Huawei Prepares 160,000 Ascend 950DT Accelerators for DeepSeek Data Center
Chinese AI lab DeepSeek has placed an order for 160,000 Huawei Ascend 950DT accelerators, each with 144 GB HBM memory and 1 PFLOPS FP8 peak compute. The system will be deployed in a single Inner Mongolia data center, achieving 160 ExaFLOPS FP8 (320 ExaFLOPS FP4) and 23 PB total HBM memory. This is one of the largest known orders for domestic Chinese AI chips, reflecting growing demand for alternatives to Western hardware.
Related Tech News
AI CuratedMemory is now 63% of an AI accelerator's component cost, up from 52% in early 2024, per the Hot Chips 2026 memory tutorials
A detailed analysis of Hot Chips 2026 memory tutorials shows that memory's share of AI accelerator BOM has risen to 63%. The post covers HBM bandwidth growth lagging behind FLOPS, DRAM wafer capacity constraints, and proposed solutions: Samsung's custom HBM base die, Oxmiq's High Bandwidth Flash, and Samsung's LPDDR5X-PIM (which achieves 3× tokens/sec on Llama 3.1 8B). Each approach has different gating factors.
Authors push back as publishers and agents seek share of Anthropic settlement
Following Anthropic's $1.5 billion copyright settlement for training AI on pirated books, authors are complaining that publishers are claiming payments for books that have reverted rights or seeking full 100% when entitled to 50%. Some attribute it to poor record-keeping, while the Authors Guild downplays intentional malice. The settlement covers nearly 500,000 titles with $3,000 per pirated work.
Huawei Prepares 160,000 Ascend 950DT Accelerators for DeepSeek Data Centar
DeepSeek has placed an order for 160,000 Huawei Ascend 950DT chips, each with 144 GB HBM memory and ~1 PFLOPS FP8. The system will be deployed in a single data center in Inner Mongolia, providing 160 ExaFLOPS FP8 and 23 PB total HBM memory for large-scale inference.
AMD’s Ryzen AI MAX+ PRO 495 Powers GMKtec’s EVO-X5 PRO To Run 300B+ Models Locally On 192 GB of Unified Memory
GMKtec's new EVO-X5 PRO flagship mini workstation features AMD's Ryzen AI MAX+ PRO 495 'Gorgon Halo' SoC with 16 cores and 192 GB of unified memory, designed to run large AI models over 300 billion parameters locally. The company claims 66% faster CPU orchestration and 40% lower cost per AI workflow versus NVIDIA's DGX Spark. The article is a brief announcement of the product launch.