HardwareFreaks

Tag: Ascend 950PR

  • Huawei Atlas chips pack 8K AI accelerators challenging Nvidia

    Featured image Huawei Atlas chips pack 8K AI accelerators challenging Nvidia

    Huawei Bets Big on South Korea’s AI Market with Disruptive Hardware

    In a high-stakes race for technological supremacy, Chinese giant Huawei is preparing to make a major splash in one of the world’s most critical AI hardware markets: South Korea. Scheduled to launch in the fourth quarter of 2026, Huawei plans to introduce its Ascend 950 series processors and the Atlas 950 SuperPod AI computing platform. This strategic entry signals an ambitious push into Nvidia’s stronghold, aiming to offer a powerful alternative that appeals directly to customers seeking to reduce their reliance on US-made chip technology.

    This isn’t just a hardware rollout; it’s a bold competitive maneuver. By positioning its technology against established leaders, Huawei is attempting to disrupt the existing pricing structure of high-end AI accelerators. The core strategy hinges on delivering extraordinary performance at a fraction of the cost.

    At the heart of this push are the Ascend 950PR and Ascend 950DT chips—specialized neural network processing units designed for both AI inference (processing) and training workloads, respectively. The accompanying Atlas 950 SuperPod platform is designed to scale these capabilities, promising a massive computing ecosystem capable of handling hundreds of thousands of specialized processors in a single deployment.

    The performance claims are what really grab attention. Huawei asserts that the Ascend 950PR accelerator can deliver nearly 2.87 times the inference performance of Nvidia’s flagship H20 chip, all while offering prices that are roughly one-quarter of the cost. This dramatic cost-performance ratio provides a compelling financial incentive for Korean buyers looking to upgrade their AI infrastructure.

    Technologically, Huawei leverages its in-house capabilities by constructing these systems using “self-developed” high-bandwidth memory (HBM). The Ascend line utilizes Huawei’s HiBL 1.0 and HiZQ 2.0 standards for memory, underscoring an integrated approach to chip design.

    Beyond raw power, the strategy extends into the software ecosystem. Recognizing that compatibility is key for adoption, Huawei is working to improve integration with the widely used Nvidia CUDA programming environment. This effort aims to smooth the transition for developers who are accustomed to established AI frameworks, ensuring that cutting-edge hardware can be utilized seamlessly.

    However, entering such a sensitive market is far from simple. While Huawei brings undeniable supply scale and deep software capabilities, industry observers point out significant hurdles. South Korean sentiment remains sensitive toward Chinese technology, raising concerns about security and potential vendor lock-in with a proprietary stack. Furthermore, there are practical considerations, including the power consumption and heat overhead associated with high-density silicon.

    The local AI scene is already vibrant, filled with agile startups. Huawei’s arrival, backed by its comprehensive hardware and software depth, introduces a genuine competitive threat to these emerging players. The success of this venture will ultimately depend not only on technological prowess but also on navigating the delicate geopolitical and market sensitivities of South Korea.

    Credit: Tom’s Hardware

  • China black market Nvidia prices rocket in wake of smuggling crackdown and customs freeze — five-year-old A100 servers triple in price, now fetching up to $82,000

    The high-stakes battle for artificial intelligence hardware is currently being fought not just in data centers, but across international shipping lanes and customs checkpoints. As global restrictions tighten around advanced chips, Chinese companies are facing an unprecedented scramble, paying exorbitant prices to secure the cutting-edge technology needed to fuel their AI ambitions.

    This market shift is most visible in the extreme inflation of specialized hardware. Servers built around Nvidia’s five-year-old A100 accelerator and modified gaming GPUs are commanding prices reaching as much as 600,000 Chinese Yuan, or approximately $82,000. The frenzy extends to flagship systems like Nvidia’s DGX B300, which has traded over $1.1 million on the black market in recent months alone.

    This escalating demand isn’t happening in a vacuum; it is driven by geopolitical maneuvers. A U.S. smuggling crackdown and various customs freezes have effectively choked off traditional supply routes for these essential components, forcing buyers to seek alternative, often more expensive, channels.

    The pressure extends beyond the A100. Nvidia’s next generation of restricted Blackwell hardware is also driving up prices. Workstation cards like the RTX 6000 Pro have seen dramatic increases in value, moving from roughly $5,580 to over $14,500. This reflects a broader trend: GPU rates within China now often match or even exceed prices seen in the United States, reversing previous smuggling discounts.

    Meanwhile, governments are actively reshaping the landscape. Following enforcement actions by Washington at the end of last year, and subsequent legal charges against some involved parties, efforts to control the flow of AI chips intensified across various jurisdictions. Authorities in Taiwan and Malaysia initiated their own investigations into illegal re-export routes, effectively drying up traditional smuggling networks.

    Beijing’s response has focused on fostering domestic resilience. While border controls were implemented for specific exports, domestic focus pivoted toward homegrown solutions. Huawei, for instance, launched the Ascend 950PR as a competitive inference chip. Although this domestic hardware is gaining traction in testing at large data center clients, it currently trails Nvidia’s superior CUDA software stack.

    This technological disparity creates a major bottleneck: even if domestic producers successfully scale up their manufacturing, they still face underlying supply chain issues. Rising costs for essential components like DRAM and HBM are compounding the difficulty of building advanced AI hardware.

    The result is a complex scenario where technological ambition collides with global trade policy and material scarcity. Until domestic chip development can fully absorb this redirected demand, the high prices for existing Nvidia inventory in China are set to continue climbing.

    Credit: Tom’s Hardware