It is worth taking time out here to give you a glimpse of the emerging national vision behind China’s cooperative AI ecosystem. GEEK ALERT! Rest assured: I have tried hard to reduce the technical jargon in what follows to an absolute minimum... though it will help if you are a neurosurgeon!
If you do not sense the enormity of the next few paragraphs, please read again until you understand how critically important my tech-rich diversion is. Until the Chinese AI vision is pictured in its entirety, its AI ecosystem cannot be fully understood.
China is currently putting in place the equivalent of an “AI body” centred on a national brain – or brAIn! The hardware element of this brain is composed of chips being purposed to interact with the other functions of China’s AI body. Each of China’s largest hardware players are contributing in different but complementary ways to the construction of the composite whole. And, to cap it all, Chinese open-source software – led by Qwen and DeepSeek – is giving life to that brain and body.
China’s AI brain is primarily made up of chips from Huawei, Cambricon and Biren.
Huawei: These chips can be likened to the brain’s hippocampus aided by its cerebral cortex, focused on performing the core functional needs of China’s AI brain. These chips are especially tasked with the formation, organisation and storing of data flows as well as converting shorter-term insights into longer-term memories that eventually coalesce into “wisdom”.
It is in these chips that the key functions of the brain’s cortex are performed: perception, sensory interpretation, memory, reasoning, thought, language, consciousness and increasingly the interpretation of emotion. These chips support both learning (training) and decision-making (inference). That is why these chips are located where centralised training and inference takes place, where superclusters (like the Atlas 950 SuperCluster with 500,000+ chips) are built and where cloud-to-Edge-to-cloud coordination is masterminded.
Examples include the Ascend 910C → Ascend 950 series, designed to learn and decide while constantly taking in fresh inputs from the rest of the body.
These fresh inputs permit upgrades to the core’s knowledge bank as aggregate inflows are fashioned and refined into “wisdom”. Being energy-intensive, much of the brain’s hippocampus and cortex are located in data centres adjacent to the low-cost electricity available from Inner Mongolia’s solar farms. This is one of the main locations where China IS indeed building data centres.
Cambricon: These chips are focused on the Edge – this refers to running AI on or near where the data is generated (phones, cameras, cars, factory gateways, retail kiosks), instead of in distant cloud data centres.
Think of the Cambricon chips as performing the brain functions of the cerebellum and its synapses – handling fast, local decisions and keeping the system in balance. Cambricon’s chips are compact and “elegant” in design, yet powerful in ability, facilitating lightweight inference on the Edge in platforms like phones, drones and robots.
Examples include the Siyuan 690, which is built for efficient routing and lightweight inference; can run Qwen-Agent locally (Qwen-Agent is an open-source framework from the Qwen/Alibaba team for building AI “agents” on top of Qwen LLMs), so some apps don’t need to “phone home”.
Biren: Think of these chips as performing the brain functions of the thalamus and its neural relay. The chips will act as the central relay station of the network, acting as a gatekeeper by filtering and prioritising information flows. The neural relay acts as a switching station or communication hub for those flows. This is where the chip’s raw power counts the most. Because of their ultra-high-performance capabilities, Biren chips facilitate computing for finance, stock exchange trading, government and research.
The main chip already in use here is the BR104 which competes with Nvidia’s imported H100 and B200; China has inventory of the BR104 supplemented by fast-rising domestic production that will cover its needs… until next year!
An upgrade to the BR104 – the BR200 – is expected in 2026: this will compete directly with Nvidia’s Blackwell B200. Biren, not Huawei, is shaping up to be Nvidia’s most feared competitor.
The BR200 will be open-weight and surpass the Blackwell in every category (teraflops, memory bandwidth, power draw, on-device latency, energy efficiency) but one: its process node will be 5 nanometres whereas the Blackwell’s will be 4 nanometres. This means Biren will trade one generation of node shrinkage for superior system-level design in every other metric.
After the BR200 upgrade, and in particular after the introduction of BIRENSA v3.0, which is the “missing link” in China’s computer chip capabilities, there will be no chip that is specifically designed and purposed for the Chinese AI hardware network that cannot be produced in China. At that point, China will not be held hostage to US intellectual property, Taiwan’s TSMC manufacturing expertise or the Netherlands’ ASML EUV knowhow... China will then be self-sufficient in every part of the chip supply chain.
China’s other chip producers either supply the Big Three above or provide chips for ancillary “bodily functions”:
- SMIC wafer chips perform the heart function of the whole system, generating the lifeblood the network body needs by ensuring the regular flow of data round the entire system. Produced domestically, SMIC supplies chip wafers with 7nm & 7nm+ nodes to the three abovementioned companies: Huawei, Cambricon and Biren;
- YMTC chips perform the lung function in that they supply the oxygen of data flows to the network. The chips allow high-speed storage for AI workloads and feed data to Huawei’s Ascend chips in a manner that permits low-latency training and inference, resulting in better speed in computation and response;
- Hua Hong services facilitate the neuromuscular junction whereby the thoughts emanating from the core are translated into the actions carried out by the rest of the AI body. It does this through integration or “packaging”: arranging chips made by others in such a way that they perform effectively. Hua Hong’s various technologies – 2.5D/3D packaging, FOWLP, MCM, BEoL – are the silent engines behind China’s AI hardware ecosystem;
- Huawei’s optical UB-Mesh nationwide wiring network acts as the nervous system connecting all the above systems;
- ModelScope, OpenXLab and other open data and software repositories provide the memory archive of the ecosystem;
- The National Data Administration in Beijing oversees the whole system, virus protection, other required immunities plus overall data governance; and
- The food, the energy that is needed to drive the whole “body”? That comes from the electricity increasingly provided from solar, wind and hydro sources.
The result? In China’s AI ecosystem, cooperation is hardwired into its DNA, with the result being a capability where the whole is much greater than the sum of its parts. There is no equivalent vision to this proto-brain and body AI network in the US.
Here ends the tech-rich interlude. DM
*Read part one here and part three here.