Global GPU Computing News【20250524】

1. Middle East AI Partnerships Accelerate

   • Saudi AI firm Humain partnered with NVIDIA and AMD to build AI factories powered by hundreds of thousands of NVIDIA GPUs over the next five years. The first phase includes 18,000 Blackwell chips, with AMD providing data center chips and software. Total investment: $10 billion.

   • UAE’s G42 secured an annual import quota of 500,000 top-tier NVIDIA GPUs (20% for its data centers, the rest for U.S. firms). It also launched a 5GW AI data center park with U.S. partners, with 1GW already under construction.

2. UAE’s "Stargate" Project Advances

   • The UAE launched the world’s largest AI data center, "Stargate," deploying 100,000 NVIDIA GB300 GPUs in its first phase (2026 launch). The project, led by G42 with OpenAI and Oracle, uses NVIDIA’s Grace Blackwell architecture and targets 1GW computing power.

3. Intel Releases New AI Inference GPUs

   • Intel unveiled the Arc Pro B-series GPUs (B60/B50), offering 3x more VRAM than NVIDIA’s counterparts. They support four-GPU systems for 100k-token LLM inference and eight-GPU configurations via Project Battlematrix, achieving 192GB pooled VRAM while complying with U.S. export controls.

4. Tesla and xAI Expand Computing Power

   • Musk’s xAI deployed 200,000 GPUs in Memphis, Tennessee, aiming to scale to 1 million. However, fossil fuel reliance sparked environmental concerns, prompting plans to integrate energy storage solutions.

5. NVIDIA’s Market Crash and Challenges

   • NVIDIA’s stock plunged 45% from its peak, erasing over $1 trillion in value. Causes include Chinese AI firms’ algorithm optimizations reducing reliance on NVIDIA’s hardware and U.S. chip export bans, slashing NVIDIA’s China market share to <5%.

6. Shanghai’s GPU Breakthroughs

   • Huawei’s Ascend 910B GPU delivers 85% of NVIDIA H100’s performance at 36% lower cost, with Q1 2025 shipments exceeding 600,000 units. Firms like Tianshu Zhixin and Biren Tech launched CUDA-compatible GPUs to accelerate import substitution.

7. Computing Cluster Construction

   • Shanghai Yidian Smart Computing Center built a 14,000-petaflops cluster using domestic GPUs (Huawei Ascend, Tianshu Zhixin). Its cross-brand heterogeneous training platform boosted utilization by 40% with 1,000-card coordination.

8. Policy and Funding Support

   • Beijing announced a 2024–2027 computing power plan, subsidizing firms that adopt domestic GPUs. Targets include 45 EFLOPS of AI compute by 2025 and centralized clusters to address fragmented infrastructure.

9. Domestic GPU Ecosystem Reshaping

   • Moore Threads’ KUAE cluster solution deployed 1,000-card clusters in Beijing and Nanjing, supporting mainstream LLM training with 91% scalability and CUDA compatibility. Software optimizations cut operational costs.

10. Challenges and Opportunities

    • Chinese GPUs lag behind NVIDIA’s Blackwell architecture (3x compute density gap) and face CUDA ecosystem barriers. However, state-backed "algorithm-for-compute" strategies and Huawei’s F5G optical networks (reducing latency) aid the "East Data, West Computing" initiative.

Overview

This week highlighted a "global resonance" in GPU computing power: the Middle East accelerated AI infrastructure via energy and capital, Intel countered with new products, while China advanced through policy-driven tech breakthroughs and ecosystem reforms. The competition will hinge on chip performance, energy efficiency, and software-hardware synergy.