구글의 실제 데이터센터에서 비용,전력 절감 효과가 검증된 OCS가 다른 하이퍼스케일러로 확산된다면 OCS는 특수 장비가 아니라 AI 데이터센터의 새로운 광학 스위칭 계층으로 자리 잡을 수 있다. $COHR $LITE

Broadcom and TOPPAN Open Singapore’s First Advanced Chip Substrate Plant U.S. and Japanese companies seek to ease the substrate shortage constraining AI growth September 29, 2026, 1:15 p.m. JST Photo: Charlie Kawwas (center left), president of Broadcom’s Semiconductor Solutions Group, explains key AI data center materials and computing chips to Singapore Deputy Prime Minister and Minister for Trade and Industry Gan Kim Yong at the new facility. Photo by Cheng Ting-Fang. SINGAPORE — U.S. chipmaker Broadcom and Japan’s TOPPAN Holdings opened Singapore’s first advanced semiconductor package substrate plant on September 29, several months ahead of schedule. The opening comes as Singapore works to strengthen its position in the global AI supply chain. #Broadcom #TOPPAN

Amazing @PhotonCap! This is the part the market is still treating as a delay. This is very bullish for companies like $BESI $AMAT $TSM $ASMPT… SEC filings stay quiet on process names until the yield is real. Forums talk when customers are already in the queue and that’s usually too late… HBM hybrid bonding slipping toward HBM5 was priced as “bonders wait.” DRAM stacked straight onto the logic die is a second line of demand that opened while that wait was happening. That’s the interposer starting to disappear. The bonders don’t care if the line sits in Icheon or Hsinchu. They care who owns the last 10 microns when the stack moves from next to the GPU to on top of it. IMO, that’s the map! Great post my friend!

- $SKHY plans to fund up to 50% of suppliers’ initial R&D costs upfront. - Failed projects or those without purchases will still receive settlement based on technical achievements and contributions. - Support includes production-line validation access and low-interest financing for mass production. - The 50% upfront funding policy was already disclosed in July; this announcement reiterates that direction. > Customer participation in suppliers’ development risk could accelerate commercialization of next-generation materials and equipment. > Investors should track qualification milestones and conversion into production orders more closely than funding amounts.

I went through every page SK hynix filed with the SEC for its US listing. The words "hybrid bonding" never come up. Then, two months on, the keynote at the company's own forum said customers began lining up this spring for one structure: DRAM stacked straight onto a logic chip. What the free section covers: The roundtable answer on why HBM stack counts are rising more slowly, and how far to trust the math that says 4-high is enough Where the hybrid bonding date that two memory makers both named in 2024 has ended up Why DRAM on logic is a much harder step than moving the HBM4 base die to a foundry Gathering the public cases in one place, I found that who ends up bonding hinges on something other than whether it's a foundry or a memory maker. Where that line falls also decides where the bonder orders go. If this is a second source of demand that opened while HBM bonding kept slipping, whose fab do the bonders end up in?

“삼성전자, 2나노 파운드리 수율 60% 근접”… 전성비 개선에 총력전 - 조선비즈 share.google/nMpBbuwpj2Qfxhf…

S’il y a bien un mot qui décrirait mon sentiment à l’égard de l’Europe suite à la conférence FOTS (par ailleurs, merci à l’équipe de Huawei pour l’invitation ), c’est: gâchis Non parce qu’on a clairement des gens talentueux et qui ont l’envie de bien faire, qu’on pourrait avoir une avance technologique mais qu’on manque de tellement de choses pour réussir… Premièrement, on manque d’ambition, et ça saute aux yeux quand on écoute des Européens puis des Américains ou des Chinois… Deuxièmement, on manque cruellement de capital… On cherche la rentabilité trop vite et on essaie de vendre pour notre petit marché plutôt que de viser plus grand. Et c’est probablement là le pire: notre marché final est trop petit. Un « géant » du cloud Européen comme OVH c’est même pas un nain face aux hyperscalers. OVH c’est 4k GPUs sur toute leur flotte, c’était la taille du cluster standard chez AWS en 2020… Et le pire c’est que pour y remédier. On va faire la manche à des gouvernements qui s’empressent de dépenser des fonds publics dans n’importe quel projet. On finit donc avec des entreprises zombies qui n’ont eu pour but que de récolter des subventions. Non, il faut simplifier et encourager le venture capital, encourager l’entrepreunariat et l’investissement de fonds privés. Et non créer de la demande artificielle venant des gouvernements tomshardware.com/tech-indust…

