Three events that once seemed like separate points have converged into a single trend: the Chinese GPU ecosystem has transitioned from the «we assemble hardware» phase to the «we compete on development speed.» For the Russian market—where Chinese GPUs are already the main alternative supply channel—this means increased availability, stack maturity, and… new challenges.
Amidst global changes in the technology and artificial intelligence markets, Chinese companies Moore Threads and Biren Tech have taken a bold step forward by completing the Day-0 adaptation of the Zhipu AI GLM-5.2 model. This event not only highlights their technical capabilities but also testifies to the maturity of the ecosystem, which is now capable of competing with Western analogues. In a context where access to modern NVIDIA graphics processors for Chinese companies is limited, such adaptation becomes a key factor for the successful implementation of AI solutions. Day-0 adaptation, as a standard, allows companies to quickly integrate new models and minimize the time for their deployment into production processes.
Moore Threads’ financial performance for the first quarter of 2026 confirms their status as one of the market leaders: revenue reached 7.38 billion yuan, which is 243% higher than the previous year. This not only demonstrates the company’s growth but also instills confidence in the stability and reliability of their solutions for clients. Meanwhile, Enflame Technology, having received approval for an IPO, plans to raise 60 billion yuan, which will accelerate the development of new generations of AI chips. These financial injections create new opportunities for Russian companies, which can take advantage of more accessible and competitive GPU solutions, especially amidst rising rental prices for NVIDIA B200.
Moore Threads and Biren Tech: Day-0 adaptation of GLM-5.2 + Q1 finances
Two of the largest Chinese GPU vendors—Moore Threads (摩尔线程) and Biren Tech (壁仞科技)—simultaneously announced the completion of the Day-0 adaptation of the flagship Zhipu AI GLM-5.2 model. This is no coincidence: in China, Day-0 has become a competitive weapon.
Technically important:
GLM-5.2 is the new generation from Zhipu (annual revenue ~6 billion yuan in 2025, IPO in Hong Kong — March 2026). The context window is noticeably wider than GLM-5.1, which loads the prefill stage.
Moore Threads MTT S5000 (80 GB HBM3): optimization for prefill and P/D separation was already implemented for GLM-5.1. For 5.2, throughput on long sequences was increased.
MUSA SDK + MUSIFY (auto-translator from CUDA to MUSA) allows transferring open-source models without manually rewriting kernels.
Biren Tech Biren 166 Series—a competing architecture—also passed Day-0. It is a race: two leaders of the «Four Dragons» demonstrate stack maturity simultaneously.
This is about marketing:
NVIDIA H100/H200 has been unavailable to China since 2023. Any major open-source model is released under CUDA. For a Chinese GPU to be in demand, adaptation must be faster than the customer gets disappointed.
When two vendors show Day-0 on the same model, it is a trend, not an exception. MUSA and BRUCLE ecosystems have reached integration speeds comparable to CUDA. Not in absolute core performance, but in time-to-production.
On June 19, Moore Threads published its Q1 2026 report: revenue of 7.38 billion yuan (243% year-over-year growth), with 450,000+ developers in the MUSA ecosystem. This is no longer a startup—it is a scale at which Day-0 becomes an operational norm, not a PR unit.
Enflame CloudBlazer: IPO on STAR Market and 60 billion yuan
On June 15, Enflame Technology (燧原科技) received IPO approval on the Shanghai STAR Market. Expected fundraising: 60 billion yuan (~8.3 billion USD). Funds will go towards:
5th generation AI chips (CloudBlazer);
6th generation AI chips;
Software-hardware co-design programs.
Enflame is part of the «Four Little Dragons» of the Chinese GPU industry (Moore Threads, MetaX, Birentech, Enflame).
Why this is a market signal:
2026 became the year of IPOs for Chinese GPU vendors. Moore Threads went public on STAR Market in October 2025 (688795SH). Enflame follows. MetaX and Birentech are preparing applications.
Public funding means accelerated R&D: 60 billion yuan is approximately 2-3 years of capex for architecture and software. For comparison: NVIDIA spends about $8-10 billion annually on R&D. Enflame is an order of magnitude smaller, but capital inflow linearly correlates with the speed of new architecture releases.
Enflame expects 1H 2026 revenue to exceed the entire 2025. This confirms demand from Chinese hyperscalers and cloud providers.
For a Russian buyer: IPO vendors undergo strict disclosure audits (CSRC regulations). This increases the transparency of the product roadmap and reduces the «black box» risk in procurement.
