China is exploring a different kind of AI hardware
In August 2025, Zhejiang University introduced Darwin Monkey, a computer built from 960 specialised chips capable of simulating more than two billion artificial neurons.
The announcement stood out at a time when much of the AI industry was focused on increasingly powerful GPUs and the computing infrastructure required for large language models.
But Darwin Monkey was not designed simply to compete with conventional AI servers.
It emerged from research into neuromorphic computing, an approach that borrows selected principles from biological neural systems.
In conventional AI computing, processors perform large numbers of numerical calculations. Some brain-inspired systems instead use artificial neurons that communicate through brief signals called spikes.
Researchers are investigating whether these different approaches can improve certain kinds of computing, particularly where energy consumption, responsiveness or hardware resources matter.
Zhejiang University is not alone in this work. Tsinghua University and other Chinese institutions have also developed experimental processors and sensing chips that approach AI computation differently.
Together, their research reveals another direction in China’s AI development, one that focuses on the hardware underlying intelligent systems rather than only the models running on it.
Zhejiang University: From Darwin3 to Darwin Monkey
Darwin Monkey is part of Zhejiang University’s longer-running Darwin research programme.
In 2024, researchers from Zhejiang University and Zhejiang Lab published Darwin3, a chip designed to simulate networks of artificial neurons.
The chip supports a large number of simulated neurons and includes mechanisms for communication and learning within these networks.
A year later, the university introduced Darwin Monkey, connecting 960 Darwin3 chips across 15 servers.
According to the institution, the resulting system can simulate approximately 2.25 billion neurons and more than 200 billion connections, with typical power consumption of around 2,000 watts.
The numbers describe the scale of the research platform. They do not mean Darwin Monkey has the intelligence of a human brain, and the reported power consumption does not independently establish an advantage over GPUs.
How researchers are using it
Zhejiang University reports using Darwin Monkey to simulate neural networks representing aspects of several animal nervous systems, including worms, zebrafish, mice and macaques.
Such work may help scientists investigate how large networks of neurons behave.
The university also reports running a brain-inspired implementation of a DeepSeek-related language model.
This is particularly relevant because it shows researchers exploring whether their architecture can support AI workloads beyond neuroscience simulations.
The performance of those demonstrations has not been independently established against equivalent conventional systems.
Nevertheless, Darwin Monkey represents a substantial engineering achievement: connecting hundreds of specialised chips into a working platform for investigating alternative forms of computation.
Sources: Zhejiang University, 2025 | Zhejiang University Darwin series, 2026 | Darwin3 research, 2024
Accessible text alternative for this figure
A diagram in three steps. On the left, one Darwin3 chip. In the middle, 960 chips linked across 15 servers. On the right, the institution-reported figures: 2.25 billion simulated neurons and about 2,000 watts of typical system power. The diagram is schematic, not a wiring plan.
Tsinghua University: From an autonomous bicycle to intelligent vision
While Zhejiang University has concentrated on large brain-inspired computing systems, researchers at Tsinghua University have explored how alternative chip designs can support machines interacting with their surroundings.
Tianjic: An AI-powered bicycle
In 2019, Tsinghua researchers introduced Tianjic, a chip designed to support different types of neural-network computation.
To demonstrate it, they built an autonomous bicycle capable of balancing, recognizing objects, following targets, responding to voice commands and avoiding obstacles.
These activities require different AI functions.
The experiment showed that Tianjic could support several of them within one physical system.
It was a research demonstration, not evidence of artificial general intelligence or a commercially available replacement for conventional processors.
Its contribution was the integration of different AI methods on a single experimental architecture.
Accessible text alternative for this figure
A bicycle drawing with one AI chip at its centre, linked to five functions shown in the 2019 research: balance, object recognition, target following, voice commands and obstacle avoidance.
Tianmouc: A different approach to visual sensing
In 2024, another Tsinghua research team introduced Tianmouc, a chip inspired by aspects of biological vision.
The researchers were investigating a practical problem.
A moving vehicle needs detailed information about its surroundings, but it must also respond quickly when something changes.
Continuously processing large amounts of visual information creates demands on computing resources and data transmission.
Tianmouc uses two complementary pathways. One captures detailed visual information, while the other responds rapidly to changes in a scene.
