With Western semiconductor exports effectively locked off, Russia's state-backed banking giant Sberbank is reportedly in talks to source Chinese-made chips to sustain and expand its homegrown large language model, GigaChat. The move underscores just how profoundly export controls have reshaped the global AI hardware landscape — and how quickly alternative supply chains are forming in response.
GigaChat launched in 2023 as Russia's answer to ChatGPT, positioning Sberbank as an unlikely AI competitor on the world stage. But without access to NVIDIA's high-end data center GPUs — the industry's dominant training hardware — scaling any frontier model becomes an engineering puzzle wrapped in a geopolitical dilemma. Chinese chipmakers like Huawei, with its Ascend AI accelerator line, have been positioning themselves as exactly the kind of fallback option Russia now needs.
This development carries real industry significance beyond the Russia-China bilateral angle. It signals that the AI race is fragmenting into distinct hardware ecosystems — one anchored by NVIDIA and its Western partners, another slowly coalescing around Chinese alternatives. For years, the assumption was that non-NVIDIA AI infrastructure simply couldn't compete on performance. Huawei and other Chinese vendors are working hard to close that gap, and deals like this one give them both revenue and real-world deployment data to improve their stacks.
For Western policymakers, the lesson is uncomfortable: export controls slow adversaries down, but they also accelerate the development of competing supply chains. Every restricted sale is an incentive for China to invest more aggressively in domestic chip capability. Sberbank's pivot to Chinese silicon may be a preview of a bifurcated AI world where geography, not just technical merit, determines which hardware powers which models — and ultimately, which nations lead in AI deployment.
Whether Chinese chips can handle GigaChat's ambitions at scale remains an open question, but the willingness to try tells you everything about where the pressure points in global AI infrastructure actually lie.