Micron 512GB DDR5, NVIDIA-powered Koa and China server plans | Daily Silicon

01 Micron’s 512GB DDR5 module makes server slots more valuable

TL;DR Micron demonstrated a 512GB DDR5 module and says AMD and Intel are validating it. Volume production is expected in the second half of 2027, so this is a development milestone rather than a shipping announcement.

The company describes up to 12TB in a 24-slot server, using TSV-stacked DRAM. This expands host memory capacity; it does not add GPU HBM.

The power comparison is 16.0W for one 512GB module versus 44.2W for four 128GB modules, not whole-server consumption. Platform qualification and pricing still matter alongside the saved slots.

Sources · Micron (2026-09-15)

02 Xingyun proposes a CNY20bn server budget, not a completed order

TL;DR Xingyun Technology (300209, 行云科技) proposes raising its annual server and component procurement ceiling from CNY7bn to CNY20bn. Shareholder approval is still required; neither completed purchases nor secured financing equal to that ceiling have been established.

The board decision is dated September 15, while the CNINFO disclosure path is September 16. Funding may include internal resources, raised funds and finance leases for AI computing and digital-trade operations.

The ceiling would be about 2.86 times larger. Translating it into chip demand requires supplier, configuration, financing and delivery details; the filing warns of nonexecution, funding shortages and higher finance costs.

Sources · 行云科技 CNINFO filing (2026-09-15)

03 Salesforce Koa adapts NVIDIA’s model to CRM work

TL;DR Salesforce announced Koa, a CRM reasoning model post-trained from NVIDIA Nemotron 3 Super. It is in selected customer pilots, with US general availability expected in winter 2026.

Salesforce says post-training uses synthetic data, not customer data. Control over weights, post-training and inference within its own boundary is part of the enterprise proposition.

The semiconductor signal is specialized inference moving toward operation. There is no disclosed GPU purchase quantity in this announcement, so a model launch cannot be treated as a hardware order.

Sources · Salesforce (2026-09-15)

04 CanSemi targets CNY7.5bn in its next IPO stage

TL;DR CanSemi issued its preliminary prospectus after registration consent, targeting CNY7.5bn. Initial price inquiry is scheduled for September 18 and subscription for September 24; the offer price and listing date are not fixed.

Planned allocations are CNY3.5bn for a 12-inch specialty analog line, CNY2.5bn for process-platform R&D and CNY1.5bn for working capital. Project investment and the IPO contribution are separate figures.

This is a capital-allocation signal for specialty processes, not simply another leading-edge logic race. The funding target is not cash already received.

Sources · CanSemi preliminary prospectus · Issuer and underwriter notice (2026-09-15)

05 MACOM targets 448Gbps lanes for 3.2T optics

TL;DR MACOM plans to demonstrate components for 3.2T optical links at ECOC on September 21–23. The announcement includes 448Gbps-per-lane drivers, TIAs and photodetectors; it does not establish completed production qualification.

Lane rate is distinct from aggregate module rate. Faster links need the transmitter, receiver and optical devices to work together.

The planned exhibits also include 1.6T active copper at 224Gbps per lane. Copper and optics therefore remain a distance, power and cost allocation problem, rather than an overnight replacement.

Sources · MACOM (2026-09-15)

06 Bulten’s Exim wins a supply role in Lam’s Asian manufacturing

TL;DR Bulten subsidiary Exim announced a C-parts agreement for Lam Research manufacturing in Malaysia, Taiwan and South Korea. Deliveries are expected to start in 2027 and ramp gradually, with annual business volume above $10m anticipated at full implementation.

C-parts include fasteners: inexpensive individually, but still requiring sourcing and inventory control. The deal concerns equipment manufacturing, not wafer capacity.

The expected annual scale belongs to the fully implemented operation. It is not revenue already earned or a number to add to the current year without qualification.

Sources · Bulten company release via GlobeNewswire (2026-09-15)

07 Tower pairs silicon photonics with SiGe process offerings

TL;DR Tower Semiconductor plans to present silicon-photonics and SiGe BiCMOS platforms at ECOC. Its emphasis is manufacturing technology for transceivers and NPO/CPO optical connectivity.

Where MACOM discusses high-speed components, this announcement concerns the process choices behind them. Optical devices and high-speed analog circuitry do not share identical optimization constraints.

Tower highlights high-volume platforms, but the announcement establishes no new order quantity or capacity increment. Design teams still need a suitable process, packaging route and manufacturing support.

Sources · Tower Semiconductor (2026-09-15)

08 India’s semiconductor build-out also needs test capacity

TL;DR Advantest plans to show SoC and memory test products at SEMICON India in Delhi on September 17–19. The exhibition offers a view of the test capabilities needed alongside manufacturing investment.

The planned lineup includes V93000 EXA Scale and the T5800 family, addressing different requirements across AI/HPC, power, analog and memory.

Exhibiting is not evidence of an Indian equipment order. As manufacturing expands, test-program development, test time and yield-data use become practical questions for the local ecosystem.

Sources · Advantest (2026-09-15)

09 Samsung’s fab-management practices reach a public nanofab

TL;DR Korea’s science ministry reported results from Samsung’s operational consulting for POSTECH’s National Institute for Nanomaterials Technology. The program addresses on-site, safety and equipment management, with wider public-nanofab rollout planned.

Catch-up on the 2026-09-11 announcement, not a new event today.

It follows the March 2025 government-industry agreement. This is an operational improvement effort, not an announcement transferring a leading-edge manufacturing process.

For researchers, reliable equipment and timely process access matter alongside the equipment list. The release provides no quantified yield improvement, so numerical impact should not be assumed.

Sources · 과학기술정보통신부 (2026-09-11)