RegulationRules, enforcement & market structure
China Puts Conditions on Nvidia H200 Purchases and Pushes AI Labs Toward Huawei Chips
Beijing has attached conditions to ByteDance, Alibaba and Tencent's Nvidia H200 purchases and is pushing labs toward Huawei. It is betting that a domestic chip supply chain is worth slower model progress now, though Huawei's 2026 compute output is estimated far below Nvidia's.
From the Odd Lots podcast episode of November 21, 2026 and cited sources: Washington cleared H200 sales, but Beijing approved purchases of more than 400,000 of the chips by ByteDance, Alibaba and Tencent only under conditions that were still being set. [1] [3] [5]
SOURCES[1] How AI Is Forcing China to Reshape Industrial Policy omny.fm[3] Reuters via Yahoo: China approves H200 purchases finance.yahoo.com[5] Al Jazeera: H200 clearance aljazeera.com
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Washington cleared H200 sales, but Beijing approved purchases of more than 400,000 of the chips by ByteDance, Alibaba and Tencent only under conditions that were still being set. [1] [3] [5]
- Original source
- Odd Lots podcast episode, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 21, 2026 podcast episode published
- Covered
- Nov 21 edition
- Evidence
- 22 sources across 17 sites; transcript sourced
- 1How AI Is Forcing China to Reshape Industrial Policyomny.fm
- 2Transcript: How AI Is Forcing China to Reshape Industrial…api.omny.fm
- 3Reuters via Yahoo: China approves H200 purchasesfinance.yahoo.com
- 4Epoch AI: Huawei's roadmap to 2031epoch.ai
- 5Al Jazeera: H200 clearancealjazeera.com
- 6Kreab: Weekly China Insightkreab.com
Reach: not measured yet

Washington cleared H200 sales, but Beijing approved purchases of more than 400,000 of the chips by ByteDance, Alibaba and Tencent only under conditions that were still being set. One option discussed was requiring buyers to take a set share of domestic chips. Brookings fellow Kyle Chan describes this as industrial policy: Beijing would give up some model progress today to build a chip supply chain that US decisions cannot cut off. [1] [3] [5]
Huawei makes up for weaker individual chips by linking thousands of them into large clusters, but production volume is the main limit. Huawei claims its Atlas 950 SuperPoD offers 6.7 times the compute of Nvidia's planned NVL144. Epoch AI rates the Ascend 950 at about half an H100 and projects Huawei's 2026 compute output at under 4% of Nvidia's. Huawei is also short of HBM memory and has reportedly told some customers to find their own. [1] [4] [9] [14]
For developers, the shift means porting work and more varied hardware and software stacks. DeepSeek V4 needed major architecture rewrites to run on the Ascend 950PR. Chinese models increasingly launch ready to run on domestic chips, and DeepSeek reportedly plans to train V5 on Ascend. [1] [15] [8]
Washington is closing the offshore route. A Commerce rule issued November 16 requires licenses to ship advanced chips to Indonesia, Malaysia, Singapore, Thailand and Brazil, targeting Chinese firms that rent Nvidia compute in data centers abroad. The rule came days after Jensen Huang resigned as Nvidia CEO following reports linking him to the Megaspeed scandal. [12] [10] [11] [20]
The reviewed sources show Beijing limiting H200 access and labs moving models onto Ascend chips.
Read the full assessment
A likely but unproven result is a lasting split in AI hardware and software, with higher porting costs for developers and a smaller China market for Nvidia.
Executive brief
Beijing is limiting its own AI companies' access to Nvidia chips. Even after Washington cleared H200 sales, China gave only conditional approvals and pushed labs toward Huawei (Reuters via Yahoo). On Bloomberg's Odd Lots, Brookings fellow Kyle Chan explains this as industrial policy: Beijing would accept slower model progress today in exchange for a chip supply chain Washington can't cut off (transcript, 25:31). The cost is large. Epoch AI estimates Huawei will produce less than 4% of Nvidia's compute output in 2026 (Epoch AI). The episode was recorded on November 11, one day before Jensen Huang resigned as Nvidia CEO.
What changed and event timeline
US clears H200 for China
Trump allowed H200 sales to "approved customers," with the US government taking a 25% fee ().
Beijing grants conditional approvals
ByteDance, Alibaba and Tencent reportedly got approval to buy more than 400,000 H200s between them. The conditions were still being set, and one option discussed was requiring a fixed ratio of domestic chips with each purchase ().
15th Five-Year Plan finalized
The plan for 2026–2030 puts semiconductor and AI self-reliance and "new quality productive forces" at its center (;).
Reports of DeepSeek on Ascend and an HBM squeeze
STAR Market Daily reported that DeepSeek plans to train its V5 model on Ascend chips (). The Information reported that Huawei told some customers to source their own HBM ().
Commerce closes the offshore-leasing loophole
A new BIS interim final rule requires licenses to ship advanced chips to Indonesia, Malaysia, Singapore, Thailand and Brazil. The target is Chinese firms renting Nvidia compute in data centers abroad ().
Capabilities and access
- Ascend 950PR: due Q1 2026, with 128GB of Huawei's own HBM at 1.6 TB/s. 950DT: 144GB at 4 TB/s, aimed at training (Tom's Hardware).
- Atlas 950 SuperPoD: 8,192 950DT chips in 160 cabinets, planned for Q4 2026 (TrendForce).
- Ascend 960DT: reportedly will rely mostly on Samsung and SK Hynix memory (TNW).
Read the full section
- Ascend 950PR: due Q1 2026, with 128GB of Huawei's own HBM at 1.6 TB/s. 950DT: 144GB at 4 TB/s, aimed at training (Tom's Hardware).
