内容来源：https://aibusiness.com/generative-ai/alibaba-aims-independence-new-ai-chips-model

内容总结：

阿里发布自研AI芯片“贞摩M890” 加速摆脱英伟达依赖

3月19日，在阿里云峰会上，中国科技巨头阿里巴巴正式推出全新AI芯片“贞摩M890”，成为又一家致力于降低对英伟达GPU依赖的中国企业。该芯片由阿里旗下半导体公司平头哥自主研发，是一款集训练与推理于一体的加速器，专为AI智能体设计，针对长上下文窗口的巨大内存需求及多模型协作进行了优化。

一同发布的还有升级版通义千问3.7-Max模型。该模型可在M890芯片上连续运行长达35小时，支持超过1000次工具调用，具备百万级Token的上下文窗口，能够处理复杂的多文件代码编辑、重构及原型设计等任务。

分析人士指出，此举不仅是中国云服务商寻求自主可控、降低对美国芯片依赖的战略体现，更是一种成本节约策略。Omdia分析师苏联杰表示：“在中国，所有超大规模云厂商都在谋求更大程度的独立与自给自足。”他强调，使用自研芯片运行AI负载，可减少对外部投资的依赖，并带来更高的定制化与优化空间。

尽管如此，M890的推出仍面临挑战。苏联杰指出，中国芯片供应链整体弱于全球水平，若完全依赖台积电代工，将不得不与英伟达正面竞争。此外，与全球竞品相比，自研芯片的效率可能偏低，长期来看，自研带来的成本节省未必能转化为整体收益。

值得注意的是，阿里发布M890的时间点恰好是谷歌在其I/O 2026开发者大会上推出第八代TPU的次日，凸显了中国AI厂商加速独立发展的决心。

中文翻译：

由谷歌云赞助
选择您的首个生成式AI应用场景
要开始使用生成式AI，首先应聚焦于能够改善人类与信息交互体验的领域。

这款芯片突显了该厂商向全栈AI战略的转型，以及其摆脱对英伟达AI芯片依赖的努力。

中国科技巨头阿里巴巴于周三推出了一款新型AI芯片，成为最新一家致力于减少对英伟达GPU依赖的中国厂商。

在阿里云峰会上亮相的"镇岳M890"是一款训推一体加速器，由该厂商的半导体部门平头哥设计。据阿里巴巴称，这款新芯片专为AI智能体打造。该处理器针对长上下文窗口的巨大内存需求以及多个AI模型间的相互通信进行了优化。该厂商在发布该芯片的同时，还推出了升级版Qwen 3.7-Max模型，该模型可在M890上持续运行长达35小时。

据阿里巴巴介绍，Qwen 3.7-Max能够长时间进行连续推理，并处理超过1000次工具调用。它专为管理复杂、多文件的代码编辑、重构和原型设计而设计，拥有100万token的超大上下文窗口。

这款AI芯片与模型的发布，展示了中国AI厂商尝试自主创新、减少对英伟达GPU依赖的另一条路径。尽管美国目前允许向中国出口英伟达H200等先进AI处理器，但越来越多中国企业正转向使用包括阿里巴巴、百度、华为在内的本土厂商的芯片。

Omdia（Informa TechTarget旗下机构）分析师苏廉杰表示："在中国，所有超大规模云厂商都有一个重大计划，即实现更高程度的独立自主。因此，这更多是一个关于自主自强的故事。"

苏廉杰指出，对中国AI厂商及其企业客户而言，本土芯片生产不仅能实现自主，还能成为一项成本节约策略。他表示："如果能在自有芯片上运行这些工作负载，就能减少在其他方面的投资需求。"

对阿里巴巴而言，这一举措既是差异化战略，也是迈向全栈化的一步——这与百度等中国竞争对手以及AWS、谷歌等美国科技巨头的做法类似。虽然该厂商此前已有芯片产品，但M890及其前代芯片"玄铁C950"均聚焦于智能体AI，表明阿里巴巴认真对待客户使用其芯片并为此进行定制化开发。

