Monitor this situation.
Unsubscribe anytime.
[SITUATION] · [QUIET] · [TECHNOLOGY]
6 clusters · 63 sources · 32 days · First seen · Last updated
AI data center expansion and domestic hardware shift
Overview
Inner Mongolia is becoming a central hub for China’s digital infrastructure due to its low energy costs, cooler climate, and proximity to Beijing. Major technology firms are increasingly establishing a presence in the region to support energy-intensive computing needs. ByteDance is reportedly in preliminary discussions to expand its computing infrastructure in Ulanqab’s Jining district, potentially integrating five to six gigawatts of new compute power. DeepSeek is advancing plans for a large-scale data center in Ulanqab that would deploy at least 160,000 Huawei Ascend 950DT AI accelerators. The order is estimated to be worth approximately $2.56 billion and marks a significant shift toward domestic hardware for the company’s inference needs.
While DeepSeek has previously relied on Nvidia hardware for model training, leadership has noted that while Huawei’s hardware may require more units to match the performance of a single Nvidia GPU, the technology is “increasingly viable” for their requirements. To support this transition, DeepSeek and Huawei have expanded their collaboration to develop a comprehensive AI software stack for Ascend processors, aiming to build an independent GPU ecosystem and reduce reliance on NVIDIA’s CUDA. This partnership has moved into hardware-level optimization, including a supernode system integrating 128 Ascend 950 chips.
To further compete with Nvidia, the collaboration is expanding development tools for the Ascend platform. This includes optimizing both computation and communication for the supernode system and utilizing TileLang, an open-source high-level programming language for AI chips. TileLang is designed to simplify code logic and improve development efficiency, seeking to lower switching costs for developers moving to architectures outside of Nvidia. Recent developments indicate that DeepSeek and Huawei continue to co-develop programming infrastructure for the Ascend platform to facilitate this transition.
Entities
DeepSeek · Huawei · Nvidia · Inner Mongolia · Ascend
Claims
What the coverage asserts, and how many sources carry each claim.
- [● 16 SOURCES] DeepSeek is collaborating with Huawei to develop programming tools for Ascend chips. www.ithome.com
- [● 14 SOURCES] The open-source components include the TileLang high-level programming language. www.ithome.com
- [● 13 SOURCES] The collaboration aims to create an independent GPU software ecosystem to reduce reliance on NVIDIA's CUDA.
- [● 7 SOURCES] DeepSeek and Huawei worked on a supernode system combining 128 Ascend 950 chips.
- [● 4 SOURCES] DeepSeek plans to deploy at least 160,000 Huawei AI accelerators at a new data center in Inner Mongolia.
- [● 4 SOURCES] DeepSeek intends to use Huawei Ascend 950DT processors primarily for AI model inference rather than training.
- [● 4 SOURCES] Huawei faces production capacity constraints and component shortages that may limit Ascend 950DT output to the low hundreds of thousands this year.
- [● 4 SOURCES] 80% of survey participants reported AI improved individual productivity
- [● 4 SOURCES] Only 37% of organizations reported that AI contributed positively to profits before interest and taxes
- [● 3 SOURCES] The open-source release includes DeepGEMM, DeepEP, TileKernels, FlashMLA, and DeepSelect. www.ithome.com
- [● 2 SOURCES] Huawei's Peerium architecture can enable up to one million processors to operate as a single computer.
Timeline
-
[TECHNOLOGY] 23 sourcesAI productivity paradox: Individual gains vs. corporate profits
A global productivity paradox has emerged where 80% of workers report AI-driven productivity gains, yet only 37% of companies see improved profits as organizational integration remains a major hurdle.
-
[TECHNOLOGY] 2 sourcesDeepSeek and Huawei drive shift toward lightweight AI hardware and reduced CUDA dependency
The AI hardware market is shifting toward lightweight integrated machines as high-end LLM units face cost and scaling issues. Meanwhile, DeepSeek and Huawei are open-sourcing software to reduce CUDA dependency.
-
[TECHNOLOGY] 27 sourcesDeepSeek and Huawei partner to develop open-source AI software for Ascend chips
DeepSeek and Huawei are open-sourcing a new software stack, including the TileLang language, to support Huawei Ascend AI chips and challenge Nvidia's CUDA dominance in China.
-
[TECHNOLOGY] 4 sourcesDeepSeek to purchase 160,000 Huawei AI chips for new data center
DeepSeek plans to purchase 160,000 Huawei Ascend 950DT AI chips for a massive new data center in Inner Mongolia, signaling a move toward domestic Chinese hardware for AI inference.
-
[TECHNOLOGY] 11 sourcesDeepSeek plans massive Huawei chip deployment in Inner Mongolia
DeepSeek plans to deploy 160,000 Huawei AI chips at a new Inner Mongolia data center to reduce Nvidia reliance, while SMIC reports record profits driven by domestic AI demand.
-
[TECHNOLOGY] 2 sourcesByteDance plans massive AI data center expansion in Inner Mongolia
ByteDance is planning a massive expansion of AI data centers in Inner Mongolia, potentially investing up to 960 billion yuan to add five to six gigawatts of computing power in Ulanqab.
Sources
appbank.net · ascii.jp · biznis.rs · blog-nouvelles-technologies.fr · businessinsider.es · businessinsurance.com · cafebiz.vn · canadanewsmedia.ca · citybuzz.co · congnghe.vn · congnghevadoisong.vn · cryptopolitan.com · dev.to · devdiscourse.com · drweb.de · finance.technews.tw · finanzen.at · htv.vn · huaweicentral.com · influencermagazine.uk · infovisionrd.com · innovation-village.com · interempresas.net · interest.co.nz · it-bg.org · it-boltwise.de · ithome.com · jamestown.org · japan.zdnet.com · klgadgetguy.com · merca2.es · news.az · news.cnyes.com · news.mynavi.jp · newsbytesapp.com · nyreport.com · ocio.diariodemallorca.es · oglobo.globo.com · onlinecourses.ooo · prnoticias.com · prtimes.jp · que.com · tahawultech.com · tech360.tv · techgenyz.com · technice.com.tw · techstory.in · techtalent.ca
This summary has been updated 9 times: see revision history