China’s first fully domestically produced 100,000‑GPU AI supercluster goes live
The National Development and Reform Commission announced that the country’s first domestically made AI supercluster with 100,000 GPUs was officially put into operation this week, marking a new stage in large‑scale deployment. Multiple computing nodes across the country have also been expanded accordingly.
Moore Threads reports 147.42% year‑on‑year revenue growth in H1 2026
On August 9, Chinese GPU maker Moore Threads released its semi‑annual report for 2026, posting revenue of RMB 1.736 billion in the first half, up 147.42% year‑on‑year and already exceeding its full‑year 2025 revenue. Net loss attributable to parent company narrowed by 95.73%. The same day, the company announced it is preparing for an H‑share listing on the Hong Kong Stock Exchange, pursuing a dual “A+H” platform.
AMD launches Instinct MI455X flagship AI GPU
At its “Advancing AI 2026” event, AMD unveiled the Instinct MI455X accelerator, featuring 320 billion transistors, 432GB of HBM4 memory, and FP4 compute of 40 PFLOPS, with certain specs surpassing NVIDIA’s Rubin architecture. It also introduced the Helios rack‑scale platform, integrating 72 GPUs to deliver up to 2.9 EFLOPS of compute power.
NVIDIA mulls cutting Rubin Ultra memory to ease HBM shortage
Industry media reported on August 6 that NVIDIA is considering reducing the HBM capacity of its next‑gen Rubin Ultra GPU from the originally planned ~384GB to 192GB, a nearly 50% cut, while also reducing the number of GPU chiplets from four to two, in order to alleviate tight supply of advanced HBM.
SpaceX partners with NVIDIA to send data‑center‑grade computing into space
On August 5, SpaceX announced a collaboration with NVIDIA to develop the Starmind AI1 satellite compute payload, with each satellite equipped with an NVIDIA Rubin GPU and Vera CPU, enabling data‑center‑class computing capabilities in space.
Meta launches GPU rental business
Meta officially announced on August 5 that it is commercialising its compute capacity by starting to rent out its own GPU resources to external customers, shifting its role from a major global purchaser of compute to a supplier, creating business‑level competition with NVIDIA.
Rental prices for NVIDIA H100 and B200 GPUs rise
Latest industry data show that hourly rental rates for NVIDIA H100 GPUs have increased by about 50% since January this year, from under $2 to nearly $3; B200 rental prices are also moving up, reflecting continued tight supply‑demand balance in the compute market.
Intel Arc Pro B70 adapts MiniMax H3 video model on day zero
On August 3, Intel announced that its Arc Pro B70 GPU has achieved day‑zero adaptation for the open‑source multimodal video model MiniMax H3. The card is based on the second‑gen Xe2 architecture, with 32 Xe cores and 32GB of GDDR6 memory, delivering peak AI compute of 367 TOPS.
Nine global cloud giants significantly raise capex guidance
Major cloud providers have recently released Q2 earnings and revised up their full‑year capital expenditure guidance. TrendForce projects that the combined capex of the top nine cloud vendors will exceed $886.7 billion in 2026, up nearly 90% year‑on‑year; the annual growth forecast for AI server shipments has been raised to nearly 31%.
China’s domestic GPU rental quotes remain generally stable
Data as of August 6 show that among six domestic GPU rental quotes, four were flat, one rose, and one fell. The A800‑80G increased by 6.86%, while the RTX 5090‑32G decreased by 2.08%; the other four remained unchanged, indicating overall stable pricing.