Global GPU Computing News【20250712】

1. CoreWeave Deploys World’s First NVIDIA GB300 Supercomputing System

   • On July 10, CoreWeave announced the commercial deployment of NVIDIA’s GB300 NVL72 platform, becoming the first cloud provider to adopt the system. It integrates 72 Blackwell Ultra GPUs, 36 Grace CPUs, and 36 BlueField-3 DPUs, using liquid cooling. Peak inference performance reaches 1.1 ExaFLOPS, with 50% higher training efficiency and multi-fold improvements in bandwidth and energy efficiency.

2. AWS Launches Most Powerful Blackwell Instance P6e-GB200

   • On July 10, AWS unveiled the EC2 P6e-GB200 UltraServer instance based on GB200, equipped with 72 Blackwell GPUs per node. It delivers 360 PetaFLOPS FP8 performance, 13.4TB VRAM, and 28.8 Tbps network bandwidth—20× faster than previous gen. Also released: 8-GPU P6-B200 instances for mid/large-scale AI training.

3. NVIDIA Blackwell Ultra Chips Enter Commercial Use, 50× Performance Leap

   • NVIDIA and CoreWeave confirmed commercial deployment of Blackwell Ultra chips, offering 50× generative AI capability over prior Blackwell chips. Mass production begins Q4 2025, boosting NVIDIA’s market cap near $4T and cementing its AI chip dominance.

4. CoreWeave Acquires Core Scientific for $9B, Secures 1.3GW Power

   • To reduce data center reliance and costs, CoreWeave acquired Bitcoin miner Core Scientific, gaining >1.3GW power capacity. Expected annual savings: $500M. This vertical integration (chip-to-power) supports 10,000+ GPU cluster expansion.

5. Microsoft Delays In-House AI Chips, Boosting Blackwell Demand

   • Microsoft postponed its Maia 200 AI chip to 2026 and high-end Braga-R/Clea chips beyond 2028. Delays may intensify reliance on NVIDIA Blackwell and expand third-party cloud providers’ (e.g., CoreWeave) market share.

6. Oracle Signs $30B Compute Deal Targeting OpenAI & Sovereign AI

   • Oracle secured a $30B/year cloud agreement (effective FY2028), reportedly for OpenAI workloads or Middle East’s G42 project. Reflects rising competition from AI cloud vendors and sovereign-state compute investments.

7. Moore Threads Files for STAR Market IPO, Advances Domestic GPU Commercialization

   • Moore Threads applied for a STAR Market IPO, promoting its "full-featured GPU" approach. Its MTT S5000 AI card supports 1,000-card clusters with efficiency surpassing global peers, already powering models like DeepSeek. AI revenue share: 77.6% in 2024, 208% CAGR over 3 years.

8. WiMi Hologram Focuses on Edge AI Chips to Lower SME Costs

   • WiMi Hologram (WIMI.US) develops low-power edge AI chips, integrating algorithms for real-time use (smart manufacturing, autonomous driving). Aims to cut AI adoption barriers for SMEs via high-compatibility solutions.

9. China SatNet Reshuffles Leadership, Accelerates LEO Compute Infrastructure

   • New Chairman Gou Ping and GM Liang Baojun took office at China SatNet, with board members from aerospace/telecom sectors. Accelerates space-ground integration and NTN tech commercialization, boosting satellite-ground infrastructure and optical module demand.

10. GB300 Drives Optical Module & Material Upgrades; 1.6T Optics/Copper Cables Surge

    • GB300 adopts 1.6T optical modules, HVLP5 copper cables, and PTFE PCBs, with liquid cooling redesigns for efficiency. TrendForce forecasts >1M GB200 shipments in 2025, GB300 volume production in Q4—propelling optical/thermal supply chain innovation.

Key Events Overview