2 Billion Yuan Raised in Six Months

The STAR Market Daily reported on the 10th that Shanghai Guangyu Xinchen Technology Co., Ltd. had completed a Series A funding round. The company said it has raised a total of 2 billion yuan over the past six months.

Participants in the round included Puquan Capital, BAIC Industrial Investment, Longteng Capital, Puyao Xinye, Dahua Venture Capital, Kaitai Capital, Shuyun Capital, Kaide Innovation, China Life Capital, Shandong Development Investment Group, Xinneng Capital, Qianhai Infrastructure Investment, Huarui Investment, Houwei Capital, Chenxi Capital, Orient Jiafu, Qiaobei Capital and Dioo Microcircuits.

Founded in 2024 and headquartered in Shanghai's Fengxian district, Guangyu Xinchen develops large-model inference chips for end devices, along with supporting hardware and software systems. Founder Zhou Qiang graduated from Fudan University's School of Microelectronics and previously worked on chip R&D, products and ecosystems at VeriSilicon, AMD, Enflame and Moore Threads.

Using Near-Memory Computing to Ease the “Memory Wall”

For smartphones, cars, robots and smart glasses to run large models locally, they must frequently retrieve model parameters from memory in addition to performing calculations. As parameter counts continue to rise, moving data can consume more time and energy than the calculations themselves, creating a “memory wall” that constrains on-device AI performance.

Guangyu Xinchen uses 3D stacking to place memory layers close to the compute chip. It combines this architecture with near-memory computing, streaming and tiered storage to reduce the distance data must travel.

The company aims to improve actual memory-bandwidth and compute-unit utilization during model inference rather than simply pursuing higher peak computing power.

TC1000 Enters Mass-Production Deployment

Guangyu Xinchen's first product is the TC1000. According to figures disclosed by the company, the chip uses a four-layer memory stack. Under specific test conditions, it runs inference on a 3-billion-parameter model at roughly 300 tokens per second and on a 35-billion-parameter model at about 70 tokens per second.

These figures depend heavily on test conditions, including the model, quantization precision, input and output lengths, batch size and power consumption. Testing environments may also differ across vendors, meaning token throughput alone does not fully reflect a chip's performance in real-world devices.

Guangyu Xinchen has also developed EdgeAIon, a full-stack hardware and software system built around the TC1000 that connects the chip, compiler, runtime and model deployment tools. The TC1000 provides the underlying computing and storage capabilities, while EdgeAIon maps model operators onto the chip and helps customers adapt and optimize their models.

For on-device AI chips, the maturity of the software toolchain directly affects how quickly customers can adopt a product. An incomplete toolchain forces customers to devote more engineering resources to model conversion, operator adaptation and troubleshooting, potentially extending the commercialization cycle.

Guangyu Xinchen said the TC1000 has entered mass-production deployment and customer application development and was unveiled during the 2026 World Artificial Intelligence Conference. The company also said the TC1000 succeeded on its first tape-out and achieved a chip yield above expectations.

However, the company has not disclosed details such as wafer batches, cumulative shipments, paying customers or order value.

Advancing Next-Generation Chip Development

Building on its first product, Guangyu Xinchen has begun the tape-out of its second-generation chip and the development of a third-generation product.

As the number of stacked memory layers increases, the next generation must not only improve computing performance but also address heat dissipation, packaging yield, storage capacity and manufacturing costs. In a 3D-stacked architecture, a defect in any layer can reduce the yield of the finished product.

Competitors in the on-device AI chip market include smartphone and automotive SoC vendors, independent AI accelerator companies and device makers developing chips in-house. Guangyu Xinchen must still prove that its architecture's performance advantages for specific models and power ranges are sufficient to offset the cost to customers of switching software platforms and supply chains.

Commercialization Still Requires Further Validation

For Guangyu Xinchen, the TC1000's actual shipment volume, number of paying customers and revenue will be key measures of its commercialization progress.

Other important tests of the company's engineering capabilities include whether its performance results can be reproduced using the same models, power levels and software environments; the yield, packaging cost and thermal performance of its 3D-stacking technology; and whether its second-generation chip completes tape-out on schedule.

Demand for on-device large-model inference is growing, but chip performance, the software ecosystem and supply-chain compatibility remain critical to large-scale adoption. Guangyu Xinchen will use the new funding primarily for further chip development, product iterations and customer application expansion. Its progress will need to be assessed through subsequent shipment, order and financial data.