# 谷歌TPU v8与华为昇腾平台：全球AI芯片竞赛开启新阶段

- 来源：Chubby♨️ (@kimmonismus)
- 发布时间：2026-04-27 18:46
- AIHOT 分数：63
- AIHOT 链接：https://aihot.virxact.com/items/cmoh2wexv02dgslwpqiiny1nx
- 原文链接：https://x.com/kimmonismus/status/2048715295226032219

## AI 摘要

谷歌在Cloud Next 2026上首次将TPU v8拆分为训练芯片TPU 8t和推理芯片TPU 8i，宣称训练速度提升2.8倍，推理性价比提高80%，并通过自研Arm架构Axion CPU实现全栈垂直控制。同时，DeepSeek V4-Pro成为首个在华为昇腾NPU平台上完成训练与推理验证的前沿大模型，其定价与昇腾950芯片量产计划挂钩，输出成本远低于主流西方模型。这标志着美国制裁试图阻止的硬件脱钩可能已不可逆转，全球AI芯片竞争进入新阶段。

## 正文

Google's TPU v8 and Huawei's Ascend NPU platform： the global Chipwar just began

At Cloud Next 2026， Google unveiled its eighth-generation TPU as two separate chips for the first time： the TPU 8t for training and the TPU 8i for inference， claiming up to 2.8x faster training and 80% higher performance per dollar for inference compared to last year's Ironwood.

The 8t was designed by Broadcom， the 8i by MediaTek， applying mobile-edge efficiency logic to inference while maximizing raw throughput on training. The 8t connects up to 9，600 accelerators via optical-circuit switches， dwarfing NVIDIA's 576-GPU NVLink domain， and a new Virgo network fabric scales beyond one million chips for a single training job.

Google is also replacing x86 hosts with its own Arm-based Axion CPUs， completing full vertical control from host to accelerator to network. The message is clear： the general-purpose AI accelerator is a fading category.

DeepSeek V4 on Huawei Ascend： China's parallel infrastructure takes shape

DeepSeek's V4 release is the more geopolitically consequential event. The 1.6 trillion-parameter V4-Pro is the first major frontier model to validate both training and inference on Huawei's Ascend NPU platform alongside NVIDIA GPUs.

The nuance： DeepSeek adapted only part of V4's training for Chinese chips and confirmed Ascend for inference， while pre-training of V4-Pro likely still relied on NVIDIA silicon.

Is this a novum？ Yes. No frontier-class model has ever publicly validated on non-NVIDIA hardware at this scale. More importantly， DeepSeek is tying future pricing to Huawei's Ascend 950 production ramp in H2 2026， making this an economic bet， not a symbolic gesture. V4-Pro costs $3.48 per million output tokens versus $30 for GPT-5.4 and $25 for Claude Opus 4.6. The real story isn't whether V4 beats Western models on benchmarks （it doesn't quite）， but whether the hardware decoupling U.S. sanctions were designed to prevent is now irreversibly underway.