Zhipu AI explores custom ASIC chip as GLM-5.2 usage surges 27x - The Information

Investing.com -- China's Zhipu AI is in early discussions with domestic chip design houses about building a bespoke AI processor for its GLM model family, The Information reported, as 27-times growth in daily token usage for its latest model collides with tightening U.S. export restrictions on advanced semiconductors.

The development is a slight headwind for NVIDIA Corporation (NASDAQ: NVDA), whose high-end GPUs have been the default compute layer for Chinese AI labs. Every Chinese developer that successfully transitions inference workloads to a custom application-specific integrated circuit represents a slice of Nvidia's addressable market in China that becomes structurally harder to recover, regardless of how export-control policy evolves.

According to The Information, Zhipu recently made preliminary inquiries with several Chinese chip design houses about the possibility of collaborating on a processor optimized specifically for running its GLM models. The Beijing-based lab has not yet selected a partner, and the conversations remain at an early stage. The whole endeavor, The Information noted, could take more than two years, requiring Zhipu to assemble or expand a semiconductor team, run the chip through design and testing cycles, and rework its software stack to take advantage of the new hardware.

The catalyst is a combination of explosive demand and constrained supply. GLM-5.2 has been the fastest-growing model on Vercel's model aggregator platform since its release last month, with daily token usage surging as much as 27 times during its first week, The Information reported. At the same time, U.S. export controls have made it increasingly difficult for Chinese AI labs to source Nvidia's most capable chips, turning compute availability into a binding constraint rather than simply a cost consideration.

The chips Zhipu is weighing are application-specific integrated circuits, or ASICs, processors engineered to perform specific tasks tied to specific models, as opposed to the general-purpose architecture of Nvidia's GPUs, as The Information explained. ASICs typically deliver superior energy efficiency and lower per-token inference costs once a model's architecture stabilizes, which makes them economically compelling for labs running large-scale, high-volume inference rather than exploratory training runs. Industry analysts have broadly noted that inference-optimized ASICs can cut operating costs for mature workloads by a substantial margin compared with general-purpose GPUs, though exact figures vary by model architecture and utilization rate.

Zhipu would be following a path already traveled by several larger peers. Google, OpenAI, ByteDance, and Alibaba have all developed proprietary custom chips to reduce dependence on outside GPU suppliers and lower the cost of running their own models. Just hours before the news, Reuters reported that DeepSeek is also seeking to develop custom chips to reduce its reliance on Huawei and Nvidia. For Zhipu, the push is notably more urgent given the export-control backdrop: unlike U.S. counterparts that can still access Nvidia's latest silicon, Chinese labs face a regulatory ceiling that makes any domestic alternative, however nascent, strategically valuable.

On the domestic chip-design side, Chinese firms including Cambricon Technologies Corp Ltd (SS:688256) and Biren Technology have been active in the AI ASIC space, though neither has been named by The Information as a prospective partner for Zhipu. The broader ecosystem of domestic designers has expanded meaningfully since the initial rounds of U.S. export restrictions took effect, giving Chinese AI labs more options than existed even two years ago.

For Nvidia investors, the cumulative picture matters more than any single announcement. NVDA shares are currently trading on the Nasdaq as one of the market's most widely held AI infrastructure names, and China has historically represented a meaningful portion of its data-center revenue. Each successive report of a Chinese AI lab exploring custom silicon, even at an early, multi-year horizon, incrementally reinforces the bear case that Nvidia's China data-center exposure faces structural, not merely political, pressure.

The immediate question for Zhipu is execution. Two-plus years is a long runway, and the lab will need to navigate chip design complexity, foundry access, and software adaptation simultaneously, The Information noted. Should GLM demand continue compounding at anything close to the pace suggested by the GLM-5.2 launch numbers, the business case for pushing through those hurdles only strengthens. The next visible milestone will likely be the selection of a chip design partner, a decision that, once announced, would signal the initiative has moved from exploratory to committed.

You May Also Be Interested In

• Alphabet Sees FY26 Capex $195B To $205B, Saw $180B To $190B
• Texas Instruments dips despite strong beat as investors sell the news
• White House accuses Moonshot of violating US export controls using Nvidia hardware