The Chinese AI infrastructure ecosystem is experiencing perhaps its busiest week in months. Three events that occurred almost simultaneously paint a picture of the rapid closing of the final technological gaps. If we used to talk about «chip substitution,» we are now talking about substituting the entire chain — from silicon to research platforms and major investment deals.
60,000 accelerators and not a single American chip
On April 14th, Sugon (中科曙光), affiliated with the Chinese Academy of Sciences, launched the country’s largest computational cluster for scientific AI (AI4S — AI for Science) in Zhengzhou (Henan Province). The cluster comprises 60,000 accelerators — its power has doubled from 30,000 in just two months.
The entire stack — chips, high-speed scaleFabric network, and software — is entirely of Chinese origin.
The cluster’s purpose is to accelerate fundamental research: protein structure prediction, new materials modeling, and aerospace engineering (simultaneous turbulence simulation on trillions of atmospheric grids). In parallel, Sugon launched the OneScience platform, China’s first integrated environment for AI4S experiments with dozens of pre-installed scientific models and datasets that requires no programming skills.
This is not just a scale record. It is a demonstration that Chinese scientific-grade computational clusters can exist entirely outside the American technology base.
For context: The analogous American project, the Genesis Mission by the Department of Energy, is still in the deployment phase.3 источника
Tencent and Alibaba invest in DeepSeek
On April 22nd, Reuters and Bloomberg reported simultaneously: Tencent and Alibaba are negotiating investments in DeepSeek at a valuation of over $20 billion.
Tencent has proposed acquiring up to 20% of the startup, although the company itself, according to sources, is not keen on giving away such a large stake.
This is a landmark event for several reasons:
Consolidation: Two of China’s largest IT conglomerates are competing for a share in a company that was a little-known hedge fund project just a year ago. DeepSeek has grown into a key player thanks to its V3 and V4 models, the latter of which (1 trillion parameters) runs on Huawei Ascend 950PR without a single NVIDIA chip.
Vertical integration: Tencent and Alibaba are simultaneously investing in AI startups and ramping up their own chip divisions. Alibaba, through Pingtouge (平头哥), has already shipped over 470,000 chips, and the division’s revenue has reached hundreds of billions of yuan on an annualized basis.
Market signal: A $20 billion valuation for a company that didn’t exist two years ago shows that the Chinese AI market no longer copies a working model — it is building its own.
Moore Threads architecture and benchmarks at world-leading levels
Last week, Moore Threads (摩尔线程) held a large-scale presentation where it revealed its full technological roadmap:
The new «Huagang» (花港) architecture — support for full precision from FP4 to FP64, 50% higher compute density, and 10x efficiency over the previous generation. The architecture is oriented towards both inference and training.
MTT S5000 in AI4S: The card has passed full validation on protein structure prediction tasks (Protenix from ByteDance) and genomic modeling (Evo 2). Result: training performance at 115% of the level of international GPUs in protein prediction tasks, inference accuracy at the level of world-class counterparts.
AI-laptop MTT AIBOOK with SoC (长江) — 50 TOPS of edge performance, an integrated development environment out of the box, support for Linux/Windows/Android.
In parallel, IDC published annual data: in 2025, the share of Chinese GPUs in the domestic market reached 41% (up from 30% in 2024). Huawei (昇腾) holds 20% of the market (812,000 units), Alibaba Pingtouge — 6.6% (265,000 units), while NVIDIA’s share has dropped to 55%. According to Bernstein’s forecast, in 2026 NVIDIA may lose another 47 percentage points — down to 8%.
What this means for the Russian market
Russian enterprises evaluating Chinese GPU solutions receive three important signals:
The stack’s maturity has passed the critical mass threshold. When Sugon launches a sixty-thousand accelerator cluster without a single American chip, and Moore Threads shows 115% performance in real scientific tasks — these are no longer «Chinese alternatives,» but full-fledged platforms. Russian customers are no longer «early adopters» — they are joining an ecosystem that has passed stress-testing in production.
Tencent/Alibaba’s investments in DeepSeek accelerate the creation of a closed AI loop. Models + chips + cloud + models — this loop is closing inside China. For Russian companies, this means: supplies of Chinese GPUs are becoming more sustainable because demand within China is already generating its own growth momentum. Deficit risks are decreasing.
Competition between the three «camps» of Chinese chipmakers works in the buyer’s favor. Huawei bets on vertical integration and the public sector, Alibaba/Baidu — on cloud-chip synergistic effects, and specialized companies (Moore Threads, Iluvatar, Cambricon) — on open ecosystems and a broad customer base. Russian customers get a real choice: from «turnkey» to «free integration.»
One nuance: Bernstein forecasts NVIDIA’s share in China will shrink to 8% by 2026. This means Chinese GPU manufacturers will be increasingly focused on the domestic market. Russians should act now while the production capacities of Chinese companies are available for export contracts.
Contact our experts to select the optimal hardware configuration and calculate rental pricing for your AI workloads.