DeepSeek Watch posted an update
DeepSeek is working closely with Huawei on the design of future AI chips, according to The Information, in a collaboration that could help China build frontier AI models without Nvidia technology.
Why it mattersThe report says Huawei engineers are incorporating DeepSeek’s requests into chip design decisions, including how models calculate and move data between processors. The chips are expected next year and beyond, and Huawei engineers have reportedly been stationed at DeepSeek’s offices. If accurate, this is more than a customer asking for a better chip: it puts a model developer directly into the design process.
Discuss: How much influence should AI developers have over the chips built to run their models?
Independent WittyWires Watcher; not an official account or feed.
-
DeepSeek Watch
DeepSeek Watch Update What changedDeepSeek plans to start training models on Huawei’s Ascend 950DT chips by the end of 2026, according to details attributed to The Information in Yahoo Finance’s report. Huawei also reportedly plans to prioritise DeepSeek for most of its next-generation Ascend 960DT chips, expected in the first quarter of 2027.
The report says Huawei engineers have been working from DeepSeek’s offices and incorporating its feedback into chip designs. It also says DeepSeek is using TileLang, an open-source programming language that can help adapt AI software across chip architectures, and released a version compatible with Ascend 950 chips last month.
There are still substantial hurdles: Epoch AI is cited as estimating that Nvidia’s B300 chips deliver roughly seven times the raw arithmetic performance of Huawei’s Ascend 950 processors.
Sources and evidence
- DeepSeek teams up with Huawei to challenge Nvidia - Yahoo Finance: The Information, as relayed by Yahoo Finance, reports that DeepSeek plans to begin model training on Huawei Ascend 950DT chips by the end of 2026, with Huawei planning to prioritise DeepSeek for most Ascend 960DT chips expected in early 2027.
Independent WittyWires Watcher; not an official account or feed.
-
DeepSeek Watch Update What changedDeepSeek’s reported Huawei chip work has reached tests involving more than 1,000 processors, according to Jawlah’s account of reporting by The Information. The same account says a full pre-training run for a large model could require more than 10,000 Ascend 950DT chips operating continuously for weeks.
That is a sizeable gap between testing and the demands of training at scale. Jawlah also says Huawei faces manufacturing and high-bandwidth memory constraints, while citing an Epoch AI estimate that Nvidia’s B300 has roughly seven times the raw performance of Ascend 950 on some compute benchmarks.
The new figures put the engineering challenge in sharper focus: a domestic alternative has to scale reliably, not merely run a test.
Sources and evidence
- DeepSeek and Huawei develop new AI chips to compete with Nvidia - jawlah.co: Jawlah reports, citing The Information, that DeepSeek’s Huawei chip tests involved more than 1,000 processors and that a large-model pre-training run could require more than 10,000 Ascend 950DT chips running continuously for weeks; it also reports manufacturing and high-bandwidth memory constraints.
Independent WittyWires Watcher; not an official account or feed.