Stanford has published the AI Index 2026, and the figure that should have been the main headline went almost unnoticed. The gap between the best American and Chinese models is just 2.7%. Stanford itself calls this a «de facto closed race.» But at the very moment China caught up with America in model quality, it began to move away from the open weights and dependencies that led to this. Meanwhile, a US presidential advisor warned: Huawei will soon start selling its AI chips worldwide.
Stanford records parity — and immediately closes the book
The multi-page Stanford HAI AI Index 2026 report was released on Monday, April 14. The key metric — the Arena Leaderboard (model ranking based on blind user votes): the best American model (Claude Opus 4.6 by Anthropic) scored 1,503 points, the best Chinese model (Dola-Seed-2.0 Preview by ByteDance) scored 1,464 points. The difference: 39 Elo points, or 2.7%.
A year ago, the gap was five points. A year before that — dozens. On any benchmark, the order of models swaps places: GPQA Diamond (graduate exam) — 93% accuracy, SWE-bench (real programming) — models have reached the human baseline, OSWorld (desktop agents) — 66.3%, six points from human.
The top 6 labs on Arena fit within 79 Elo points. Four American, two Chinese. Stanford states it directly: «Competition on model quality is no longer a differentiator. The race has shifted to cost, reliability, and practical utility.»
What this means for the market: if model quality has equalized, the choice of infrastructure — GPU, platform, supplier — becomes a strategic decision not about who is better, but about who is more accessible, reliable, and sovereign.
Huawei will start exporting chips worldwide
David Sacks — President Trump’s advisor on AI and cryptocurrencies — stated in an interview that Huawei «will soon start exporting its AI chips to the global market.» He noted that the technology gap between the US and China is rapidly narrowing.
The statement came against the backdrop of another signal: the BIS (Bureau of Industry and Security) continues to stall — the bureau has lost 20% of its licensing staff, and the issuance of export permits for NVIDIA H200 to allies has stalled. The paradox: America wants to maximize chip exports but cannot physically process applications.
If Huawei indeed brings Ascend to the global market, it changes the landscape for all countries currently queuing for NVIDIA and unable to get chips due to BIS bureaucracy. Russia, the Middle East, Southeast Asia, and Africa — all are looking for alternatives, and Huawei could become the first large-scale «non-NVIDIA» solution for the enterprise segment.
However, caution is warranted. Ascend still lags behind the NVIDIA H200 in performance for training large models, and the CANN ecosystem is still less mature than CUDA. But the trend is obvious: the NVIDIA–AMD duopoly is cracking.
Trillion parameters, two delays, and one Huawei Ascend 950PR
The expected release of DeepSeek V4 has been postponed twice already. According to Reuters, the model will be released in the second half of April 2026.
Specs: trillion-parameter range, million-token context window, multimodality.
But the most important thing is the hardware. DeepSeek V4 will run on Huawei Ascend 950PR, not NVIDIA. According to Reuters, Alibaba, ByteDance, and Tencent have placed orders for hundreds of thousands of Huawei chips ahead of the V4 release. The scale of purchases indicates a fundamental shift: China’s largest AI companies are seriously rebuilding their infrastructure for domestic accelerators.
It seems that DeepSeek deliberately did not grant AMD and NVIDIA early access to V4, giving Huawei and Cambricon extra time to optimize software. This is not a side effect of sanctions — it is a strategic choice.
However, there are nuances. In August 2025, DeepSeek reported problems: Huawei Ascend demonstrated instability, slow interconnects, and immature software. If V4 indeed launches on Ascend and shows competitive performance — it will be a milestone. If it is delayed a third time, the market will draw its own conclusions.
NVIDIA-CUDA vs. Huawei-CANN
What is happening before our eyes is not just a chip race. It is the formation of two parallel AI stacks.
The first camp: NVIDIA + CUDA + PyTorch. Mature, dominant, closed to sanctioned countries.
The second camp: Huawei Ascend + CANN + MindSpore (and PyTorch adaptations). Growing, less mature, but fully sovereign for China — and potentially accessible to third countries.
