In just one year, four Chinese GPU startups have gone from IPO to billion-dollar revenue. Their reports show: China’s AI accelerator market has definitively become bipolar.
In March–April 2026, four leading Chinese GPU manufacturers published their first annual reports after going public. The figures they showed are not a story about preparing to compete with NVIDIA. It is already a story about real sales, real clusters, and real clients who have already switched to Chinese chips.
MetaX, Moore Threads, Biren Technology, and Iluvatar CoreX — the so-called «Four Little Dragons» of the Chinese GPU market — reported a combined revenue of 5.2 billion yuan (about $750 million) for 2025. Growth ranged from 92% to 243% year-over-year. All four remain unprofitable, but three of the four have reduced losses. The industry has moved from the R&D stage to the stage of commercial scaling.
Numbers that speak for themselves
MetaX (Shanghai) — Revenue 1.64 billion CNY ($239 million); Growth +121%; Loss -789 million CNY (reduced by 44%);
Moore Threads (Shanghai) — Revenue 1.51 billion CNY ($219 million); Growth +243%; Loss -1.024 billion CNY (reduced by 37%);
Biren Technology (Hong Kong) — Revenue 1.03 billion CNY ($145 million); Growth +207%; Loss -874 million CNY (expanded by 14%);
Iluvatar CoreX (Hong Kong) — Revenue 1.03 billion CNY ($145 million); Growth +92%; Loss -438 million CNY (reduced by 32%).
MetaX leads in absolute revenue — 1.64 billion yuan. But Moore Threads shows the most impressive growth: +243% for the year. The company demonstrated that it can not just sell individual GPU cards, but win large contracts for turnkey infrastructure: in March 2026, Moore Threads signed a CNY 660 million ($95.5 million) contract to supply the KUAE intelligent cluster. This is a shift from «selling chips» to «selling AI factories» — a model that NVIDIA has refined on tens of thousands of orders worldwide.
Biren Technology is the only «dragon» whose loss expanded (by 14%). The company attributes this to continued investments in R&D and production scaling. For a startup at the «revenue growth, no profit» stage, this is a normal picture — what matters is that the revenue growth rate (+207%) significantly exceeds the expense growth rate.
Inference — the new main engine
The most important insight from the reports: the main driver of growth is not clusters for training models, but infrastructure for inference — running and serving already trained models.
MetaX recorded an 866% year-over-year growth in inference GPU card sales — far exceeding the growth of universal «training + inference» cards.
Iluvatar reported a 238% revenue growth from its «Zhikai» inference lineup, also significantly outpacing training products.
Biren Technology formulated the reason for this shift in its annual report:
«The proliferation of AI agents creates a multiplier effect for computing demand. One agent call can generate dozens of inference requests, leading to an exponential growth in demand for inference computing power.»
This is a key observation. When AI transitioned from chatbots to autonomous agents, the demand architecture changed: one user request no longer means one model pass. An agent plans, checks, corrects — and each step requires a separate inference.
Result: demand for inference capacity is growing faster than demand for training capacity.
For the market, this means: companies investing in inference infrastructure today are in a more advantageous position than those building training clusters.
Inference is recurring revenue, predictable load, high utilization.
Training is burst capacity, expensive downtime, and dependence on the «someone is training a new model» cycle.
NVIDIA is back — but is the market returning?
At GTC 2026 in March, Jensen Huang announced the resumption of H200 production for Chinese clients. NVIDIA received export licenses, orders are already in progress. Blackwell remains banned — China is getting the second most powerful generation, not the first.
On paper, this should weaken the position of Chinese manufacturers. But the annual report figures say the opposite: when NVIDIA was cut off from the market, Chinese companies did not just «fill the void» — they built a full-fledged customer base and brought products to commercial scale.
The share of Chinese chips in the domestic AI server market reached 41% in 2025:
Huawei (24%);
Cambricon (5%);
Baidu Kunlun (4%);
MetaX (3%);
Iluvatar (3%).
