Global GPU Computing News【20260705】

1. Meta Plans Cloud Computing Business to Lease Idle AI Compute Capacity

Meta announced it is building a cloud computing business called "Meta Compute" for external customers, planning to lease underlying GPU compute capacity directly from its data centers or provide AI model services via APIs. The news sparked market concerns about "compute overcapacity," causing sharp drops in shares of CoreWeave and other cloud service providers, as well as memory chip makers. However, some analysts believe the issue is not absolute overcapacity but a structural mismatch, with high‑end intelligent compute still in short supply.

2. NVIDIA Introduces "Revenue‑Share" Model to Support AI Startups

NVIDIA unveiled a revenue‑sharing and credit support model to help AI cloud service providers obtain large‑scale infrastructure, which they can then offer to startups and model builders. Instead of leasing GPUs directly to startups, NVIDIA provides GPUs, software platforms, and credit lines. The first two partners plan to deploy up to 210,000 GPUs combined, with one in Australia and the other in Indonesia.

3. Kingsoft Cloud Accelerates GPU Build‑out; Xiaomi Boosts Budget, Alibaba Signs Five‑Year Contract

Kingsoft Cloud announced it will accelerate its GPU computing cluster construction in the second half of the year. Xiaomi's demand for Kingsoft Cloud's compute power has been upgraded to a super‑large‑scale cluster, with investment budget increased to over RMB 10 billion. Kingsoft Cloud raised its 2026 capital expenditure plan to RMB 15 billion. Separately, Alibaba's large‑model team signed a five‑year compute lease contract with Kingsoft Cloud involving more than 3,000 eight‑GPU servers.

4. Meta and AMD Reach Five‑Year, 6‑Gigawatt Compute Capacity Agreement

Meta and AMD signed a five‑year agreement covering 6 gigawatts of compute capacity. Meta will purchase multiple generations of AMD Instinct GPUs, while AMD issued performance‑based warrants to Meta. AMD's CEO stated that the collaboration is "very special" and that customer engagement with MI450 and Helios has exceeded initial expectations.

5. AWS Raises Prices for AI Compute Capacity Reservations by Approximately 20%

Starting July 2026, Amazon Web Services will increase prices for its EC2 Capacity Blocks for ML reservations by about 20%. This service allows enterprises to lock in GPUs in advance, primarily for large‑model training and fine‑tuning tasks that cannot be interrupted.

6. GPU Lease Prices Continue to Rise; A100/H100/B200 Show Multi‑Month Increases

June data shows that GPU lease prices in the non‑hyperscaler market continued to climb. Average lease prices for A100 rose 6.3% month‑over‑month to $1.63 per GPU‑hour, marking the fifth consecutive monthly increase; H100 rose 3.7% to $2.72, the seventh straight monthly rise; B200 rose 2.7% to $5.33. Separately, monthly rental rates for the RTX 5090 are concentrated in the range of RMB 11,500 to 12,000.

7. Iluvatar CoreX Secures Order for 50,000 AI Inference Chips from ByteDance

ByteDance is in talks to purchase at least 50,000 AI inference chips from Iluvatar CoreX, mainly from the Zhikai series of cloud‑side inference GPUs. In ByteDance's compute strategy, Huawei Ascend and Cambricon handle heavy pre‑training workloads, while Iluvatar CoreX targets high‑volume consumer‑facing inference scenarios. Domestic chips now account for 41% of China's AI server GPU market.

8. CETC 58th Institute Achieves Major Breakthroughs in Intelligent Computing Chips

The 58th Research Institute of China Electronics Technology Group Corporation made significant progress on several core intelligent computing chips. The ZQ300 series AI SoC completed key technology validation and successfully taped out, while the ZQ500 general‑purpose GPU has entered mass production. The ZQ300 features a heterogeneous computing architecture combining parallel computing frameworks and AI acceleration units; the ZQ500 offers strong compute performance and excellent power control, suitable for unmanned systems, radar systems, and smart terminals.

9. Jingjia Micro Discloses Progress on Next‑Generation GPU Architectures

Jingjia Micro revealed on its investor interaction platform that its high‑performance general‑purpose GPU R&D and industrialization project is developing two chip models. Model 1 will adopt a new GPU architecture with significantly improved rendering and general‑compute capabilities, targeting graphics workstations and cloud rendering workstations. Model 2 will develop a supporting software stack compatible with inference frameworks and mainstream training, enabling an efficient integrated inference‑training compute model.

10. State Council Meeting Outlines AI Development; Eight Ministries Promote Compute Infrastructure Construction

On July 1, Premier Li Qiang chaired a State Council executive meeting to review AI development progress, emphasizing the need to accelerate key technology breakthroughs and the construction of ultra‑large‑scale intelligent computing clusters. Separately, eight ministries including the Ministry of Industry and Information Technology issued implementation opinions on promoting high‑quality development of the industrial internet, calling for integrated planning and simultaneous construction of industrial internet infrastructure with intelligent and supercomputing facilities.