Global GPU Computing News【20261101】

Metax’s Q3 revenue rose 128.45% year on year, and its 10,000-card-class “Token Super Factory” entered production. In the first three quarters, revenue reached RMB 2.824 billion, up 128.45% year on year, and single-quarter revenue reached RMB 1.5 billion. The MXMACA software stack was upgraded to version 3.9.0. The 10,000-card-class compute cluster built with Morphling Intelligence entered the production environment, with the first batch of clusters completing the transition from readiness to business model launch in one to two days.

Moore Threads turned profitable on a single-quarter non-recurring-adjusted basis and plans a 100,000-card full-function GPU cluster with JD Cloud. Q3 single-quarter revenue was RMB 1.55 billion, up 1,773.9% year on year, and net profit attributable to parent after deducting non-recurring gains and losses was RMB 20.072 million. Its P/D-disaggregated heterogeneous inference solution based on the MTT S5000 has carried real token traffic from model vendors. The two sides’ cooperation covers training, inference and embodied intelligence scenarios.

Ascend 950DT chips were lit up in Shantou, and domestic high-end chips entered the token production link. On October 29, the Shantou Overseas Chinese Economic and Cultural Cooperation Pilot Zone disclosed that Ascend 950DT compute equipment completed rack installation and was lit up, connecting to the local token factory production line. The chip focuses on large-model training and high-throughput decoding, features large-capacity HBM and 4 TB/s memory bandwidth, and is designed for the KV Cache bandwidth bottleneck in token production.

An Ascend 950DT supernode cluster went online in Wuhan. On October 29, a liquid-cooled Ascend 950DT 1,024-card supernode cluster built by Hubei Ketou went online. A single node is equipped with 1,024 Ascend NPU accelerator cards, reaches 1 EFLOPS of FP8 compute, and has 256 TB of global unified memory. Inter-card interconnect provides terabyte-level bandwidth and microsecond-level latency, targeting trillion-parameter large-model training and agent inference. The 950DT 10,000-card intelligent computing cluster in Horinger, Inner Mongolia, also entered the final stage of joint commissioning.

Qingyang, Gansu, signed a RMB 5.5 billion domestic 100,000-card compute cluster. On October 27, Gansu Yisuan, a wholly owned subsidiary of Yitian Intelligent, Enflame Intelligence and the Qingyang Municipal Government signed a strategic cooperation framework agreement to jointly invest RMB 5.5 billion to build a domestic 100,000-card compute cluster. The construction scale is no less than 100,000 domestic compute cards, providing no less than 25,000 P of compute services, and the project is planned to be completed and put into use by the end of 2027.

Nvidia disclosed an OOI optical interconnect prototype, seeking to reduce electrical signal loss between GPUs. On October 31, Nvidia showcased an “Optics-on-Interposer” co-packaged optics technology prototype at SEMICON West 2026, integrating optical engines directly onto the silicon interposer carrying the GPU, bypassing traditional organic substrates and long-distance copper traces, and using dense wavelength division multiplexing to achieve 256 Gb/s throughput per fiber. The solution still faces thermal management and manufacturing yield challenges from co-packaging optical engines with high-power GPUs, and Nvidia did not give a mass production timetable.

AMD’s Helios rack completed installation and testing at Meta, and Meta will deploy customized Instinct MI450 GPUs. AMD announced that its Helios rack-scale AI infrastructure solution completed installation and testing at Meta and will be expanded. Meta, as an early architecture definition customer for Instinct MI450, will deploy AMD GPUs customized for its workloads. The two sides also plan gigawatt-scale AI infrastructure, and AMD will make a USD 1 billion equity investment in Meta.

Nvidia will invest USD 3 billion in Meta and expand cooperation on Vera Rubin and AI agent infrastructure. On October 27, Nvidia announced plans to subscribe to Meta shares for USD 3 billion, and the two sides will expand cooperation around Muse series AI agent models. Meta plans to expand deployment of the Nvidia Vera Rubin platform and Vera CPU, and introduce Spectrum-X Ethernet and BlueField-4 to improve network efficiency. The focus of cooperation extends from pure chip supply to end-to-end infrastructure optimization for AI agent workloads.

SpaceX completed about USD 40 billion in GPU equipment finance leasing, deepening the trend of debt-financed compute assets. On October 26, Musk said SpaceX had completed about USD 40 billion in debt financing for GPU procurement, structured as equipment finance leasing. Related entities issue debt to buy Nvidia AI chips and lease them to SpaceX, allowing SpaceX to secure compute without directly putting the debt on its own balance sheet. In the same period, Meta disclosed an equity financing plan of up to USD 95 billion, and Broadcom led USD 63 billion in financing to procure chips for Anthropic. AI infrastructure-related financing disclosed on a single day approached USD 200 billion.

HBM and advanced packaging supply constraints continued to become more visible, and Samsung’s 2027 HBM supply is sold out. Samsung Electronics said Q3 HBM4 sales tripled quarter on quarter, its 2027 HBM supply was fully sold out, and it plans to mass-produce HBM4E in the first half of 2027. SK hynix expects DRAM and HBM supply tightness to persist through around 2030, with HBM capacity effectively sold out through 2029. On the advanced packaging side, TSMC reportedly is considering raising CoWoS quotes by about 20%, and Anthropic also listed HBM shortages and TSMC capacity constraints as constraints on AI expansion in its IPO filing. The compute leasing market also showed signs of divergence: SpaceX’s large-scale financing contrasted with tightening financing for small and medium-sized neocloud companies, and Musk said more than 200 neocloud companies may face consolidation pressure.