawei Ascend as the structural alternative

China's self-blocking of Nvidia purchases is credible because Huawei's domestic alternative has reached meaningful capability:

  • Ascend 950PR (Q1 2026 availability): 1.56 petaflops FP4, ~2.8x Nvidia H20 performance, in-house HBM (HiBL 1.0 — 2.5x prior-gen bandwidth), manufactured by SMIC on N+2 (7nm-class) (TrendForce, March 2026, via snippet)
  • Atlas 950 SuperPoD: 8,192 Ascend chips, Q4 2026 availability; claimed scale to 1 million NPUs (TrendForce 2025 announcement via snippet)
  • DeepSeek V4: Reuters confirmed (April 2026) DeepSeek V4 runs on Huawei Ascend 950PR — the first frontier-class AI model built entirely on Chinese domestic silicon ([search snippet aggregate])
  • Roadmap: Ascend 960 (Q4 2027), Ascend 970 (Q4 2028) — committed multi-year domestic roadmap

Beijing's claim that domestic chips "now match H20 and RTX Pro 6000D" is backed by the 2.8x H20 performance claim for Ascend 950PR — though independent verification is limited. If accurate, the functional gap between China's domestic AI compute and what Nvidia's H200 would offer has narrowed materially.

Contradictions and open questions

  • Will China ever clear H200 purchases? The US cleared 10 firms; Jensen Huang is in Beijing seeking a deal; but China's domestic-chip-first industrial policy appears structural, not transactional. The H200 deal may never close.
  • B30A commercial viability: If China is blocking H200 purchases, why would a downgraded Blackwell sell? Nvidia's bet seems to be that once China's policy thaws, a compliant product is available. But the timeline risk is high.
  • Ascend 950PR claim of 2.8x H20 needs independent benchmark: The TrendForce/Huawei claims are from Huawei itself. Actual inference performance parity with independently-tested chips has not been confirmed.
  • Does China's self-block kill the smuggling thesis? If domestic chips are competitive, Chinese firms have less incentive to pay smuggling premiums for H100/H200. The smuggling channel may be winding down organically even without US enforcement success.
  • SMIC as the winning public company: If Ascend 950PR is manufactured at SMIC and China continues its domestic-chip-first policy, SMIC (SMICY) is the most directly benefited public company. This is a new hypothesis worth opening.

Provenance

Rounds run: 2 of 3 (early-exit — round 3 sub-questions would not materially change synthesis given WebFetch failure)

Sub-questions by round:

Round 1 (broad survey):

  1. New BIS enforcement actions or rule changes for AI chip exports to China, May 2026
  2. Third-country routing schemes (Singapore, Malaysia, UAE) for Blackwell/H100 reaching China
  3. Nvidia China revenue exposure and compliance posture changes Q1-Q2 2026
  4. Chinese domestic AI chip alternatives (Huawei Ascend, Cambricon) as compute substitute

Round 2 (drill-down):

  1. China self-blocking Nvidia H200 purchases — mechanism, firms covered, Trump/BIS reaction — targeted the surprising policy reversal
  2. Nvidia "B30A"/China Blackwell chip — specs, timeline, revenue opportunity
  3. Huawei Atlas 950 SuperPoD / Ascend 950PR timeline and scale

Anchor source: fetch failed: HTTP 403 on Grokipedia United_States_New_Export_Controls_on_Advanced_Computing_and_Semiconductors_to_China

URLs fetched: 0 successful, 11 failed (all HTTP 403 — WebFetch systemically blocked, sixth consecutive run). Synthesis built entirely from search snippet extracts. Confidence: moderate — snippets contain specific figures, dates, and named sources, but primary text could not be verified.

Attempted fetches (all 403):

  • bis.gov/news-updates
  • thewirechina.com/2026/03/01/chasing-the-chip-smugglers...
  • sec.gov/...nvda-20260426.htm
  • tech-insider.org/super-micro-nvidia-chip-smuggling...
  • tomshardware.com/...deepseek-research-suggests-huaweis-ascend-910c...
  • manufacturingdive.com/news/nvidia-q1-2026-earnings...
  • morganlewis.com/pubs/2026/01/bis-revises-export-review...
  • sanctionsnews.bakermckenzie.com/bis-revises-license-review...
  • chathamhouse.org/2026/04/ai-export-controls...
  • thenextweb.com/news/taiwan-detains-three-ai-server...
  • spectrum.ieee.org/china-ai-chip

Tools used: WebSearch, WebFetch (all failed), grokipedia-fetch (403 on fetch). Generated: 2026-05-22

Referenced by