Insightful AI World

Sign in Subscribe

hardware

DeepSeek V4 on Huawei Ascend: Open Weights, MoE at Trillion Scale, and the Self-Hosting Path

DeepSeek V4 ships under MIT license in two MoE sizes (1.6T Pro and 284B Flash) with 1M-token context. Huawei's Ascend 950 SuperNode handles inference. Here is what readers can do with it — and what comes next.

Close-up of a computer chip under cool studio light, dark studio background, polished silicon surface with fine die pattern visible.

Cerebras IPO at $86B: What the 168x Multiple Underwrites

Cerebras priced May 13 and closed day one at a ~168x revenue multiple. The first-day pop is the smaller story. The capex signal underneath it is the bigger one.

Close-up macro photograph of an HBM2 DRAM die on a Radeon RX Vega 64 GPU — the actual silicon that stores the bytes a model has to load and reload during inference.

What is HBM memory, and why does AI need so much of it?

HBM is the high-bandwidth memory glued to every modern AI accelerator — and the single component most likely to be the bottleneck the next time someone says 'we need more GPUs.'