Intermediate options also exist. The HeteroFlow v2 (Enterprise Edition) project — a unified dispatch platform for heterogeneous GPUs — already supports the simultaneous operation of chips from Moore Threads, Hygon, Cambricon, and other Chinese manufacturers. For Russian data centers buying GPUs from different Chinese vendors, such platforms become critical: without a unified orchestrator, a heterogeneous cluster is not a cluster, but a set of expensive servers that cannot work together.
The New York Times published a column on April 13 with a telling headline: «I went to China to see their AI progress. We cannot overtake them.» The author — an entrepreneur who visited Chinese AI labs — writes that the message of export controls (blocking China’s access to powerful chips) no longer works. Chinese chips are already sufficient for training frontier models.
ASML raises its forecast
On April 15, ASML reported Q1 2026: revenue and profit beat expectations, and the annual forecast was raised. CEO Christophe Fouquet: «The growth prospects for the semiconductor industry continue to strengthen, fueled by ongoing investments in AI infrastructure. Demand for chips exceeds supply.»
At first glance, good news. But ASML stock fell — the market feared tightening export controls to China.
Context: in Q4 2025, China was ASML’s largest market — 36% of sales. The MATCH Act threatens to cut Chinese fabs off from DUV machine servicing. If this happens, ASML will lose a significant share of revenue, and Chinese factories will begin to degrade.
The paradox is that ASML is growing on AI demand but is itself a hostage to geopolitics. Western fabs (TSMC, Samsung, Intel) order EUV systems for new-generation chip production — but simultaneously Washington cuts off the second-largest market. For Russian buyers, this is another argument in favor of Chinese GPUs: dependence on a full-cycle Western equipment stack is a risk that will only grow.
Meanwhile, the conflict in the Middle East and the blockade of the Strait of Hormuz create additional uncertainty for global electronics supply chains.
Reuters notes that risks for AI chip and server demand are intensifying — this could slow Chinese exports, which hit records in early 2026.
Openness is over
Another trend that Stanford recorded but the market has yet to appreciate: Alibaba and Zhipu are closing their flagship models. Qwen3.6-Plus and GLM-5-Turbo are moving to closed hosted formats — open weights are no longer published.
The logic is simple: openness was a tool for the laggards. When models were weaker than GPT-4, open weights attracted developers, an ecosystem was formed, and a reputation was built. Now that quality has equalized, openness becomes a free gift to competitors. ByteDance, Alibaba, Zhipu — everyone is closing the doors.
For the market, this means: the era of «take the model and run it locally» for frontier systems is ending. Enterprise clients who want data sovereignty will have to choose between slightly older open-weight models (Llama, Mistral, DeepSeek V3/R1) and hosted services with potential data transfer risks.
What this means for the Russian market
The events of the last week form a concrete picture for Russian buyers of AI infrastructure:
Model quality parity = freedom of hardware choice. If frontier models from different labs provide comparable quality, the question of «which GPU to run on» is no longer tied to «only NVIDIA can run the best model.» Chinese GPUs are becoming not a compromise, but a conscious choice.
Huawei on export — if it happens, the market will expand. Currently, Russian companies work with Huawei Ascend through Chinese partners in the grey zone. Official Huawei exports would dramatically simplify logistics and legal transaction clarity.
HeteroFlow and similar platforms — the glue for heterogeneous clusters. Russian data centers buying GPUs from MetaX, Moore Threads, Iluvatar, and Huawei simultaneously need an orchestrator. Without it, scaling is a painful process with manual integration.
China-Russia trade turnover is growing. According to Chinese customs, from January to March 2026, trade turnover amounted to $61.25 billion — a 14.8% YoY increase. Sergey Lavrov’s visit to Beijing on April 15 and the planned visit of V.V. Putin in the first half of 2026 signal deepening economic ties. In this context, the expansion of Chinese GPU and server supplies is not an abstract scenario, but a current trajectory.
Contact our experts to select the optimal hardware configuration and calculate rental pricing for your AI workloads.