NVIDIA B200: rental prices to double, orders postponed to Q2 2027
On June 18, literally the next day, an answer came from Silicon Valley: Baseten (an AI inference platform) at the Altimeter event reported that the rental cost of B200 GPUs will double starting from October 2026. New orders are postponed to the second quarter of 2027.
Context:
B200 is the next generation after H200. Production is limited by the CoWoS packaging shortage from TSMC. NVIDIA prioritizes hyperscalers (Microsoft, Google, Amazon, Meta), while the secondary cloud rental market remains without volume.
The rise in B200 rental prices creates a new price floor for GPU as a Service. If B200 rental in the USA rises to $7-9/hour (estimate), alternative GPUs (including Chinese) become competitive at $2-4/hour on comparable inference workloads.
For Russia, the effect is amplified: Russian companies cannot buy H100/B200 directly anyway. High rental prices through intermediaries make Chinese GPU clusters the economically optimal choice for LLM inference.
EU sanctions, 47 entities under new restrictions
On the same day Enflame received IPO approval, the European Union introduced a new sanctions package against Russia: 34 individuals and 47 legal entities. Among the restrictions is the tightening of control over dual-use technologies and equipment, including microelectronics.
Practically, this means:
GPU transit through third countries (including former Central Asian republics, UAE, Turkey) is under intensified monitoring. The EU blocks not just direct supplies but also gray schemes through proxies.
Chinese GPUs do not fall under direct US EAR (if produced in China without US IP), but the supply route is complicated. It requires heavier loads on logistics.
On the other hand: tightening Western sanctions only accelerates import substitution in Russia. Chinese GPUs become not an alternative, but the «main channel» for enterprise AI infrastructure.
How this will affect the Russian market
Ecosystem maturity. Day-0 adaptation of GLM-5.2 by two vendors (Moore Threads and Biren Tech) simultaneously is not PR, it is a trend. Chinese GPUs support modern open-source LLMs with the same integration timelines as NVIDIA. For Russian integrators (banks, ministries, telecom), this reduces vendor lock-in and TCO.
Ecosystem scale confirmed. Q1 2026 Moore Threads: 7.38 billion yuan revenue, 450,000+ developers. This means the MUSA SDK has sufficient critical mass for the rapid adaptation of new models without dependence on NVIDIA.
Funding equals scale. Enflame’s IPO and Moore Threads’ already completed entry to STAR Market mean that Chinese GPU vendors are receiving public investments to accelerate roadmaps. Expect new architectures every 12-18 months.
Price window for the Russian market. The rise in B200 rental prices makes Chinese GPU clusters competitive not only due to sanctions but also purely economically—approximately 40-60% cheaper in TCO on inference.
Logistics risks. EU sanctions on June 15 and tightened BIS control complicate transit. A 3-6 month procurement horizon and reliable supply chain partners with cross-border experience are required.
ARKA as an agnostic platform. The growth in the number of Chinese GPU vendors strengthens the need for a unified platform. The localized version of MoArk ARKA supports 7+ GPU vendors and allows running inference, fine-tuning, and pipelines without being tied to specific hardware. Amidst the growing fragmentation of the GPU market, this becomes critically important.
Conclusions
The combination of events—from the synchronous Day-0 adaptation of the GLM-5.2 model by Moore Threads and Biren Tech vendors to the successful IPO of Enflame Technology and the impressive quarterly report of Moore Threads—finally forms a new industry landscape. The Chinese graphics processor industry has made a qualitative leap, moving from simple copying and hardware assembly to direct competition in the speed of bringing innovations to market. This is no longer the isolated success of individual companies, but a systemic shift where the speed of software-hardware integration becomes a key competitive advantage, comparable in importance to the raw performance of chips. For the global market, this is a signal of a new force emerging, and for Russia—a confirmation that the alternative technology stack has reached the necessary maturity to support critical loads.
For the Russian market, where Chinese GPUs have de facto become the main supply channel, this transition opens a window of strategic opportunities but simultaneously presents new challenges for business and the state. The growth in availability and stack maturity reduces dependence on Western vendors and the total cost of ownership (TCO) of infrastructure, but requires a revision of planning approaches. The complication of logistics due to tightening Western sanctions dictates the need to form long-term partnerships and create equipment reserves. In these conditions, the key success factor becomes the implementation of universal agnostic platforms like ARKA, which allow abstracting from a specific hardware manufacturer and ensuring flexibility amidst growing market fragmentation. Thus, the task of import substitution is transforming: it is no longer just about finding a replacement, but about building a sovereign, scalable, and adaptive AI ecosystem capable of quickly absorbing innovations that now originate not from the West, but from the East.
Contact our experts to select the optimal hardware configuration and calculate rental pricing for your AI workloads.