Back

Cambricon Surges >110% as China's AI Strategy Pivots to Self-Reliance

7 min read
By Silicon Analysts

Executive Summary

China's focus is shifting from competing on foundational AI models to building self-reliant, scalable hardware ecosystems for inference. This strategic pivot, driven by geopolitical realities and the upcoming 15th Five-Year Plan, is creating significant demand for domestic hardware champions like Cambricon and reshaping global semiconductor supply chains.

1Cambricon Stock Growth: The company's stock has surged over 110% in the past year, reflecting market confidence in its role in China's AI strategy.
2Hardware Demand Shift: Market focus is moving from training-centric supremacy to scalable, efficient inference platforms, benefiting specialized hardware providers.
3Optical Component Boom: Demand for 800G and 1.6T optical modules for AI data centers has made it one of the strongest performing segments in China's domestic markets.
4Strategic Self-Reliance: The upcoming 15th Five-Year Plan is expected to intensify investment in domestic semiconductor production, from lithography to advanced packaging, to mitigate US trade restrictions.

Supply Chain Impact

The tectonic shifts in China's technology strategy, crystallized by the upcoming 15th Five-Year Plan (2026-2030), are a direct response to escalating technology sanctions imposed by the United States and its allies. These policies have fundamentally altered the landscape for Chinese technology firms, forcing a strategic and costly pivot from reliance on global supply chains to the development of a vertically integrated, domestic ecosystem. The over 110% appreciation in Cambricon Technologies' stock is not an isolated event but a clear market signal endorsing this national strategy. Investors are pricing in the immense long-term value of companies that form the backbone of this self-reliance push.

The most acute pressure point is access to leading-edge semiconductor manufacturing. With restrictions blocking Chinese entities from accessing TSMC's and Samsung's most advanced nodes (sub-7nm), domestic foundry SMIC has become a national champion by necessity. While SMIC has reportedly achieved a 7nm process, its production faces significant hurdles in terms of yield, scale, and cost-effectiveness compared to global leaders. Yields for this node are estimated to be in the 40-50% range, far below the 80-90% typically seen at TSMC for a mature process. This directly inflates the per-die cost and limits the viable supply of high-performance processors.

Furthermore, the inability to procure state-of-the-art EUV (Extreme Ultraviolet) lithography equipment from ASML locks Chinese foundries out of nodes below 7nm, creating a durable technology gap. This forces Chinese chip designers like Cambricon and Huawei's HiSilicon to innovate within the constraints of DUV (Deep Ultraviolet) lithography, focusing on architectural improvements, chiplet designs, and advanced packaging to extract more performance from relatively mature process nodes.

This bottleneck extends to critical components like High-Bandwidth Memory (HBM). While China has domestic DRAM manufacturers like CXMT, they are several generations behind global leaders SK Hynix, Samsung, and Micron in HBM development and mass production. Access to the latest HBM3 and HBM3e stacks is heavily restricted, forcing Chinese AI accelerator designers to either use older HBM2e or develop innovative interconnect solutions with high-speed GDDR6 memory, which presents a performance trade-off.

The Rise of Domestic AI Accelerators

In this constrained environment, the market has shifted its focus from a singular pursuit of raw training performance, epitomized by large-scale foundational models, to the practical and scalable deployment of inference workloads. Inference, which involves running trained models to make predictions, is often less computationally demanding but requires highly efficient, cost-effective, and readily available hardware. This is the market segment where companies like Cambricon thrive.

Cambricon's MLU (Machine Learning Unit) series of accelerators is designed specifically for cloud and edge inference. By optimizing for these workloads, Cambricon can deliver competitive performance-per-watt and performance-per-dollar using mature manufacturing processes (e.g., 12nm, 7nm) that are accessible domestically. This aligns perfectly with the national strategy of building a vast, scalable AI infrastructure independent of foreign technology.

The table below provides an approximate comparison of leading domestic Chinese AI accelerators against Nvidia's export-controlled equivalents. While direct benchmarks are scarce, these estimates are based on architectural specifications and publicly available data.

