appConfigEntityId

bodyType

customContents

entityId

entityUuid

isVideo

articlePageQuery

article

articleSchemaContent

content

hooksArticlePageAdvertisementConfigArticle

hooksArticlePageFloatingButtonConfigArticle

appConfig

filter

json

liveArticle

Wency Chen

Ann Cao

Chinese AI labs are increasingly pursuing proprietary chips, mirroring a global trend towards software-hardware integration, but industry insiders and analysts warn that the strategy carries risk due to the massive upfront investments required.
“The core motivation for choosing in-house chips lies in pursuing [greater] hardware-software synergy and lowering long-term operating costs,” said Arisa Liu, chief director and research fellow at Taiwan Industry Economics Services.
Paul Triolo, a partner and technology policy lead at DGA-Albright Stonebridge Group, said in his personal newsletter AIStackDecrypted that the proprietary efforts underscored how China’s leading model developers increasingly viewed silicon as a strategic extension of the model stack rather than simply another infrastructure input.
DeepSeek has been quietly hiring chip-design talent without posting public job openings, according to two people familiar with the situation, who declined to be named as the matter was private.
The Hangzhou-based start-up’s plans for a customised AI inference chip began roughly a year ago and remained at an early stage, according to a Reuters report on Tuesday.
Meanwhile, Zhipu AI, the Beijing-based developer of the high-performance GLM-5.2 model, was in early talks with domestic chip-design companies about tailored AI processors amid a sharp increase in its daily token usage, The Information reported on Tuesday.
The chip design aspirations of Chinese model makers come amid surging computing demand, rising inference costs and continued restricted access in China to Nvidia’s most advanced AI chips.
While processors from external suppliers like Huawei Technologies were available, Liu said custom silicon allowed for hardware optimisation tailored to specific architectures, such as the DeepSeek-R1 model.

Triolo added that the goal was not to replace Nvidia or even Huawei, but to reduce the rapidly rising cost of AI inference as token volumes explode and enterprise deployments accelerate.
“Software-hardware integration is an inevitable trend,” said Hui He, semiconductor research director at Omdia. “For model companies, running their own models on customised chips can improve efficiency and help build a moat around their ecosystem.”
US tech giants Google, Amazon, Microsoft and OpenAI have all pursued custom or semi-custom AI chips to reduce reliance on off-the-shelf processors and tailor hardware to their own workloads.
JP Morgan estimated in June that shipments of AI application-specific integrated circuits, or ASICs, and XPUs to major US hyperscalers would exceed graphics processing unit (GPU) shipments from Nvidia and AMD by 2027, accounting for 53 per cent of AI accelerator units.
Custom AI chips are not necessarily more powerful than Nvidia’s GPUs across all workloads, but JP Morgan said they offered better performance per dollar and per watt when designed around a hyperscaler’s own software stack and data-centre architecture. Such projects require large, stable workloads and high-volume deployments to justify the cost.
Model developers’ efforts do not always have to take the form of a fully internal chip-design unit, Omdia’s He said. They could also build chip teams, invest in chip start-ups or co-develop processors with specialist designers, sharing the results of such projects.
Still, building in-house-design capabilities did not necessarily mean a company had already taped out a chip, she said. Companies have to decide whether and when to tape out based on their expected deployment scale, future shipment volumes and whether those volumes can cover the high upfront cost. Tape out is the final stage of chip design when it is sent to the foundry for fabrication.

