Huawei details Ascend 960 SuperPoD, Anthropic targets 5 GW, CXMT plans a NAND line

Huawei puts near-package optics inside the Ascend 960 SuperPoD

At Huawei Connect 2026 in Shanghai on September 17, the company introduced the Ascend 960 SuperPoD, described as the first supernode built with near-package optics. The machine pairs Huawei's Lingqu interconnect with the Hi-ONE optical engine, a 7.2T module that the company says is the only NPO part with a light source built in. One supernode scales to 4,096 cards and delivers 8 EFLOPS of FP8 compute, or 16 EFLOPS at FP4.

Mounting optics close to the package cuts part count and power. Huawei says 5,500 Hi-ONE modules replace roughly 48,000 800G optical transceivers, bringing power draw down by more than 550 kW and doubling mean time between failures, with availability put at 99.8 percent. Multiple supernodes linked over Lingqu or RoCE reach 512,000 cards, and a multi-rail topology takes that to one million.

The event also brought Peerium, a computing architecture built on nested parallelism, unified memory addressing and peer-to-peer interconnect, replacing earlier master-slave designs with what it calls Nested BSP. Atlas 950 SuperPoD is the first product built on Peerium, a 256,000-card Atlas 950 cluster is being deployed, and the NPO-based Atlas 960 is in testing. "In the AI era, Huawei will keep building supernodes and clusters on the Peerium architecture to meet customer training and inference needs," rotating chairman Eric Xu said. Source: Huawei

Anthropic expects about 5 GW of compute by December

Anthropic has told investors it plans to reach roughly 5 GW of available computing capacity by the end of 2026, against about 1.5 GW last year, and about 10 GW by the end of 2027, according to people familiar with the numbers. The path tracks OpenAI, which people with knowledge of its plans put at around 5 GW this year and 10 GW next. Anthropic agreed to pay SpaceX about $1.25 billion a month for compute through May 2029 and is discussing a lease from Meta that could total $10 billion over two years. CNBC reported that both labs are also chasing 20 to 30 MW blocks in the UK, the Nordics and the US, since inference can run across separate sites and a powered site comes online faster than a new campus. Source: Sina Finance

CXMT prepares a NAND line in Beijing

ChangXin Memory Technologies is preparing to move into flash memory, Reuters reported on September 18, citing three people with knowledge of the plan. The DRAM specialist intends to set up an R&D production line for NAND at its new Beijing plant and has opened a research institute in the city with NAND among its projects. It has spoken with customers including a newly created startup that wants the chips for storage in AI systems and supercomputers; the person declined to name it. When the line starts running, and whether CXMT follows it with volume manufacturing, is not clear. TrendForce put Samsung's second-quarter NAND revenue share at 29.3 percent, ahead of SK Hynix and Micron. Source: TechNode, citing Reuters

GSMA links memory prices to a widening AI gap

The State of Mobile Internet Connectivity Report 2026, released in New York on September 15, puts 3.4 billion people outside mobile internet use even though more than 90 percent of them live inside mobile broadband coverage. Roughly 4.8 billion people use mobile internet on their own device, but about 160 million came online in 2025, down from 190 million the year before.

Handset affordability ranked as the biggest barrier in the low- and middle-income countries surveyed, ahead of digital skills. At the end of 2025 an entry-level internet-capable handset cost the poorest 20 percent of people the equivalent of 44 percent of average monthly income, rising to 76 percent in Sub-Saharan Africa. Citing Counterpoint Research, the report says memory prices more than doubled between the third quarter of 2025 and the first quarter of 2026, then rose another 80 to 90 percent in the second quarter as suppliers prioritize AI infrastructure. A $30 entry-level phone could have reached almost 1.6 billion people a year ago and a $20 one about 2.2 billion; the GSMA now calls both out of reach, with global smartphone shipments forecast to post their largest annual decline on record as the sub-$100 segment collapses. Source: GSMA

TSMC returns to Longtan for sub-1.4nm capacity

TSMC has decided to build at Longtan after all, supply chain sources told Cailian on September 18. The phase-three expansion is oriented toward nodes at 1.4nm (A14) and below, with three fabs sketched out and the first targeted for completion and production readiness around 2030. Total investment is estimated at more than NT$1 trillion, about $31.4 billion.

TSMC walked away from the site in October 2023 after local opposition to land acquisition; the Hsinchu Science Park administration approved a revised expansion plan on September 17 with the area cut to about 104 hectares from 158.59. Source: Cailian

onsemi raises prices across a broad range of parts

onsemi has notified customers of price increases covering a wide set of products, effective October 10, 2026, its second round this year after one in the first half. The company attributed the move to higher costs for raw materials, manufacturing, energy and infrastructure, and said demand has been rising across power, industrial and data center applications, leading it to invest further in manufacturing capacity, technology and operational resilience. Specific part numbers and increases will go out through authorized distributors. Source: Securities Times

Waymo sets 2028 for a Singapore robotaxi launch

Waymo said on September 18 that it will run a fully driverless, all-electric ride-hailing service in Singapore from 2028, its first market in Southeast Asia, working with the Ministry of Transport and the Land Transport Authority. An initial fleet of Jaguar I-PACE vehicles arrives in the coming months, and through 2027 trained specialists will drive them manually while the Waymo Driver is adapted to local road geometry and monsoon weather. Testing begins in western Singapore, around Labrador and Harbourfront. Waymo reports more than 20 million fully autonomous rides, with injury-causing crashes 94 percent below the rate for human drivers in the US cities where it operates. Source: Land Transport Authority

NASA and IBM open a lunar foundation model

NASA and IBM Research trained the NASA-IBM Lunar Foundation Model from scratch on SomBench, a multimodal dataset of close to two million co-registered bundles across 11 modalities. Most of the training data comes from 17 years of Lunar Reconnaissance Orbiter observations: about two million image tiles, including more than one million narrow-angle camera images at 1-meter resolution and nearly 964,000 multispectral images at 100 meters, along with GRAIL gravity, Lunar Prospector and JAXA SELENE data.

The model was tested on crater detection at regional and meter scales, segmentation of irregular mare patches, and regression of lunar polar ice prospectivity. It matched or beat ImageNet-pretrained comparison models on all three, with strong label efficiency on crater detection, where a benchmark covers more than 49,000 manually identified craters. Weights, fine-tuning code and benchmark datasets are public, and the model is on Hugging Face with code on GitHub. Source: USRA