For decades, artificial intelligence (AI) has pursued scale: more parameters, more data, more GPUs. That approach delivered rapid breakthroughs but also revealed hard limits in power consumption, cost, and accessibility. In China, a different path is taking shape, one that looks beyond silicon scaling laws to the human brain itself.
At a recent forum in Shanghai, Chinese researchers presented what they see as a step toward "next-gen AI": brain-inspired large models trained entirely on domestic computing platforms, alongside brain-computer interface (BCI) systems already moving into clinical trials. Taken together, the advances signal a broader ambition: to rethink AI around energy efficiency, temporal intelligence, and tighter integration with human cognition, as TMTPost, Chinastarmarket.cn, and East Money reported.
Why the brain matters more than bigger GPUs
The case for brain-inspired AI is stark in its simplicity.
Training GPT-3, with 175 billion parameters, reportedly required around 1,000 GPUs and roughly 300 kilowatts of power. The human brain, by contrast, operates tens of billions of neurons and vastly more synaptic connections on about 20 watts — less than a household light bulb.
That efficiency gap is becoming impossible to ignore. As AI models scale, rising energy demand, hardware constraints, and geopolitical frictions around advanced GPUs are turning compute into a strategic bottleneck. Brain-inspired computing, long seen as experimental, is now being reconsidered as a potential way out.
It is against this backdrop that the Tianqiao and Chrissy Chen Institute (China) has launched its Spiking Intelligence Lab (SIL), led by Li Guoqi of the Institute of Automation, Chinese Academy of Sciences.
Shunxi 1.0: A brain-inspired large model on domestic computing
At the center of the announcement is Shunxi 1.0, described as China's first large-scale spiking, brain-inspired model.
Unlike mainstream Transformer-based models, Shunxi 1.0 draws on neuroscience principles such as spike-based communication, event-driven computation, and neural dynamics that encode information over time rather than static tokens. The goal is not only lower power consumption, but also improved handling of long sequences and generalization.
Researchers say the open-source 7-billion-parameter version of Shunxi 1.0 achieved roughly 90% of the performance of Alibaba's Qwen-7B while using less than 2% of the pre-training data typically required. More notably, both training and inference ran entirely on domestic GPU platforms, without reliance on foreign accelerators.
The significance goes beyond benchmarks. It suggests advanced AI research in China can continue even as export controls tighten and that architectural innovation, not just access to cutting-edge hardware, can drive progress.
Closing the stack: From models to chips
Turning brain-inspired AI into something practical requires more than software. Spiking neural networks behave very differently from conventional deep-learning workloads, demanding new hardware and system-level designs.
Shunxi 1.0 was developed in close collaboration with MetaX, a domestic GPU company, to align model design with local compute capabilities. Researchers describe the work as a full-stack effort, linking brain-inspired base models, domestic GPU platforms, and future neuromorphic chips into a single pipeline.
The long-term goal is co-design: models that naturally map onto low-power, event-driven hardware, potentially pushing inference power consumption down to milliwatt levels for certain tasks. If achieved, it would mark a sharp break from today's energy-hungry AI infrastructure.
Brain-computer interfaces leave the lab
If brain-inspired models rethink how machines compute, brain-computer interfaces test how directly AI can interact with the human nervous system.
Shanghai has emerged as one of China's most active BCI testing grounds. In a recent clinical milestone, BrainTiger Technology, incubated by the Tianqiao and Chrissy Chen Institute, completed a clinical trial at Huashan Hospital using a fully implanted, wireless brain-computer interface.
The system, described as "fully implanted, fully wireless, and fully functional," allowed a patient with complete paralysis below the shoulders to control a cursor, browse the web, and play videos using only neural signals. The recorded decoding rate reached 5.2 bits per second, close to international front-runner performance.
A key design choice was implanting the battery module under the chest skin rather than in the skull, following established deep-brain stimulation (DBS) pathways. This reduces thermal risk near the brain and improves long-term maintainability, underscoring how clinical realities are shaping engineering decisions.
When AI and neuroscience start shaping each other
What stands out in China's recent progress is not individual breakthroughs, but the feedback loop forming between AI and neuroscience.
Advanced AI models are improving the decoding of noisy brain signals, accelerating BCI development. At the same time, growing insight into neural efficiency and temporal processing is pushing AI researchers away from brute-force scaling and toward biologically inspired architectures.
Clinicians involved in the projects point to a shift in innovation dynamics. Instead of research flowing one way from lab to hospital, real clinical needs are now shaping model design, algorithms, and hardware requirements, with engineers, neuroscientists, and doctors increasingly working side by side.
What this means for Western AI and BCI players
Globally, companies such as Neuralink, Synchron, and Blackrock Neurotech remain leaders in invasive brain-computer interfaces, while US and European labs dominate large-model research. China's trajectory, however, points to a different competitive angle.
Rather than chasing ever-larger models on the most advanced GPUs, Chinese researchers are betting on energy efficiency, architectural change, and system-level integration, areas where latecomers can potentially leapfrog incumbents.
For Western rivals, this raises uncomfortable questions. If brain-inspired models mature faster than expected, the advantage of massive GPU clusters could erode. And if BCI systems begin integrating tightly with low-power, brain-like AI, the boundary between human and machine intelligence may shift in ways today's AI stacks are ill-equipped to handle.
A parallel track to "next-gen AI"
Brain-inspired AI and brain-computer interfaces remain far from mainstream deployment, with hurdles ranging from scalability to reliability and ethical oversight. Still, China's recent progress suggests these technologies are no longer confined to theory or isolated demonstrations.
Instead, they are forming a parallel track to conventional AI, one that favors efficiency over scale, integration over abstraction, and biology over brute force. Whether this path converges with or disrupts today's dominant AI paradigm remains an open question.
What is clear is that the race for "AI's next generation" is no longer defined by who has the biggest models or the most GPUs. It is increasingly about who understands intelligence itself.
Article edited by Jack Wu
