Terafab Storage Capacity Targets Samsung Scale as Google Expands TPU Lineup

Highlights

• Terafab plans 1 million wafers per month by 2032, 700,000 for memory
• Google releases two new TPUs, AMD EPYC provides host infrastructure
• Domestic compute adaptation: LongCat-2.5 deeply optimized for Chinese chips

Manufacturing and capacity: Terafab sets its sights on Samsung's scale

Musk's semiconductor project Terafab now has a more concrete capacity timeline. According to Nikkei, Terafab has told semiconductor equipment suppliers it plans to reach 1 million wafers per month by 2032, with 700,000 for memory chips and 300,000 for logic chips. For reference, Samsung's current DRAM capacity is about 650,000 to 700,000 wafers per month, meaning the newcomer's memory capacity target directly matches Samsung's current scale. (Source: ITHome)

On timing, Terafab, a joint venture between SpaceX and Tesla, plans volume production in the third quarter of 2028. Musk previously said the internal target is to produce the first test chip using SRAM circuits at its Austin R&D base around May or June next year. The team also plans to build and test about 10 chips a year at the R&D facility, using faster test iterations to compress development cycles. (Source: ITHome)

A measured view is warranted: from the first test chip to 1 million wafers per month in 2032 lie equipment delivery, process ramp-up and yield validation. Replicating Samsung's current scale within six years is a hard constraint for any new player.

Compute chips: Google expands TPU lineup, domestic adaptation advances

Competition in AI compute continues to heat up. Google launched two new TPU chips, intensifying its rivalry with Nvidia; AMD said its fifth-gen EPYC processors provide host infrastructure for Google's TPU 8tx. (Source: TMTPost)

Model-chip collaboration is also tilting toward domestic compute. Meituan's LongCat open-sourced LongCat-2.5, with 1.6 trillion total parameters, about 48 billion activated per token, and support for a one-million-token context. The team said it made deep collaborative optimizations in model architecture, chip adaptation and deployment strategy for domestic compute chips constrained by memory and bandwidth, and adopted a W4A4 inference scheme that quantizes both expert weights and activations to 4 bit. (Source: ITHome)

On cloud scheduling, Alibaba Cloud's patent for "a method and device for processing user requests" won the gold award at the 26th China Patent Awards. The technology targets on-demand allocation and multi-tenant isolation for hyperscale intelligent compute. Alibaba Cloud said its global patent portfolio exceeds 10,000 filings, covering cloud computing, AI, large models and chips. (Source: Leiphone)

Device chips: 6nm, GaN and cockpit-driving integration land

Consumer and automotive devices continue to absorb mature-node and application-specific chip solutions. Edifier's W900NB 30th Anniversary Edition uses a 6nm six-core heterogeneous chip with an integrated Arm Ethos U55 dedicated NPU for end-to-end AI pickup noise reduction, priced at 1,159 yuan in a limited run of 3,000 units. (Source: ITHome)

Colorful's iGame Sigil 100W GaN charger carries Southchip Semiconductor's new-generation GaN co-packaged controller chip, supports the PD 3.2 AVS protocol, and is priced at 129 yuan for first-day buyers. (Source: ITHome)

On the automotive side, Wey's second-gen Blue Mountain packs Qualcomm's new-generation 8797 cockpit-driving integrated flagship chip and Nvidia DRIVE Thor-U driver-assistance chip; Arcfox Beta S3 uses Qualcomm's Snapdragon 8155 cockpit chip and Horizon Journey J6M smart-driving chip. (Source: ITHome)

Other news

• AMD graphics drivers were reported to cause sustained SSD writes, with some users reporting write speeds of 5MB per second, or 432GB over 24 hours; the company has opened an investigation. (Source: ITHome)
• The top 15 of the first Ant Lingbo Embodied Large Model Challenge were announced; preliminary-round GPU input rose from 200 to 350 cards, providing over 340,000 cumulative card-hours of compute. (Source: Leiphone)