AI Inference is Reshaping Chip Architecture, Markets, and Geopolitics

news14052026

Spring became the inflection point for the semiconductor industry. Not another war over training clusters — oh no, this time the tectonic shift happened elsewhere. AI inference, meaning the deployment of models to production, has for the first time surpassed training in terms of both compute volume and, more importantly, hardware demand. The consequences — ranging from the return of the CPU to the center of architecture to the diverging trajectories of Chinese GPU vendors and new records in smuggling — are unfolding simultaneously.

CPU returns, super-level for server processors

GlobalFoundries, one of the largest contract chip manufacturers, officially declared in May 2026 the onset of a supercycle in the server CPU market. The cause is the explosive demand for AI inference and agentic applications (AI Agents).

The architectural logic is simple: in the training era, the GPU-to-CPU ratio in a server was 8:1. The processor was the support staff for the video cards. During inference, the ratio shifts to 4:1 or even 2:1. And in AI agent scenarios, where the system continuously coordinates multiple models, processes context, and makes decisions, the CPU may find itself in a 1:1 ratio with the GPU.

AMD reacted instantly: at its May earnings call, the company doubled its forecast for the total addressable market of server CPUs by 2030 — from $60 billion to $120 billion. Intel’s stock nearly doubled in 2026: its own fabs (IDM model) give Intel an advantage in scaling production speed compared to competitors dependent on TSMC.

The AI compute market is ceasing to be a GPU monopoly. Infrastructure investments are now distributed among GPUs, CPUs, and memory — and this changes the rules of the game for all participants.

The «Six Dragons» of Chinese GPU

The combined market capitalization of six leading Chinese GPU companies — Cambricon (寒武纪), Hygon (海光信息), MuXi (沐曦), Moore Threads (摩尔线程), Iluvatar (天数智芯), and Biren (壁仞科技) — has exceeded ¥2 trillion ($280 billion). But behind this figure lies a sharply diverging reality.

Cambricon — the profit leader. Q1 2026: revenue ¥2.885 billion (+160% YoY), net profit ¥1.01 billion (+185%). Almost half of the entire 2025 profit was earned in a single quarter. But there is an alarming signal: inventory on the balance sheet is about ¥4.5 billion (1.6 quarters of revenue), plus a write-down of inventory impairment of ¥246 million. Whether this is a strategic purchase or overstocking will be shown by the next quarters.
Moore Threads — first quarterly profit in history. Q1 2026: revenue ¥738 million (+155%), net profit ¥29.4 million. But excluding one-off items (扣非), the company is still loss-making. The key driver is the S5000 card (80 GB), adapted for DeepSeek V4 in Day-0 mode. The KUAE cluster contract ($95.5 million) continues to generate revenue.
MuXi — growth exists, but losses persist. Q1 2026: revenue ¥562 million (+75%), net loss ¥99 million (reduced from ¥233 million a year ago). The IPO on the STAR Market in December 2025 brought in ¥4 billion, but the path to profitability is still ahead.
Iluvatar and Biren (traded in Hong Kong) reported revenue growth but remain in the deep red. The market, however, values them on potential rather than current profit.

The Chinese GPU sector has passed the point of no return — several companies are already generating positive cash flow. But the «Six Dragons» are not a homogeneous force, but a spectrum from maturity to startup. And the distance between the first and the last will only grow.

Huawei Atlas 950 SuperPoD, a systemic answer to sanctions

At MWC Barcelona 2026, Huawei presented the Atlas 950 SuperPoD — a new generation supernode built on Ascend 950DT chips and the Lingqu 2.0 (灵衢) interconnect.

The numbers speak for themselves:

8,192 Ascend 950DT cards (20+ times more than the predecessor — Ascend 384).
8 EFLOPS FP8 / 16 EFLOPS FP4.
1,152 TB of unified memory with a single address space.
16 PB/s of interconnect bandwidth (62 times more than NVIDIA NVL144).
Based on Atlas 950 SuperCluster: 524,288 cards, FP4突破1 ZFLOPS.

The key innovation is not in the chip, but in the system. Lingqu 2.0 is a fully optical interconnect protocol with a latency of 200ns per hop (10 times lower than Lingqu 1.0). Huawei has already opened the Lingqu specifications (600+ pages), allowing third-party manufacturers to develop compatible equipment.

