Orbital Foundries for Advanced Semiconductors
The Cost of Compute Amidst the HBM Shortage

The specialized compute market is experiencing significant volatility: over the past three months, the price of Huawei's flagship Ascend 950DT AI accelerators has surged by approximately 60%. A single chip now commands 250,000 yuan (roughly $37,300), effectively placing it in the same price bracket as Nvidia's powerhouse B200 server GPU. This trend underscores a sobering reality: even with proprietary developments, the manufacturer remains vulnerable to the intricacies of global supply chains.
The primary driver behind this price hike is the acute shortage of High Bandwidth Memory (HBM). In modern AI accelerator architectures, memory has evolved from a mere storage repository into a critical performance bottleneck. HBM enables massive data throughput between the memory and the compute core with minimal latency—a prerequisite for training and deploying Large Language Models (LLMs). This global deficit has persisted for over a year, with repercussions felt across the entire spectrum, from server racks to premium consumer smartphones.
The situation is further exacerbated by a surge in demand for domestic Chinese alternatives to Nvidia's solutions. Restricted access to Western technology has pushed local players to aggressively scale their compute capacity, placing immense pressure on Huawei's production lines. The company has openly acknowledged that HBM supply constraints are directly impacting both the cost of goods sold and equipment availability, effectively turning memory into the most expensive and scarce resource in the AI hardware stack.
Despite pricing pressures and manufacturing hurdles, the strategic ambitions of the region's tech giants remain undiminished. A prime example is the DeepSeek project, which plans to deploy a massive compute cluster in Inner Mongolia. Realizing this vision will require no fewer than 160,000 Ascend 950DT accelerators. Such monolithic orders are forcing Huawei to strictly prioritize shipments, favoring the largest infrastructure players capable of driving systemic growth in the AI market.
Current production volumes for the 950DT series are estimated at several hundred thousand units for the year. However, access to this capacity is tightly controlled: equipment is supplied in batches of at least eight units, with final pricing determined on a bespoke basis. Costs depend not only on order volume but also on the complexity of the accompanying networking hardware required to interconnect thousands of chips into a unified compute fabric. Ultimately, the market is revealing a stark correlation: the cost of intelligence today is determined not just by the quality of the code, but by the physical availability of silicon and memory.

