Russia's flagship AI model GigaChat was developed by Sberbankthe country's largest bank that is driving Russia's expansion in artificial intelligence.
Russia hopes its national AI chat will be powered by Chinese chips because Western sanctions continue to block the country's access to advanced equipment abroad. “We hope that we will be able to use Chinese microprocessors in GigaChat,” German Gref, president of Sberbank, said in an interview on state-run Channel One.
Russian GidaChat – in appearance it is not much different from Western AI chats
|
GigaChat
GigaChat – allows you to talk via Vkontakte
|
neiro.unisender.com
Gref did not specify which Chinese arrangements Sberbank is interested in, but he is the most likely candidate Huawei Ascend 950the most advanced Chinese AI chip, which, however, is still inferior to the American H200 chip from Nvidia.
The bank's efforts to buy advanced chips from China, however, face stiff competition as China's biggest internet companies also rush to place chip orders Huawei Ascend 950 AI.
Huawei Ascend 950
|
Own materials
Sberbank's schedule undoubtedly constitutes a problem for Huawei, which must first fulfill huge orders from ByteDance, Alibaba and Tencent. ByteDance itself allocated $5.6 billion. on orders for the Ascend 950PR earlier this year. Huawei plans to sell 750,000 950PR units in 2026.But production at the SMIC plant is limited by poor performance of the N+2 DUV process (7nm equivalent) and an estimated eight-month cycle from the start of wafer production to processor completion.
Huawei Ascend 950 – the most advanced Chinese AI chip
|
Huawei.com
The Ascend 950PR falls between the Nvidia H100 and H200 models in inference performance and outperforms the limited H20 by a factor of 2.8, although this value is difficult to verify because Hopper-era hardware does not support native FP4 support. Despite this, every chip produced by Huawei has huge domestic demand, so it is subject to sanctions on Western chips, a Russian buyer would have to compete for allocation with companies that collectively constitute the backbone of China's network economy.
Sberbank launched GigaChat Ultra in March with a new inference mode, and the family of base models has grown over the past year with GigaChat 2.0 and GigaChat Max.
GigaChat MAX – home page
|
neiro.unisender.com
Launching these models at scale requires both inference and training hardware. Model Ascend 950PR is optimized for inferencespecifically the prefill stage of the LLM model. However, its counterpart specialized in training Huawei 950DTis scheduled to hit the market only in Q4 2026 and features 144GB of Huawei's patented HiZQ 2.0 memory with a bandwidth of 4TB/s. Sberbank's current infrastructure relies on a combination of hoarded Western GPUs, Chinese alternatives, and domestic Russian production that has yet to achieve competitive performance for pioneering AI workloads. If Sberbank wants an all-Chinese AI computing stack for GigaChat, will need both systems in large quantities.
AIDOL – the fall of a Russian anthropomorphic robot during a presentation
|
Own materials
In January, Sberbank acquired 41.9 percent. shares in Element, the largest Russian electronics manufacturer, for 27 billion rubles ($356 million). Element produces chips and semiconductors that make up about half of Russia's microelectronics output, but its production is focused on defense and industrial applications, not AI accelerators for data centers.
Sberbank – logo
|
sberbank.ru
Russia's most advanced chipmaker aims to introduce 65nm lithography by 2030, a delay of about 25 years in relation to the latest technologies.
AIDOL – fall of a Russian anthropomorphic robot during a presentation, 11/13/25.
|
AOL.com
Signed on Wednesday Putin and Xi's joint declaration called for closer bilateral cooperation in the field of AI and supported China's proposal to create a global AI governing body. Whether this will translate into actual chip sales for the sanctioned country, which is fighting for a place in Huawei's already bursting at the seams order book, remains unknown for now.