Another critical technology is the unified memory address space: all 8,192 cards address shared memory directly, without serialization/deserialization of transfers. This eliminates the main «bottleneck» in training and inference of large-scale models.

Huawei is literally implementing a new strategy: compensating for the lithography lag with system architecture. Atlas 950 enters the Chinese market in Q4 2026.

H3C UniPoD S80000, the multi-vendor supernode

At NAVIGATE 2026 (May 8, Beijing), H3C unveiled the UniPoD S80000 supernode line — for trillion+ parameter models. The key difference from Huawei is the multi-vendor architecture. UniPoD S80000 supports GPUs from several manufacturers, including Chinese vendors. This is important for customers who do not want to be tied to a single ecosystem.

H3C claims to have solved the GPU utilization problem: according to Cast AI, up to 95% of purchased AI-GPUs in corporate clusters sit idle. UniPoD S80000 promises full-stack optimization — from the network to orchestration — to increase the actual utilization of equipment.

Smuggling, sanctions, and the real price of NVIDIA in China

An NVIDIA B300 server (8×GPU) costs ¥7 million ($1 million) in China. In the US, the same server is $550k — a difference of almost two-fold. The reason is not only sanctions — the supply channel narrowed after the arrest of Supermicro’s co-founder.

Supermicro case (May 2026): Supermicro co-founder Wally Liaw was arrested on charges of organizing the smuggling of servers with NVIDIA H200 into China and Russia via the Thai company OBON — an estimated volume of $2.5 billion. The scheme involved creating shell companies in Southeast Asia, serial number substitution (they «blew the labels with a heat gun» and re-glued them), and using connections with Thai state structures.

Parallel to this — Jensen Huang in an interview with Congress stated: «China should not get the most advanced GPUs.»

And confirmation from the Secretary of Commerce: H200 is permitted for export to China, but not a single Chinese customer has bought it.

Sanctions did not stop the penetration of chips — they made it more expensive and criminal. And the price for the end customer in China has doubled.

Vera Rubin: summer testing, autumn — batch

NVIDIA confirmed the schedule: Vera Rubin — trial release in June, first deliveries to key customers in July, mass production in the second half of 2026. But TrendForce lowered the forecast for Rubin’s share in 2026 from 29% to 22% due to HBM4 supply issues and TSMC’s 3nm capacity constraints.

This means: the window for Chinese vendors remains open at least until the end of 2026. Until Rubin reaches full capacity, inference solutions on Chinese chips retain a price advantage.

What this means for the Russian market

Inference is a window of opportunity. The shift from training to inference lowers the requirements for homogeneous H100/B200 clusters. For inference, heterogeneous systems can be used — Chinese GPUs + domestic CPUs.
Huawei and H3C show the way. Atlas 950 SuperPoD and UniPoD S80000 are not just «Chinese analogs of NVIDIA.» These are new-type system products where the competitive advantage is created by the interconnect architecture, not just chip power. For the Russian market, where sanctions limit access to top-tier GPUs, system optimization is a strategic response.
Smuggling is a risk, not a solution. The Supermicro case shows that gray supply channels are closing. Legally supplied Chinese GPUs, delivered through partnerships, are the only sustainable path for Russian enterprise.
CPU supercycle is a trump card for import substitution. The growing demand for server CPUs during inference means that domestic processors (Baikal, Elbrus) gain a real role in AI infrastructure — not as a GPU replacement, but as a mandatory system component. This dilutes the monopoly of GPU vendors and opens space for hybrid solutions.
Divergence of Chinese vendors — a filter for partners. Not all «Six Dragons» will reach maturity. For a Russian customer, it is critical to choose those who have already turned a profit (Moore Threads, Cambricon) or have strong strategic support (Huawei), rather than those who are still burning cash.

Let's start a conversation

Start Scaling Today

Contact our experts to select the optimal hardware configuration and calculate rental pricing for your AI workloads.

Location: Moscow, ext.ter.g., sh. Ochakovskoe, 28, p.2
GPU Servers High-density compute
Облачная аренда Быстрый старт ИИ моделей
Enterprise SLA Guaranteed availability
Low Latency Optimal network routes