However, Nvidia chips would still be necessary for general-purpose AI computing, He added, describing its GPUs as a “token generator” for the industry. Custom chips, by contrast, are more likely to be used for proprietary or dedicated workloads, as AI companies seek multiple sources of computing power amid persistent supply constraints.
The trend is also expected to benefit ASIC design partners and related supply-chain companies, including Broadcom, Marvell and MediaTek, while also becoming a key driver of Taiwan Semiconductor Manufacturing Co’s CoWoS advanced-packaging expansion and high-bandwidth memory demand.
DeepSeek, which closed a 51 billion yuan (US$7.5 billion) fundraising round in June that valued it at around US$60 billion, has been aggressively expanding its team. Its latest V4 model has been widely deployed for inference on Huawei’s Ascend 950 chips.
China’s major cloud service providers, including Alibaba Group Holding, Baidu and ByteDance, have also been pouring resources into chip units.
At the same time, as demand for computing power surges and access to advanced foreign semiconductors remains constrained, Chinese AI companies are increasingly sourcing processors from domestic chip designers such as Cambricon, Iluvatar CoreX and MetaX to diversify their supply. Alibaba owns the South China Morning Post.
“The integration of domestic chips and domestic algorithms or models is inevitable,” said Wei Shaojun, vice-president of the China Semiconductor Industry Association and founder and chief executive of Dongfang Suanxin, a Shanghai-based chip-design unicorn.
Wei said his company was working with model developers to build large models “based on domestic chips,” rather than developing them first on foreign processors and then adapting them to Chinese hardware – a process he said inevitably resulted in efficiency losses.
Triolo said US export controls had pushed the Chinese AI ecosystem “to co-develop and optimise solutions across the manufacturing, hardware integration, software integration, model development, and harness or application development stack all at once in parallel.”
Meanwhile, analysts also warned that the pivot from software to hardware carried immense risks, as the heavy upfront investment might take years to generate positive returns.
Even if a chip successfully replaced external hardware and significantly lowered electricity and computing costs, it generally took two to three years of mass production to recoup the high R&D costs, Liu said.
“For AI software companies such as DeepSeek, the initial financial and time investment required to develop in-house chips is undoubtedly a huge gamble,” Liu said. The recruitment of scarce talent in chip design and wafer engineering could be intensive, she added.
Additional reporting by Howard Liu and Coco Feng


Philippines tribe vows to stand ground against US-led Pax Silica: ‘listen to us’

As the US weighs restrictions on cloud computing, how will China’s AI sector adapt?

Chinese AI chipmaker Hygon plots expansion from data centres to robotics

China’s CXMT touts new chip manufacturing platform as close to ‘world’s most advanced’

Semiconductors

Chinese AI labs pursue custom chips to lower costs but heavy upfront investment a risk

Custom AI chips offer better performance per dollar and per watt when designed around a hyperscaler’s own software stack and data-centre architecture. Photo: Shutterstock Images

DeepSeek has been quietly hiring chip-design talent without posting public job openings. Photo: Shutterstock Images

Zhipu AI is reportedly in early talks with domestic chip-design companies about tailored AI processors. Photo: CFOTO/Future Publishing via Getty Images

China weighs open-weight AI’s security risks against national tech innovation strategy: researchers

Person holding mobile phone with logo of Zhipu Huazhang Technology on screen in front of web page. The company recently marked a breakthrough with its GLM-5.2 model. Photo: Shutterstock Images

Anthropic hits back after China warns of Claude Code ‘backdoor’ risks

Anthropic says Claude Code is not officially available in China, despite the coding tool’s popularity among Chinese developers. Photo: Reuters

Tech

Tech Trends

Tech War

News

China Future Tech

Custom silicon allows for hardware optimisation tailored to specific architectures, such as the DeepSeek-R1 model


Custom silicon allows for hardware optimisation tailored to specific architectures, such as the DeepSeek-R1 model.


Chinese AI labs pursue custom chips to lower costs but heavy upfront investment a risk

.css-1c6uqr6{color:inherit;font-weight:inherit;font-size:inherit;font-family:inherit;line-height:inherit;overflow-wrap:break-word;}Custom silicon allows for hardware optimisation tailored to specific architectures, such as the DeepSeek-R1 model

Custom silicon allows for hardware optimisation tailored to specific architectures, such as the DeepSeek-R1 model