Inside a modern data centre, most of the energy is spent not on thinking but on shuttling data between memory and processor — the famous von Neumann bottleneck. The brain does not have that problem: it computes where it stores. A Beijing company has spent eight years trying to turn that biological trick into a shipping chip, and in late 2025 it stood up what it calls China's first brain-inspired (类脑) compute cluster.
From a Tsinghua lab to a chip
Lynxi (灵汐科技) was founded in 2018, growing out of Tsinghua University's brain-inspired computing centre (类脑计算研究中心). Its flagship is the KA200 neuromorphic chip, first released in 2021 and later put into mass production. The architecture is deliberately different from a GPU: it fuses memory and computation, runs many cores in parallel, and supports both deep neural networks and spiking neural networks (脉冲神经网络, SNN) — the latter being the closer imitation of how biological neurons fire.
The company says a credit-card-sized module delivers around 32 TOPS of compute, and that its designs target the four hallmarks of brain-like hardware: in-memory computing, many-core parallelism, sparse computation, and event-driven operation.
A Nature cover, then a cluster
Lynxi's credibility got a lift in May 2024 when research on a brain-inspired complementary visual chip — developed with Tsinghua professors Shi Luping (施路平) and Zhao Rong (赵蓉) — landed on the cover of Nature. The work demonstrated a chip that blends precision computing with brain-like efficiency for open-world perception, a sign that "brain-inspired" is more than a marketing adjective.
The bigger commercial move came on 11 December 2025, when Lynxi, with China Telecom's Hangzhou unit and partners, launched what it calls the country's first brain-inspired (类脑) intelligent-computing cluster for large-model inference. The company says the cluster exceeds 10 billion neurons in scale and more than 200 POPS of compute, and that it drives single-user large-model inference latency down to the millisecond range while cutting power use by roughly half to two-thirds versus comparable conventional inference. Those cluster figures are company disclosures from the launch announcement.
Why brain-like chips matter for the brain
Neuromorphic computing is not brain-computer interface (脑机接口) in the implant sense, but the two share a foundation. Spiking neural networks (脉冲神经网络, SNN) and event-driven hardware are the natural fit for the sparse, asynchronous spikes that real neurons produce. A cheap, efficient spiking chip is exactly what you want sitting next to an electrode array, decoding neural signals in real time without a power-hungry GPU. Lynxi explicitly markets its gear for brain-science research and next-generation AI algorithm development, and its edge products target robots, drones, and cameras — places where power and latency, not raw FLOPs, decide what is possible.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the strategic bet behind neuromorphic (类脑) silicon is efficiency under real-world power budgets, not benchmark-topping peaks.
The power problem real AI keeps hitting
The reason brain-like design keeps returning is mundane: moving data costs energy. Every time a processor fetches a weight from memory, it spends more than the arithmetic itself. Data centres feel this as electricity bills and heat; a drone or a wearable feels it as dead batteries. Biological brains avoid the trap by storing and computing in the same wetware, which is why a human brain runs on roughly the power of a dim light bulb while outperforming racks of servers at perception. Neuromorphic (类脑) hardware is the engineering attempt to copy that trick — sparse, event-driven, in-memory. Lynxi's pitch is not that it beats a GPU on a benchmark, but that for the long tail of always-on, power-bound devices, the brain's architecture is the better starting point.
The honest competitive picture
The field Lynxi plays in is crowded and global. Intel's Loihi, IBM's TrueNorth, and a wave of academic spiking chips all pursue the same goal. Lynxi's edge is claimed to be heterogeneous fusion — running both DNN and SNN on one chip and staying compatible with PyTorch and TensorFlow — which lowers the adoption friction that killed earlier neuromorphic efforts. Whether that translates into volume outside niche research and edge deployments is still unproven.
Honest limitations
Lynxi's cluster scale (10 billion-plus neurons, 200-plus POPS) and efficiency claims (half to two-thirds power cut, millisecond latency) come from the company's own launch materials, not independent benchmarks. The KA200's 32 TOPS figure is a vendor specification. The 2024 Nature cover confirms genuine research credibility but describes a specific visual chip, not the commercial product line. Neuromorphic (类脑) computing remains a small share of deployed AI infrastructure, and real-time neural decoding on Lynxi hardware is a stated application, not a shipped, approved medical function. Figures here are qualitative-to-vendor-grade; treat them as directional.
What readers can do now
- Engineers building edge AI for robots or drones should evaluate spiking or neuromorphic (类脑) accelerators where power and latency dominate cost — not just GPU-class parts.
- Neuroscience labs can use brain-inspired (类脑) hardware as a low-power substrate for real-time decoding experiments, a cheaper alternative to GPU racks.
- Readers should separate genuine heterogeneous (DNN plus SNN) fusion from pure-marketing "brain-like" labels when judging neuromorphic (类脑) vendors.