The researchers reported high sensing speeds, improved handling of challenging lighting conditions and reduced data transmission in their experiments.
They also tested the chip in an experimental vehicle-perception system.
However, the reported advantages attracted scientific criticism.
In June 2025, independent researchers published a technical comment in Nature, questioning aspects of Tianmouc’s novelty and performance comparisons with existing vision sensors.
The criticism does not establish that the chip failed to operate. It means that claims of superiority require further evidence.
Tianmouc remains a useful example of Chinese research into specialised AI hardware, particularly for machines that must interpret rapidly changing environments.
Original research: Nature, 2024 | Independent technical comment, 2025
Accessible text alternative for this figure
Two panels of the same street. The left panel shows the detailed scene. The right panel highlights only what changes and moves. This is a conceptual explanation, not actual Tianmouc sensor output.
New research into more efficient learning
Chinese researchers are also investigating how brain-inspired hardware can learn and adapt.
In March 2026, researchers associated with Pengcheng Laboratory, Peking University and collaborating institutions published a computing architecture designed to train spiking neural networks.
The team tested its design using configurable hardware rather than a finished commercial processor.
Their experiments demonstrated training across 20 processing cores and a collaborative-learning arrangement involving five participants.
The researchers reported performance and memory-use advantages under selected experimental comparisons.
These results do not establish that the architecture is generally superior to GPUs, but they demonstrate continued research into making brain-inspired systems more practical.
One possible application is AI operating on devices that have limited computing resources and cannot always depend on a remote data centre.
The research remains experimental, and its wider advantages require further evaluation.
Source: Nature Communications, 2026
What these chips could change in how AI systems are built
The Chinese projects examined here address different problems.
Darwin Monkey investigates large-scale neural simulations and selected AI workloads. Tianjic explores combining different computing methods. Tianmouc focuses on visual sensing, while the newer Pengcheng research examines efficient learning.
Their achievements therefore cannot be compared through a single specification.
A system with billions of simulated neurons is not necessarily more capable than one designed to process visual information quickly.
Likewise, a processor consuming less power during one experiment may not provide an advantage for another type of calculation.
The international NeuroBench initiative, published in Nature Communications in 2025, highlights the need to compare neuromorphic systems using equivalent tasks, accuracy requirements and energy measurements.
That remains an important challenge for this field.
The evidence examined here establishes working Chinese research prototypes and experimental systems, but not widespread commercial deployment or general superiority over conventional AI processors.
GPUs will remain relevant to the article because they provide a familiar reference point, not because the Chinese projects have demonstrated that they can replace them.
Source: NeuroBench, Nature Communications, 2025
China’s contribution to alternative AI computing
China’s research institutions are exploring several approaches to computing beyond the architectures commonly associated with modern AI.
Zhejiang University has developed specialised chips and connected them into a large experimental platform.
Tsinghua University has demonstrated processors that combine AI functions and sensors that handle visual information differently.
More recent research is investigating ways to make brain-inspired systems learn efficiently.
The results remain largely experimental, and their commercial value has yet to be established.
But the direction is worth understanding.
Much of today’s AI development focuses on improving models through additional data and computing resources. Brain-inspired hardware research asks a different question: whether certain tasks could be performed more effectively by changing how information is processed in the first place.
For applications involving sensors, robotics and devices operating with limited power, that could eventually become valuable.
China has not demonstrated a universal successor to the GPU. Its researchers are developing and testing alternative computing approaches that may prove useful for particular AI applications.
Their practical importance will depend on how well those systems perform beyond the research laboratory.
Sources and further reading
- Zhejiang University, Darwin Monkey announcement, 2025
- Zhejiang University, Darwin series and Darwin Monkey specifications, 2026
- Darwin3: A Large-Scale Neuromorphic Chip with a Novel ISA and On-Chip Learning, 2024
- Towards Artificial General Intelligence with Hybrid Tianjic Chip Architecture, Nature, 2019
- A Vision Chip with Complementary Pathways for Open-World Sensing, Nature, 2024
- Dynamic Range and Precision of Hybrid Vision Sensors, Nature, 2025
- A Highly Energy-Efficient Multi-Core Neuromorphic Architecture for Training Deep Spiking Neural Networks, Nature Communications, 2026
- The NeuroBench framework for benchmarking neuromorphic computing algorithms and systems, Nature Communications, 2025