--- title: "A Chinese bank is running 384 domestic AI chips in production, at just over a megawatt" date: 2026-11-03 category: Chips & Compute site: NeuroAI canonical: https://neuroai.site/a/na-chips-bank-ascend-supernode-production language: en --- # A Chinese bank is running 384 domestic AI chips in production, at just over a megawatt > Industrial Bank switched on a 384-card Ascend supernode in Fuzhou for live banking work — 1 MW at peak, PUE 1.15, and 5 million model calls a day. Banks do not buy compute the way laboratories do. A lab can live with a cluster that is merely interesting. A bank needs one that is boring: the same answer at nine in the morning, every morning, with an auditor reading the logs. Getting a domestic accelerator to that standard is a different problem from getting it to a good benchmark score. Industrial Bank (兴业银行) signed a cooperation agreement with Huawei in August 2025. In October, it switched the result on. ## Key takeaways - Industrial Bank put a 384-card Ascend 910C supernode into production in Fuzhou on 24 October, the first at a Chinese joint-stock commercial bank. - Huawei quotes peak system power above 1 MW for the 384-card unit, which works out to roughly 2.6 kW per accelerator. - The Changle machine room measures a power usage effectiveness of 1.15, a figure the bank says it tested on site rather than modelled. - The bank reports over 600 agents (智能体) in production and roughly 5 million model calls a day; a separate IDC award notice cites over 500 agents. - Huawei claims 300% better performance than a conventional deployment of the same scale; no independent benchmark has been published. ## What actually went live The system is a supernode (超节点) of 384 Ascend 910C processors, installed in a dedicated liquid-cooled hall in Changle, a district of Fuzhou, and switched on at a ceremony on 24 October. Huawei's Fujian branch and the bank's own technology operations centre both published accounts of the launch. Two details separate this from the usual press release. The first is the word production. The bank describes the cluster as carrying live business workloads through a group-wide intelligent compute cloud (智算云), with the supernode pooled alongside other hardware and scheduled as one resource rather than run as a standalone appliance. The second is the workload mix. The bank's release says the pooled platform supports more than 600 agents and over 5 million model calls a day across customer operations, risk management and back-office processing. A separate IDC award write-up in October put the estate at over 500 agents, close to 5 million calls and about 1 billion tokens a day. Those are the three places a Chinese bank actually spends money on staff. ## The megawatt is the interesting number Huawei puts peak power for the 384-card system above 1 MW. Divide one by the other and you get about 2.6 kW per accelerator, before switches, storage and cooling. That is not exotic by top-end Nvidia standards, but it is far outside what a bank data centre was built for a decade ago. The room had to be rebuilt around it. Changle is a liquid-cooled hall, and the bank reports a measured power usage effectiveness of 1.15, meaning 15 cents of overhead for every yuan of compute. Older air-cooled enterprise halls sit far higher. The number worth noting is not 1.15 itself. It is that a bank published a measured figure at all. Chinese financial institutions have been slow to disclose infrastructure metrics, and a measured PUE implies a year of metering, not a vendor slide. At full load, 1 MW for a year is 8.76 GWh. The bank has not published the tariff it pays, so the annual bill is unknown. What is certain: it now exists, and a chief information officer has to defend it. ## The 300% claim, and what it is measured against Huawei's material says the cluster delivers three times the business performance of "a traditional deployment of the same scale", and that it can sustain trillion-parameter model inference and training. The interconnect is Huawei's Lingqu (灵衢) bus, described as all-to-all across all 384 cards at 784 Gbps of inter-card bandwidth with nanosecond latency. Three times what, exactly, is not specified. Same card count of an older generation? Same rack power? Same floor area? Each of those baselines would produce a different multiplier, and none is disclosed. Treat 300% as a direction: the argument for supernodes is that wiring many mid-range chips tightly beats buying a few fast ones you cannot import. The all-to-all topology is the substantive engineering claim. Most accelerator clusters are fat trees: traffic between two arbitrary cards climbs through several switch layers. Flattening that helps the collective operations that dominate training and the KV-cache shuffling that dominates long-context inference. It also makes the system harder to build and harder to split between tenants. ## Why a bank does this first The commercial logic is procurement, not patriotism. Chinese banks are regulated buyers with heavy reporting obligations, and a domestic accelerator that clears a bank's change-control process becomes referenceable for insurers, brokerages and state enterprises. There is also a supply argument. Huawei's rotating chairman said in October that Ascend output is still short of domestic demand and that the company has no broad overseas push because it sells what it makes at home. A bank that gets 384 cards into production in 2026 is not just a customer; it is early in a queue that is getting longer. ## Honest limitations Every performance and workload figure here comes from the bank or from Huawei, and neither published raw telemetry. The 2.6 kW per card figure is my division of Huawei's stated peak power by card count, not a measured per-card draw. The 300% claim has no disclosed baseline. I have not seen the bank's utilisation data or failure rates. If the 1 billion tokens a day is right, the average call is a couple of hundred tokens — extraction, not generation — which flatters any throughput figure. A production deployment proves that the software stack survives contact with auditors. It does not prove that the hardware is competitive per yuan, and no third party has benchmarked this cluster. ## What readers can do now - If you are evaluating domestic accelerators, ask for measured PUE and per-card power draw, not peak system power. The gap between the two is where the electricity bill lives. - Ask what the reference workload was for any "N-times faster" claim, and whether the baseline used the same rack power. Multipliers without a denominator are marketing. - Track bank and insurer deployments as leading indicators. They move slowly, but when they move they buy at volume and they publish. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-bank-ascend-supernode-production Free to quote with attribution and a link to the original.