Chinese AI Labs Develop Own Chips for Synergy, Efficiency

Chinese artificial intelligence laboratories are increasingly developing their own chips. This trend reflects a global shift towards integrating software and hardware, though industry sources warn of substantial upfront investment risks. The objective is to achieve greater hardware-software synergy and reduce long-term operating expenses.

Photo: Shutterstock Images

Paul Triolo, a DGA-Albright Stonebridge Group technology policy lead, stated in AIStackDecrypted that proprietary efforts show China's leading model developers view silicon as a strategic extension of the model stack. DeepSeek, a Hangzhou based start up, has hired chip design talent without public job postings, Reuters reported. Its customised AI inference chip plans began approximately a year ago.

So, Zhipu AI, a Beijing based GLM 5.2 model developer, has engaged in initial discussions with domestic chip design companies for tailored AI processors, The Information reported. This follows a sharp rise in the organisation's daily token usage.

The chip design goals emerge during increased computing demand, rising inference costs, and restricted access to Nvidia's advanced AI chips. Arisa Liu, Taiwan Industry Economics Services' chief director, noted custom silicon permits hardware optimisation tailored to specific architectures, like the DeepSeek R1 model, even with external suppliers.

Paul Triolo added the aim is not to replace Nvidia or even Huawei, but to reduce quickly increasing AI inference costs as token volumes surge. Hui He, Omdia semiconductor research director, called software hardware integration an inevitable trend, noting customised chips improve efficiency and build competitive advantage.

US Big Tech companies have pursued custom or semi custom AI chips to reduce reliance on standard processors and allow hardware customisation. JP Morgan estimated in June that shipments of AI application specific integrated circuits, ASICs, and XPUs to major US hyperscalers would surpass GPU shipments from Nvidia and AMD by 2027.

But custom AI chips do not necessarily exceed Nvidia's GPUs in power across all workloads. JP Morgan reported they offer better performance per USD and per watt when designed around a hyperscaler's software stack and data centre architecture. Such projects require large, stable workloads and high volume deployments to justify their cost.

Model developers' efforts do not always need a fully internal chip design unit, Omdia's He observed. Organisations could establish chip teams, invest in chip start ups, or co develop processors with specialist designers. Companies decide when to tape out based on expected deployment scale, future shipment volumes, and whether those volumes cover the initial high cost before fabrication.

However, Nvidia chips will remain necessary for general purpose AI computing, He explained, calling its GPUs a token generator for the industry. And custom chips are more likely for proprietary or dedicated workloads, as AI companies seek varied computing power sources amid ongoing supply limitations.

DeepSeek closed a significant fundraising round in June and has aggressively expanded its team. Its V4 model is extensively deployed for inference on Huawei's Ascend 950 chips. China's major cloud service providers have also directed substantial resources into chip units.

So, as computing power demand surges and access to advanced foreign semiconductors stays limited, Chinese AI companies increasingly source processors from domestic chip designers. Wei Shaojun, China Semiconductor Industry Association vice president and Dongfang Suanxin chief executive, stated that domestic chip and algorithm integration is inevitable. Wei's company collaborates with model developers to build large models based on domestic chips, avoiding efficiency losses from adapting foreign processors.

Paul Triolo observed that US export controls have prompted the Chinese AI ecosystem to co develop and optimise solutions across several areas simultaneously, including manufacturing, hardware integration, software integration, model development, and harness or application development.

But analysts also warned this shift from software to hardware carries significant risks. The substantial upfront investment might require years to yield positive returns. Arisa Liu stated that even if a chip successfully replaces external hardware and significantly reduces electricity and computing costs, it generally takes two to three years of mass production to recoup high research and development costs. She explained that for AI software companies such as DeepSeek, the initial financial and time investment required for in house chip development represents a considerable gamble, and the recruitment of scarce talent in chip design and wafer engineering could also be demanding.

• Chinese AI laboratories are developing proprietary chips to achieve hardware software synergy and lower operating costs.

• DeepSeek and Zhipu AI are among the Chinese firms pursuing customised AI processors.

• This trend is driven by surging computing demand, rising inference costs, and restricted access to advanced foreign AI chips.

• US Big Tech companies have also pursued custom AI chips to reduce reliance on standard processors.

• Analysts warn that the pivot from software to hardware involves significant upfront investment risks and long payback periods.

Source: SCMP