Is the China-via-Malaysia Blackwell compute routing loophole durable, and does it structurally expand Nvidia's TAM?
Is the China-via-Malaysia Blackwell compute routing loophole durable, and does it structurally expand Nvidia's TAM?
The question
MAJOR REFRAME (May 22, 2026): The original question focused on whether the US would close the offshore-compute-routing loophole. The more important dynamic as of May 2026 is the reverse: the US relaxed H200 export controls (Jan 2026 BIS policy change + May 2026 clearance of 10 Chinese firms), but China itself is blocking purchases to force domestic tech firms onto Huawei Ascend. The question has shifted from "will the US close the loophole?" to "will China ever allow Nvidia back in, and does Huawei Ascend make US chips structurally irrelevant to China's AI build-out?"
Original framing: ByteDance is accessing a 36,000-chip Blackwell B200 GPU cluster physically located in Malaysia (owned by Aolani Cloud) as a cloud tenant — a legally compliant arrangement under export controls that regulate where chips are shipped, not where their compute output is accessed.
Why it matters
If China's domestic-chip-first policy is structural:
- Nvidia's China data center revenue is effectively zero regardless of US export policy.
- Huawei Ascend (SMIC-fabricated) becomes the de facto standard for Chinese AI compute.
- The offshore-compute-routing thesis collapses — Chinese firms won't pay offshore premiums if they're mandated onto domestic silicon.
- SMIC (SMICY) is the emerging public beneficiary of China's AI compute buildout.
If China eventually allows H200 (or B30A) purchases:
- Nvidia recovers a potentially large market. H200 deal terms reportedly include a 25% US revenue cut — unusual structure suggesting the deal, if it closes, would be highly negotiated.
- The B30A (single-die, $6.5–8K, June 2026 production target) is positioned as a compliant Blackwell chip for China.
What we currently believe
- The US relaxed, China blocked: BIS revised policy Jan 15, 2026 from "presumption of denial" to "case-by-case review" for H200/MI325X to China. By mid-May 2026, ~10 Chinese firms (Alibaba, Tencent, ByteDance, JD.com) were cleared for H200 purchases. But not a single H200 has shipped — Beijing blocked domestic firms from buying, steering them to Huawei Ascend.
- China's rationale is credible because Ascend 950PR is real: Huawei Ascend 950PR (Q1 2026) claims 1.56 PFLOPS FP4, 2.8x H20 performance, in-house HBM (HiBL 1.0), manufactured by SMIC on N+2 (7nm-class). DeepSeek V4 (Reuters, April 2026) is the first frontier-class model trained entirely on domestic silicon. Beijing's claim that domestic chips "now match H20" is not fully independently verified, but is directionally credible.
- Nvidia $4.5B H20 charge: Nvidia took a $4.5B charge for H20 excess inventory and purchase obligations in Q1 FY2027 (reported May 2026) and guided zero China data center revenue for the next period. The H20 era is closing regardless of export-policy outcome.
- Full Blackwell remains banned: B200/B300 are fully restricted to Tier 3 (including China). B30A (single-die Blackwell, lower spec, ~$6.5–8K) is being developed for China market; June 2026 production target. Commercial viability uncertain given China's purchase block.
- Offshore routing (Malaysia loophole) was real: ByteDance/Aolani Cloud Malaysia (36K B200 GPUs, ~$2.5B) was a real arrangement. But if China mandates domestic silicon, the offshore-access model becomes less attractive even without US closure.
- Smuggling at scale: 290K–1.6M H100-equivalents smuggled to China through end of 2025 (median ~660K, ~⅓ of China's total AI compute). DOJ Operation Gatekeeper (Dec 2025, $160M). Three detained in Taiwan in Super Micro supply chain case. BIS enforcement budget +23% in FY2026.
- Jensen Huang in Beijing: Huang traveled to China seeking a deal (Trump reportedly involved in brokering at Alaska stop). No resolution as of mid-May 2026.