- Atlas 950 SuperPoD: 8,192 950DT chips in 160 cabinets, planned for Q4 2026 (TrendForce).
- Ascend 960DT: reportedly will rely mostly on Samsung and SK Hynix memory (TNW).
- H200 in China: available only under Beijing's conditional approvals.
Technical analysis for researchers and developers
- Quantity over quality: Chan says Huawei uses its networking expertise to link many weaker chips into "super pods" large enough for pretraining runs (transcript, 16:11).
- Porting cost: DeepSeek V4 needed "extensive adjustments and rewrites" to its architecture, done with Huawei and Cambricon, to run on Ascend 950PR, which is mainly an inference chip (TrendForce).
- Reproducibility gap: Zhipu trained GLM-Image (a 9B generator plus a 7B decoder) on Atlas 800T A2 servers using MindSpore.
Read the full section
- Quantity over quality: Chan says Huawei uses its networking expertise to link many weaker chips into "super pods" large enough for pretraining runs (transcript, 16:11).
- Porting cost: DeepSeek V4 needed "extensive adjustments and rewrites" to its architecture, done with Huawei and Cambricon, to run on Ascend 950PR, which is mainly an inference chip (TrendForce).
- Reproducibility gap: Zhipu trained GLM-Image (a 9B generator plus a 7B decoder) on Atlas 800T A2 servers using MindSpore. It did not disclose how many accelerators it used or how long training took (The Register).
Claims and evidence
- Huawei vs. Nvidia NVL144: Huawei says the Atlas 950 SuperPoD has 6.7× the compute of Nvidia's planned NVL144 system.
- Independent estimate: Epoch AI puts Ascend 950 at roughly half an H100, with about 1.5M Huawei units in 2026 against about 5.9M for Nvidia (Epoch AI).
- DeepSeek V5 on Ascend: this rests on anonymous industry sources and has not been confirmed by DeepSeek or Huawei (TechNode).
Read the full section
- Huawei vs. Nvidia NVL144: Huawei says the Atlas 950 SuperPoD has 6.7× the compute of Nvidia's planned NVL144 system. This is Huawei's own claim, not an independent benchmark (TrendForce).
- Independent estimate: Epoch AI puts Ascend 950 at roughly half an H100, with about 1.5M Huawei units in 2026 against about 5.9M for Nvidia (Epoch AI).
- DeepSeek V5 on Ascend: this rests on anonymous industry sources and has not been confirmed by DeepSeek or Huawei (TechNode).
- Effect of export controls: Chan argues they "massively" galvanized China's chip push, and acknowledges the counterargument that the effort began before them (transcript, 21:46).
Context and prior work
Chan has written on full-stack competition for Brookings, and on how chip bans can backfire (Brookings). CFR argues the opposite case: Huawei cannot catch Nvidia, so controls should stay (CFR). Megaspeed arranged nearly $12B of Nvidia chips for Chinese firms between March 2024 and June 2025 (Fortune).
Limitations, safety and contested findings
- In Epoch's domestic-HBM scenario, Huawei's compute output stays around 1% of Nvidia's through 2028. SMIC is capped at roughly 5–7nm-equivalent processes (Epoch AI).
- Problems with Huawei chips reportedly nearly derailed DeepSeek's R2 model in 2025 (The Register).
- Chan calls the Shenzhen EUV prototype only a rumor, and labels his own view that Xi deprioritizes consumers a "controversial take" (transcript, 28:28).
Read the full section
- In Epoch's domestic-HBM scenario, Huawei's compute output stays around 1% of Nvidia's through 2028. SMIC is capped at roughly 5–7nm-equivalent processes (Epoch AI).
- Problems with Huawei chips reportedly nearly derailed DeepSeek's R2 model in 2025 (The Register).
- Chan calls the Shenzhen EUV prototype only a rumor, and labels his own view that Xi deprioritizes consumers a "controversial take" (transcript, 28:28).
- The auto-captions garble names: "CSMT" is CXMT and "Heineck" is SK hynix.
Business and practitioner implications
- Hardware fragmentation: Chinese models increasingly ship ready for Ascend and other domestic chips on day one, so developers face more varied hardware and software stacks (transcript, 16:11).
- Memory cost: one Shenzhen broker says a single HBM3E unit now sells for about $4,000 in China, up from several hundred dollars a year earlier (TNW).
- Southeast Asia compute: offshore GPU capacity in the region now needs licenses (CNBC).
Read the full section
- Hardware fragmentation: Chinese models increasingly ship ready for Ascend and other domestic chips on day one, so developers face more varied hardware and software stacks (transcript, 16:11).
- Memory cost: one Shenzhen broker says a single HBM3E unit now sells for about $4,000 in China, up from several hundred dollars a year earlier (TNW).
- Southeast Asia compute: offshore GPU capacity in the region now needs licenses (CNBC).
- Markets: Nvidia shares reportedly fell to eight-month lows during the scandal (Al Jazeera). That is a price move, separate from fundamentals. On profits, Chan says low barriers to entry erode Chinese firms' competitive moats (transcript, 38:22).
The source trail.
Sources (22)
How AI Is Forcing China to Reshape Industrial Policy
Transcript retrieved via published_transcript; language en. Timestamped text, not direct audiovisual review. Source: https://api.omny.fm/orgs/e73c998e-6e60-432f-8610-ae210140c5b1/clips/bb74684d-fcbe-4c36-9779-b5070115b930/transcript?format=TextWithTimestamps&t=1795107304. Automatic captions/transcription may contain errors.
omny.fmTranscript: How AI Is Forcing China to Reshape Industrial Policy
published_transcript
api.omny.fm