Global GPU Computing News【20261011】

Nvidia’s market cap approaches $6 trillion, with DGX Spark and Rubin Ultra advancing on multiple fronts. Nvidia’s stock hit a record high this week, bringing its market cap back to $5.8 trillion, just short of $6 trillion, ending a two-month pullback that had erased more than $1 trillion. On the hardware side, the DGX Spark 64GB version will go on sale on October 23 starting at $4,999; a single unit can run a 100-billion-parameter model, while two units can support 200 billion. Rubin Ultra is confirmed to ship next year as planned, with comprehensive optimizations in cooling, PCB, and HBM design. The OCP Global Summit concurrently focused on AI interconnect, 800VDC power delivery, and liquid cooling solutions.

RTX Spark laptops open pre-orders, bringing local AI agents to Windows PCs. On October 7, Microsoft and Nvidia unveiled the RTX Spark laptop series in San Francisco, with the first products going on sale on October 16. The platform uses a Blackwell RTX GPU and Grace CPU, with up to 6,144 Blackwell RTX cores, a 20-core Arm CPU, 128GB unified memory, and about 1 PFLOP of FP4 AI compute, and can run models with more than 120B parameters locally. Acer, Asus, Dell, Lenovo, and other manufacturers will launch related products. Microsoft’s Surface Laptop Ultra starts at $2,599.

AMD’s Lisa Su says AI chip demand far exceeds supply, plans major capacity expansion in 2027. AMD CEO Lisa Su said on October 6 that the company has already increased supply and capacity in 2026, but demand is even higher than supply, and it plans to significantly increase supply in 2027. AMD’s market cap recently crossed $1 trillion, and its stock has risen about 30% over the past month. Su also warned that memory chip supply is tight, and the company is actively coordinating to secure the memory volumes needed for AI servers. The industry expects AMD to launch a new wave of memory procurement.

Biren Technology’s BR20X successfully lights up, with commercial mass production of a domestic GPU imminent. On October 8, Biren Technology disclosed that its next-generation AI chip BR20X has successfully lit up and completed startup testing, entering the post-tapeout verification stage, faster than previously expected. BR20X is based on a self-developed instruction set and Chiplet architecture, supports FP8/FP4 low-precision computing, and features the self-developed BLink 2.0 interconnect protocol, supporting up to 1,024-card Scale-up expansion. The company also raised about HK$4.02 billion through a placement, about 70% of which will be used for supply chain procurement and commercialization of next-generation products.

Huawei Ascend surpasses Nvidia in China market share, with super-node deployment reaching 256,000 cards. Huawei Rotating Chairman Xu Zhijun said at HC 2026 that, based on available statistics, Ascend AI chips have surpassed Nvidia in China market share. According to Mysteel statistics, Ascend’s domestic AI accelerator share has reached about 50%, while Nvidia’s is about 8%. Huawei also released the Peerium computing architecture and Lingqu unified bus solution. The Atlas 950 SuperPoD super-node has reached a deployment scale of more than 256,000 cards and is undergoing cluster testing.

China Mobile’s full-stack domestic computing power center lands in Beijing, with Ascend 950DT thousand-card super-node online. China Mobile Beijing Branch and China Mobile Cloud announced that the Huawei Ascend 950DT thousand-card super-node has officially landed at the Changping Information Port Data Center. A single node with 1,024 cards delivers 1 EFLOPS (FP8), supports 256TB of unified memory, and microsecond-level latency. China Mobile Cloud has completed compatibility with mainstream domestic AI chips including Ascend, T-Head, Kunlunxin, Moore Threads, Biren, Hygon, and MetaX, enabling unified access and management of diverse computing power.

AI memory shortage to last for years, Nvidia Rubin GPU specs forced to be downgraded. A Morgan Stanley report points out that the memory shortage has become the most persistent structural constraint in this AI buildout cycle. Nvidia’s Rubin architecture single-GPU HBM capacity has been cut from a planned 288GB to 192GB, and per-rack LPDDR5 capacity has been compressed from 54TB to 28TB. Morgan Stanley believes that model sizes doubling every six months, context windows expanding about fivefold annually, and continuously rising inference concurrency are three structural factors that make the downgrade only a temporary measure; the memory bottleneck will shift downstream to NAND and network interconnect.

Bernstein breaks down AI infrastructure costs: up to $39.5 billion investment per GW data center. A Bernstein report released on October 5 shows that building a 1GW data center using different AI accelerator architectures requires capital expenditure of about $34.6 billion to $39.5 billion. Nvidia’s Rubin architecture has the highest construction cost, AMD Helios is about $35.7 billion, Google TPU v7 is about $39 billion, and OpenAI’s custom ASIC is about $34.6 billion. The expected cost of Nvidia’s Rubin NVL72 per rack has been lowered from $9.1 million to $7.52 million, a drop of about 17%, mainly reflecting adjustments to HBM price and NAND storage capacity expectations.

Central bank includes computing power networks in PSL support, local computing power policies land intensively. The People’s Bank of China adjusted and improved monetary policy tools, including computing power networks along with new-type power grids and next-generation communications networks—the “six networks”—in the pledged supplementary lending (PSL) support area. The PSL interest rate was cut by 0.25 percentage points, and the sci-tech innovation relending quota was increased to 1.4 trillion yuan. At the local level, Chengdu High-tech Zone’s computing power policy took effect on October 8, using computing power vouchers and model vouchers to reduce enterprise usage costs; Hong Kong’s Sha Ling data park broke ground, and the AI funding scheme was injected with HK$1 billion.

Nvidia-backed Lambda plans to raise $4 billion for IPO, GPU computing power leasing financing model accelerates. Nvidia-backed cloud startup Lambda is raising up to $4 billion, led by Blackstone and Coatue, at a pre-money valuation of $14.5 billion, and plans to list in 2027. Lambda’s AI computing power backlog has surged from $15 billion in June this year to $50 billion. Meanwhile, Amazon is considering transferring about $8 billion worth of Nvidia GPUs into a special purpose entity and leasing them back, as innovative financing models in the GPU computing power leasing sector accelerate.