This is no longer a «niche for those who couldn’t get NVIDIA» — it is a structural market share that has taken root.
Global Times quotes analysts: NVIDIA chips have shifted from «must-have» to «optional» status. The H200 will return to a market where it has real competitors running real workloads for real clients. And it will return without Blackwell — the most attractive product.
Manufacturing base: SMIC reaches 5nm
Scaling the Chinese GPU stack is impossible without a manufacturing base. Important events have happened here as well.
SMIC — China’s largest foundry — is conducting pilot runs of a 5nm process (designated N+3). According to TechInsights, N+3 is a «scalable evolution» of the 7nm process and approaches a «true 5nm equivalent» without EUV lithography.
In 2026, SMIC also plans to double its 7nm capacity. In January 2026, SMIC acquired its Beijing branch for $5.79 billion — consolidating control over key production capacities.
Hua Hong — China’s second-largest foundry — has, according to Reuters, developed a 7nm technology suitable for AI chip production. This is a backup manufacturing base in case SMIC cannot cope with growing demand.
It is worth noting: the yield level of Huawei Ascend 910C chips at SMIC is estimated at only 20–40% — significantly below industry standards. But here the «state acceptance» model works: the strategic necessity of obtaining sovereign AI chips outweighs commercial efficiency. China is ready to pay for low process efficiency because the alternative is not to pay at all.
Software-Defined Chips: the bet on bypassing CUDA
In addition to ramping up production capacity, China is making an architectural maneuver. At the Strategic Semiconductor Summit in early April, Tsinghua Professor Wei Shaoqjun presented the concept of Software-Defined Chips (SDC) — an alternative to directly copying NVIDIA’s CUDA ecosystem.
The essence of the approach: instead of a fixed GPU architecture with proprietary software, SDC assumes that the chip is dynamically reconfigured for a specific task. The compiler generates a bitstream directly controlling data flows — and the developer gets control at a level unavailable in traditional GPUs.
This is not an attempt to «build a better CUDA.» It is an attempt to make CUDA unnecessary.
SDC is particularly advantageous for inference, and inference, as we showed above, is becoming the main market for AI accelerators.
For edge devices, industrial IoT, and embedded systems, SDC can provide efficiency unavailable in general-purpose GPUs.
The approach is risky: it requires new development tools, new frameworks, and a new approach to optimization. But if China can create a mature SDC ecosystem, it is not just an alternative to NVIDIA — it is a paradigm shift.
What this means for the Russian market
The annual reports of the «Four Dragons» are not abstract Chinese news. They are a practical benchmark for Russian companies choosing AI infrastructure.
First, inference is what you should plan for. Most Russian enterprises do not train models from scratch — they run and adapt existing ones (DeepSeek, Qwen, Kimi). Inference load is the main scenario. Chinese GPUs, especially lines optimized for inference (MetaX C550/C600, Moore Threads S4000/S5000, Iluvatar BiV150), suit this scenario better than universal NVIDIA cards, which cost 2–3 times more.
Second, the «turnkey cluster» model works. Moore Threads proved this with a $95.5 million contract — not just GPUs, but ready-made infrastructure with software and support.
Third, the manufacturing base is stabilizing. SMIC at 5nm, Hua Hong at 7nm, doubling capacity — this means the next generation of Chinese GPUs (MetaX C700, Moore Threads Lushan/Huashan, expected in 2026–2027) will be produced on more mature processes with better yields. And therefore — a more predictable supply chain and more predictable pricing.
Fourth, 41% is the point of no return. When four companies publish billion-dollar revenues, when inference sales grow by 866%, when SMIC builds 5nm — this is no longer a «temporary replacement for NVIDIA during sanctions.» It is a parallel ecosystem developing by its own laws. Russian companies that connect to it now will enter a mature market. Those who wait for the «return of NVIDIA» risk being left behind.
Contact our experts to select the optimal hardware configuration and calculate rental pricing for your AI workloads.