苏廉杰分析道："如果他们能在自有芯片上运行工作负载，就能获得更多定制和优化的机会。这将使他们在性能上显著优于其他超大规模云厂商。"

不过M890的发布时间点颇为有趣——阿里巴巴发布新模型的前一天，谷歌刚在其I/O 2026开发者大会上推出了第八代TPU。

尽管如此，苏廉杰表示阿里巴巴仍面临障碍，尤其是供应链制约。他指出："中国芯片供应链整体弱于国际或全球水平。"

阿里巴巴面临的另一挑战是：若完全依赖台积电，就必须与英伟达竞争。目前，该厂商及其阿里云等子公司主要转向使用自有AI芯片。

苏廉杰总结道："与全球竞争对手相比，他们的芯片效率将相对较低。这意味着从长远来看，自研芯片所节省的成本未必能有效转化为整体节约。"

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
The chip highlights the vendor’s move toward a full-stack AI strategy and its efforts to wean itself off Nvidia AI chips.
Chinese tech giant Alibaba on Wednesday introduced a new AI chip, becoming the latest Chinese vendor to push to reduce its dependence on Nvidia GPUs.
Unveiled at the Alibaba Cloud Summit, Zhenwu M890 is a training-and-inference-integrated accelerator designed by the Chinese vendor’s semiconductor unit, T-Head. The new chip is built for AI agents, according to Alibaba. The processor is optimized for the enormous memory demands of long context windows and for multiple AI models communicating with one another. The vendor unveiled the chip alongside its updated Qwen 3.7-Max model, designed to run on the M890 for up to 35 hours.
Qwen 3.7-Max can perform continuous reasoning for a long time and handle over 1,000 tool calls, according to Alibaba. It’s designed to manage complex, multi-file code editing, refactoring and prototyping. It has a large 1-million token context window.
The release of the AI chip and model shows another way Chinese AI vendors are trying to venture out on their own and be less dependent on Nvidia GPUs. Although the U.S. is currently allowing exports of advanced AI processors such as the Nvidia H200 to China, more Chinese firms are relying on chips from local vendors, including Alibaba, Baidu and Huawei.
“Within China, this is a major plan among all the hyperscalers to be a lot more independent, to be a lot more self-sufficient,” said Lian Jye Su, an analyst at Omdia, a division of Informa TechTarget. “So, it's more of a self-sufficient, independent story.”
For Chinese AI vendors and their enterprise customers, not only does locally producing chips enable independence, but that route also becomes a cost-saving strategy, Su said.
“If you were able to run these workloads on your own chipset, it reduces your need to invest elsewhere,” he said.
For Alibaba, this move is both a differentiation strategy and a step toward a full-stack approach, similar to Chinese competitors such as Baidu and U.S. tech giants such as AWS and Google. While the vendor already has chips, M890 and its predecessor chip, XuanTie C950, focus on agentic AI, showing that Alibaba is serious about customers using its chips and customizing them for that purpose.
“If they can run their workloads on their own chipset, which gives them a lot more opportunity to do customization and optimization. That will give them a lot better performance compared to other hyperscalers,” Su said.
The timing of M890 is interesting, though, because Alibaba released the new model a day after Google launched its eighth-generation TPU at its I/O 2026 developer conference.
Despite this, Alibaba still faces hurdles, most notably supply chain constraints, Su said.
“China’s chipset supply chain overall is weaker than the international one or the global one,” Su said.
Also presenting a challenge for Alibaba is the issue of having to fight Nvidia if it were to rely solely on Taiwan Semiconductor Manufacturing Company. Currently, the vendor and its subsidiaries, such as Alibaba Cloud, mostly turn to its own AI chips.
“Their chipset efficiency will be rather poor as compared to global competition,” Su said. “Which means in the longer run, the savings that they incur from having their own chipset may not necessarily translate well into the overall savings.”