Evidence we have
- From 2026-05-19-autoresearch-ai-infrastructure-macro-scan-may-19-2026: ByteDance Aolani Cloud Malaysia: 36K Blackwell B200 GPUs, 500 NVL72 rack-scale systems, ~$2.5B total. Nvidia confirms no export violation. ByteDance not on Entity List; hardware physically in Malaysia.
- From 2026-05-19-autoresearch-ai-infrastructure-macro-scan-may-19-2026: Loophole mechanism: US export controls regulate where hardware is shipped, not where compute output is accessed. A Chinese company can legally access Blackwell compute if hardware is offshore.
- From 2026-05-19-autoresearch-china-geopolitics-ree-taiwan-may-2026: Huawei Ascend 910C at ~60% H100 inference performance; CloudMatrix 384 beats GB200 NVL72 in some workloads at 5x more processors.
- From 2026-05-22-autoresearch-china-blackwell-compute-routing-may-2026: BIS revised review policy Jan 15, 2026 (presumption of denial → case-by-case); ~10 Chinese firms cleared for H200 by May 2026; not a single H200 has shipped; Beijing blocked domestic purchases to favor Huawei Ascend. Trump confirmed China "chose not to" buy because "they want to develop their own."
- From 2026-05-22-autoresearch-china-blackwell-compute-routing-may-2026: Nvidia $4.5B H20 inventory/obligation charge; zero China DC revenue guidance. $44.1B Q1 FY2027 revenue (+69% YoY), $39.1B data center, but China impact fully provisioned.
- From 2026-05-22-autoresearch-china-blackwell-compute-routing-may-2026: Ascend 950PR — 1.56 PFLOPS FP4, 2.8x H20, in-house HBM (HiBL 1.0), SMIC N+2 7nm-class, Q1 2026 available. DeepSeek V4 (Reuters April 2026) trained entirely on Ascend 950PR — first frontier model on domestic silicon.
- From 2026-05-22-autoresearch-china-blackwell-compute-routing-may-2026: B30A development confirmed (Nvidia denies the name) — single-die Blackwell, $6.5–8K, June 2026 production target. Also second cheaper Blackwell inference chip for September start.
- From 2026-05-22-autoresearch-china-blackwell-compute-routing-may-2026: Smuggling scale: 290K–1.6M H100-equivalents via Malaysia, Singapore, Thailand, UAE, Taiwan-HK routes through 2025. DOJ Operation Gatekeeper (Dec 2025); Super Micro $2.5B case; Taiwan detentions. BIS budget +23%.
Evidence we need
- Does China ever clear the H200 purchase block? Jensen Huang's Beijing visit, Trump's involvement — watch for Commerce/State announcement or Nvidia guidance update.
- B30A commercial traction: Will Chinese firms buy a compliant downgraded Blackwell if offered? Revenue materiality vs. $4.5B H20 charge context.
- Ascend 950PR independent benchmark: Huawei's 2.8x H20 claim is self-reported. Independent performance data would confirm whether China's domestic-chip-first policy is sustainable.
- Malaysia/SE Asia loophole durability: If Chinese firms are forced onto Ascend domestically, does the offshore-routing model (ByteDance/Aolani) persist or wind down?
- SMIC revenue from Ascend: Does SMIC disclose Ascend 950PR fabrication volume and revenue contribution?
How to resolve
- Watch Nvidia Q2 FY2027 (calendar Q2 2026) guidance — any change from "zero China DC revenue" would be the clearest signal the H200 deal is advancing.
- Watch Commerce/BIS for B30A export license framework — if the US licenses a Blackwell-class chip for China sale, it's a structural change.
- Watch Huawei/Ascend quarterly shipment estimates from TrendForce for scale of domestic uptake.
- Watch SMIC quarterly earnings for any disclosure of Ascend-related foundry revenue.
Related
- china-rare-earth-november-2026-deadline — parallel China/US tech confrontation
- huawei — domestic alternative that gains if Nvidia blocked
- nvidia — primary beneficiary if China deal closes; B30A is the near-term vehicle
- section-232-phase-2-us-fab-premium — parallel export-control / industrial-policy framing
- smic — potential new entity: SMIC fabricates Ascend 950PR at N+2; China's domestic AI compute buildout flows through SMIC