Written by: Agata Wojtas, Chief Commercial Officer, Digital Colliers
Huawei moving its Ascend 960DT launch forward to Q1 2027 is not just a competitive response to Nvidia. It's a signal that the US-China AI supply chain is splitting in earnest. For mid-market banks running AI in production, the problem is immediate. You cannot model the risk of a vendor relationship when you do not know which chip ecosystem sits underneath. If your transaction monitoring or credit decisioning relies on a service whose inference runs on hardware subject to export controls, you need a data model that maps the dependency now. The bifurcation is not theoretical. It is a procurement reality that will hit your vendor conversations in the next twelve months.
Why the chip layer determines operational resilience
Banks buy AI as a service. You do not specify the silicon. But the silicon determines your operational resilience. A vendor running inference on Chinese-origin chips faces different supply-chain exposure than one running on US-origin hardware. Export controls can turn a stable vendor relationship into a single point of failure overnight.
The pattern we keep seeing is that teams treat AI vendors as software providers when the underlying dependency is hardware. Your AML false-positive rate might sit at 85-95% in a typical mid-market setup, but if the model that was supposed to improve that stops working because the vendor lost chip access, you are back to manual review backlogs with no path forward. The software contract does not protect you from hardware supply-chain failure.
What to map in your vendor stack
The data model needs four layers. First, vendor stack. Which AI services sit in your production environment. List the vendor, the service, and what business process it touches. If it is AML monitoring, credit scoring, fraud detection, or customer service automation, note it.
Second, chip dependencies. Which inference hardware powers each service. Does the vendor run on Nvidia A100s, H100s, or on Huawei Ascend chips. If they claim to be chip-agnostic, ask for the actual deployment configuration. Most vendors are not agnostic in practice.
Third, geographic exposure. Where the chips are designed, manufactured, and where the vendor's data centers operate. A vendor running US chips in European data centers carries different risk than one running Chinese chips in Asian data centers. Both scenarios need distinct mitigation plans.
Fourth, fallback options. If the primary vendor loses access to its hardware supply chain, what alternatives exist and how long would migration take. Can you revert to the legacy system. Do you have a secondary vendor under contract. How much re-integration work does the switch require. Most teams do not ask these questions during procurement because they assume AI is fungible. It is not.
The regulatory window is tightening
DORA has been in force since 17 January 2025. It requires financial entities to manage ICT third-party risk and maintain operational continuity. If your AI vendor cannot deliver because export controls severed its chip supply, that is an operational resilience failure under DORA. You need documentation showing you understood the dependency and had a mitigation plan ready.
The EU AI Act high-risk obligations apply from 2 December 2027. Credit scoring and transaction monitoring both fall into the high-risk category. Fines reach up to €15M or 3% of global turnover. You cannot wait until late 2027 to map chip dependencies when Huawei's roadmap acceleration tells you the bifurcation is happening now. The regulatory asks are coming. The geopolitical split is already here.
Build the dependency map in 2026
Start with your production AI services. For each one, document the vendor, the inference infrastructure, and the chip family in a spreadsheet. Then map the supply-chain lineage. US-designed chips fabricated in Taiwan carry different risk than Chinese-designed chips fabricated domestically. Neither is zero risk. Both need fallback plans documented with owners and timelines.
The winning operators are building this map in 2026 because they know export policy moves faster than vendor migration timelines. If you are running a mid-market bank with AI in production, the chip-layer split is not a geopolitical curiosity. It is an operational dependency you need visible in your risk register with mitigation plans that have been tested. Huawei accelerating its roadmap is the signal to act. Do not wait for the export control announcement before you discover which vendors cannot deliver.