FeatureNvidia H20 (Export)Huawei Ascend 910BCambricon MLU590Biren BR100
Process NodeTSMC 4N (Custom)SMIC 7nm (Est.)TSMC 7nmTSMC 7nm
FP16/BF16 Perf.~294 TFLOPS (dense)~320 TFLOPS (dense)~256 TFLOPS (dense)~1024 TFLOPS (dense)
INT8 Performance~588 TOPS~640 TOPS~512 TOPS~2048 TOPS
Memory TypeHBM3HBM2eHBM2eHBM2e
Memory Bandwidth~4.0 TB/s~1.2 TB/s~1.2 TB/s~2.3 TB/s
InterconnectNVLink (~900 GB/s)HCCL (~800 GB/s)MLU-Link (~300 GB/s)BCC (~896 GB/s)
Target WorkloadGeneral Purpose AITraining & InferenceInference FocusTraining Focus

Note: Performance figures are theoretical peak values and can vary significantly based on workload and software optimization. Data compiled from public sources and analyst estimates.

Huawei's Ascend series, particularly the 910B, represents another pillar of China's AI hardware strategy. Manufactured by SMIC on an estimated 7nm process, it serves as a direct, albeit slightly less performant, alternative to Nvidia's restricted A800/H800 GPUs for training workloads. The development of a proprietary software ecosystem (CANN) around the Ascend platform is a critical component of building a competitive moat and reducing reliance on Nvidia's CUDA.

This domestic hardware boom is complemented by surging demand for adjacent technologies. The need to connect thousands of these accelerators within data centers drives the market for 800G and **1.**6T optical transceivers, explaining why this segment has seen such strong performance in China's A-share market. Companies like Zhongji Innolight and Eoptolink have become essential enablers of this domestic AI scale-up.

Strategic Implications for Procurement and Roadmaps

The bifurcation of the global semiconductor supply chain presents both profound challenges and unique opportunities for technology companies worldwide. Procurement teams must now navigate a landscape with two increasingly divergent ecosystems.

1. Dual-Sourcing & Risk Mitigation: For multinational corporations operating in China, developing a dual-sourcing strategy is no longer optional, but essential. This involves qualifying and integrating domestic hardware like the Ascend 910B or Cambricon MLUs for China-based data centers, while continuing to use state-of-the-art Nvidia or AMD hardware elsewhere. This mitigates the risk of future sanctions disrupting operations but introduces significant engineering overhead in maintaining software compatibility across different hardware architectures.

2. Focus on Software Abstraction: The rise of multiple hardware architectures underscores the importance of software. Companies that invest in hardware-agnostic AI software layers, such as those leveraging OpenXLA or TVM, will gain a significant competitive advantage. The ability to deploy models seamlessly across Nvidia, AMD, Intel, and Chinese domestic accelerators without extensive re-engineering will be a key determinant of agility and cost-efficiency.

3. Opportunities in Mature Nodes: The intense focus on domestic production in China is creating a massive demand for equipment and materials related to mature process nodes (28nm and above). For global suppliers not directly impacted by the most stringent sanctions, this represents a significant growth market. However, it comes with the geopolitical risk of being targeted in future regulatory expansions.

4. Long-Term Roadmap Planning: Hardware roadmap planning must now account for this geopolitical divergence. Product development cycles, particularly for hardware with lead times of 18-24 months, must anticipate the potential for further trade restrictions. This may involve designing systems with modular accelerator components, allowing for swaps between different suppliers based on regional requirements and supply availability. Lead times for advanced packaging services like CoWoS remain extended, typically in the 20-30 week range, and any geopolitical flare-up could exacerbate these bottlenecks significantly.

Ultimately, China's determined push for technological self-reliance, underscored by market phenomena like Cambricon's surge, is a permanent feature of the modern semiconductor industry. The era of a single, globalized supply chain is over. Strategic decision-makers must adapt to this new reality of technological spheres of influence, building resilience, flexibility, and architectural foresight into their roadmaps and procurement strategies.

References & Sources