AI model developers are increasingly moving to secure their own AI chips, a push that appears to be becoming mainstream rather than a niche effort.

Major tech companies across the United States and China, including Google, Amazon Web Services (AWS), ByteDance and Alibaba, have been developing custom AI chips to reduce reliance on outside suppliers and cut costs. More recently, leading AI model companies have also begun to join the race to develop AI chips in earnest.

ㆍ[Tech Inside] Selling like Nvidia... Big cloud firms' 'AI chip war' enters a new phase

OpenAI recently co-developed an AI inference server chip called Jalapeno with Broadcom. It is part of a strategy to reduce dependence on Nvidia chips and secure its own hardware. It is part of OpenAI's strategy to reduce dependence on Nvidia chips and secure its own hardware.

OpenAI plans to bring its first Jalapeno servers online at year-end and then expand its use of the chip. Unlike Nvidia's flagship GPUs, which handle both training and inference, Jalapeno is designed only for inference. OpenAI said early tests showed Jalapeno delivered much higher performance per watt than products that are "currently state-of-the-art".

OpenAI is also developing custom server racks equipped with Jalapeno and Broadcom networking gear. For this, it is working with Toronto-based data center equipment design services company Celestica.

ㆍOpenAI unveils in-house AI inference server chip... first operation at year-end

Signs have also emerged that Anthropic is interested in developing its own chip. The Information reported in early July, citing sources, that Anthropic has begun early-stage work to develop an in-house AI chip and is also discussing a manufacturing partnership with Samsung Electronics.

Anthropic is discussing what role the chip should play, what performance level it should target, and how it would be incorporated into servers or server clusters. The Information said Anthropic has discussed the project with multiple chip design companies but has not yet moved to detailed design, testing or manufacturing.

ㆍAnthropic also reviews in-house AI chip development... discusses partnership with Samsung Electronics

Chinese AI startups are also showing interest in developing their own chips. China-based AI company DeepSeek has already begun in-house development of an inference AI chip, according to reports.

According to Reuters, DeepSeek is reviewing development of an inference chip used to run models after training.

ㆍDeepSeek pushes in-house AI chip development… reduces reliance on Nvidia and Huawei

It is part of a strategy to lower dependence on external chip suppliers such as Huawei and Nvidia. DeepSeek is already in talks with manufacturing partners and has reportedly begun hiring engineers to support the development.

Z.ai, which has recently expanded its presence in the global AI arena with the open-source AI model GLM5.2, is also reviewing in-house AI chip design as surging demand and U.S. export controls have made it harder to secure computing resources.

According to The Information, Z.ai recently asked several Chinese chip designers about the possibility of developing a custom AI processor optimized to run its models. Discussions are at an early stage and no design partner has been selected yet, The Information said.

Z.ai has used a mix of Huawei chips, other Chinese-made chips and some Nvidia chips. In January, it unveiled an image-generation model trained using only Huawei chips. It was the first major image model trained solely on Chinese-made chips.

But there are concerns about limitations of Huawei chips. The Information said U.S. export controls have left Chinese semiconductor foundries continuing to struggle to secure advanced equipment needed to produce Huawei chips.

The U.S. Commerce Department placed Z.ai on a technology procurement ban list. The Information said if Z.ai moves ahead with an AI chip development project, it is likely to produce the chips at a Chinese foundry.

But being blacklisted does not mean U.S. companies cannot use Z.ai models. Coinbase CEO Brian Armstrong recently said on X he is testing GLM-5.2 as part of efforts to cut costs. GLM-5.2 is also available on Oracle Cloud.