Huawei and Cambricon raise prices for AI accelerators due to rising HBM memory costs
Read more
Pandaily
pandaily.com

Huawei and Cambricon raise prices for AI accelerators due to rising HBM memory costs

According to sources familiar with customer pricing offers, Huawei and Cambricon have increased the estimated prices for current and future artificial intelligence accelerator cards amid rising costs for high-performance memory (HBM).

This report provides an update on the supply chain and product release schedule: HBM stack components now constitute a significant portion of the cost of finished cards. Meanwhile, Huawei's stated release date for the Ascend 950DT in the fourth quarter of 2026 remains unchanged.

Reuters reported that Huawei has raised the estimated price of its Ascend 950DT accelerator above 250,000 yuan. This is approximately 20%–50% higher than the prices quoted about two months ago, depending on contract terms. Bloomberg separately noted that the projected prices have risen by about 60% in three months, reaching around 250,000 yuan, with one source comparing this price to the Nvidia B200 for similar multi-card deals.

Both publications emphasize that the final cost still depends on the order volume and how many components customers purchase, including boards, network equipment, and other auxiliary elements. Huawei stated that the 950DT, which is the most advanced Ascend product for high-load decoding model development and inference, is still scheduled for the fourth quarter of 2026. The company described proprietary memory stacks for the Ascend 950 series—HiBL 1.0 for 950PR and HiZQ 2.0 for 950DT—but did not disclose details about the memory production.

Cambricon also revised the price for its next model, preliminarily named 690, raising it by approximately 20%–30% compared to the level indicated two months ago, although the exact amount in yuan was not disclosed. A similar price increase is also observed among smaller competitors MetaX and Iluvatar. Older Huawei boards have also seen changes: the Ascend 950PR has become more expensive, rising from approximately 60,000 yuan at the beginning of the year to over 80,000 yuan, and the Ascend 910C has increased from about 90,000 yuan to over 110,000 yuan, according to sources.

The main limiting factor is memory. Supplies of advanced HBM remain concentrated among a small number of global suppliers, and Chinese accelerator manufacturers are increasingly forced to use 'gray market' channels with higher costs, which directly affects the final card price. Huawei, Cambricon, MetaX, and Iluvatar did not comment on the pricing discussion for the publications.

As additional information, Iluvatar has roughly doubled its GPU supply to ByteDance to 100,000 units this year, according to one Reuters source, and redirected internal GPUs to meet this demand. Sources also reported that Huawei remains the largest domestic supplier for ByteDance, followed by Cambricon and Iluvatar. For buyers planning clusters for 2026, the immediate signal is the higher estimated prices and tighter memory logistics before the Q4 window for the Ascend 950DT, rather than a change in the product development plan.

Similar stories

UniPat raises $300 million with Alibaba support to expand AI testing and benchmarking
Read more
ventureburn.com

UniPat raises $300 million with Alibaba support to expand AI testing and benchmarking

As artificial intelligence develops, the problem of testing has emerged. While many focus on creating impressive large language models to attract investors, there are only a small number of verified platforms to check the functionality of these complex systems after training is complete.

AI models are prone to hallucinations and can drift from the norm over time. To solve this significant operational problem, which constantly deters large corporate users, UniPat has raised $300 million. UniPat actively conducts stress tests on these complex systems to ensure a full level of trust.

UniPat differs from other AI companies; they do not just create new algorithms to demonstrate at expensive tech conferences. Instead, the startup specifically tests models based on real-world scenarios, rather than sterile laboratory conditions.

This can be compared to a rigorous training camp for artificial intelligence. Before a model gains access to real consumer data or begins performing automated financial operations, UniPat subjects it to intensive trials. This allows for the generation of high-quality system performance data that developers need to eliminate critical flaws, as what hasn't been thoroughly broken first cannot be fixed.

The tech giant Alibaba led this major funding round, causing a stir in the industry. The financial details of this round are quite impressive: the new influx of capital boosted the startup's valuation to an impressive $2.5 billion post-investment. It is also interesting that this specialized center is headed by a former company employee.

It is clear that Alibaba is interested in retaining its top talent. Significant funds continue to flow despite the caution of the broader venture capital market. Moreover, this specific deal ranks among the top three percent of all registered late-stage venture capital rounds.

The high cost of the testing platform is due to basic corporate economics and the need for competitive survival. Early investors included representatives from Sequoia China. This demonstrates that leading financial players are looking for infrastructure-related enterprises amid the current AI race. We have moved past the phase of simply admiring smart chatbots.

Now, large international corporations demand flawless analytics. They need solid proof that implementing multi-million dollar AI will not lead to public embarrassment. UniPat provides this necessary insurance policy. This is not just another routine technology investment. It is a calculated move. As global powers fiercely compete for dominance in artificial intelligence, the basic infrastructure for evaluating these tools becomes infinitely valuable.

Alibaba's massive bet signals a clear shift in market priorities. The future belongs not only to those who can build the largest model but also to those who can prove their model is the safest, fastest, and most reliable in real-world conditions.

Popular