TechInsights teardown points to global parts inside Huawei’s Ascend 910C AI chip
Huawei’s latest data center AI processor is back in the spotlight after a fresh teardown. According to semiconductor research firm TechInsights, the Ascend 910C still includes components tied to Taiwan Semiconductor Manufacturing Company (TSMC), Samsung Electronics, and SK Hynix. The finding underscores how deeply intertwined the world’s chip supply chains remain, even amid years of export controls and efforts to localize advanced semiconductor production.
What TechInsights found matters because the Ascend 910C sits at the heart of Huawei’s push into high-performance AI computing. The Ascend line targets training and inference workloads for data centers, powering everything from large language models to computer vision and scientific computing. Any indication that the platform relies on parts from leading global players sheds light on the practical realities of sourcing advanced components in a constrained environment.
It’s important to note that “components” in a teardown can encompass a range of parts and processes beyond the main compute die. That can include memory chips, power management ICs, interface controllers, passive components, and packaging or fabrication services for supporting chips on the board. In other words, a system can incorporate parts from multiple vendors even if the primary processor is fabricated elsewhere.
Why this is significant:
– It highlights the durability of legacy inventory and the role of intermediaries. Components from companies like Samsung and SK Hynix—especially memory—are widely used and can reappear via existing stock or third-party channels.
– It reinforces how difficult it is to completely unwind cross-border tech dependencies. Modern AI accelerators rely on complex stacks of silicon, memory, power delivery, and packaging—often sourced globally.
– It offers a window into the pace of domestic substitution. Observers are closely watching how quickly Huawei and its partners can replace imported parts with homegrown alternatives across compute, memory, and supporting ICs.
The context around Huawei’s hardware makes this finding especially noteworthy. Over the past few years, export controls have reshaped how and where high-end chips can be made and sold. While companies have adapted—by tapping domestic foundries, optimizing designs, and reworking supply chains—teardowns like this one suggest the transition is uneven and still unfolding. The presence of components linked to TSMC, Samsung, and SK Hynix does not necessarily indicate new shipments; it may reflect pre-existing inventories or complex procurement paths that are common in the electronics industry.
For data center operators and AI developers, the big picture is clear: performance leadership in AI silicon depends on more than the core compute die. High-bandwidth memory, efficient power delivery, thermal design, and reliable packaging all contribute to overall throughput and energy efficiency. If a platform like the Ascend 910C integrates well-known memory brands or supporting ICs manufactured by established foundries, it can influence real-world performance and availability.
What to watch next:
– Follow-up teardowns and yield analyses that clarify which subsystems rely on non-domestic parts and how that mix evolves in upcoming revisions.
– Shifts toward domestic memory or controller ICs as supply chains adjust, potentially affecting bandwidth, latency, and thermal profiles.
– Software stack and system-level optimizations that help mitigate hardware constraints, keeping AI training and inference competitive even when component sourcing is in flux.
Bottom line: TechInsights’ disassembly of Huawei’s Ascend 910C suggests that, despite sustained efforts to localize production, global components still play a role in the company’s AI hardware. It is a reminder that the semiconductor ecosystem remains deeply interconnected—and that the race to build competitive AI accelerators hinges on a complex, multilayered supply chain that is not easily or quickly reconfigured.