- Samsung's "vertical stacking" patent is not necessarily advanced packaging. - It vertically stacks superconducting waveguide/resonator, dielectric, and shielding layers on the same substrate. - That is closer to multilayer thin-film fabrication or BEOL 3D integration than HBM-style die stacking. - Advanced packaging becomes more relevant when separate qubit, cryogenic-control, or interposer dies are bonded together. > The key distinction is not whether the structure is vertical, but what is being stacked. > This patent does not yet imply a direct read-through to hybrid-bonding equipment such as $BESI. > Thin-film deposition, etch, lithography, and metrology are currently the more direct manufacturing links. > As quantum systems scale toward multi-die 3D integration, advanced packaging could eventually become a major bottleneck. 삼성전자, 양자컴 메모리도 '수직 적층'…한국·미국 특허 등록 share.google/3aulj0k6OiEu5Q9…

HOM in space!! 우주서 양자컴퓨터 가동 성공…"지구 궤도서 데이터 처리" - 지디넷코리아 share.google/0E6Ek9jFyXBDLPw…

- CXMT plans RMB 34.9B of investment: RMB 24.1B for R&D and RMB 10.8B for DRAM back-end testing. - The testing project will build a dedicated facility for DRAM chip testing and memory-module assembly over 37 months. - CXMT explicitly warned that external OSAT test capacity and related investment could become insufficient as its output and product mix expand. - RMB 8.18B, roughly 76% of the testing project, is allocated to equipment purchase and installation. > More wafer capacity alone is not enough. Test throughput must scale with DRAM output. > As HBM and DDR5 increase product complexity and SKU count, testing is becoming a structural memory CAPEX bottleneck. > This is directionally positive for the memory test/probe ecosystem including $FORM and Advantest, but CXMT has not disclosed equipment suppliers. > CXMT is also vertically integrating testing into its IDM model, treating test capacity as part of manufacturing capacity rather than a simple outsourced back-end step.

- $AVGO and Toppan's JV, Advanced Substrate Technologies, is opening Singapore's first advanced FC-BGA substrate manufacturing facility. - The plant targets large-body, high-layer-count FC-BGA substrates for AI accelerators and network switches. - Investment is expected to exceed ¥100B, with more than 300 engineers and technicians planned. - It is Toppan's first FC-BGA production site outside Japan, creating a dual-site structure with Niigata. - $AVGO is Toppan's largest FC-BGA customer and plans to begin using the Singapore capacity in FY2027. - Hock Tan explicitly identified substrates as an AI supply-chain bottleneck that the Singapore capacity should help address. > The key point is that $AVGO now treats substrates as a strategic constraint on AI XPU scaling, not merely a purchased component. > $AVGO is increasingly controlling critical supply from custom AI silicon and networking to InP lasers and substrates. > This is not automatically positive for incumbent FC-BGA suppliers, as captive/JV capacity could reduce $AVGO's external sourcing dependence. > AI packaging bottlenecks are expanding beyond CoWoS and HBM into large-body FC-BGA substrates.

Citi sees an $11B OCS market by 2030, with $LITE and $COHR capturing a combined 80% of the merchant market. I recently wrote a deep dive into the technology and economics behind this opportunity. Why can slow MEMS mirrors be fast enough for AI training? Why are vendors converging around 300 ports? Who is moving beyond $GOOGL? I followed the papers and patents through cost savings, next-generation SiPh competition, and listed-company OCS exposure. You've seen the market forecast. Here's the technology and demand behind the numbers. 👇

As CPO scales, laser thermal management and serviceability become increasingly valuable. $LITE announced plans to demonstrate an eight-wavelength DWDM external laser module, or ELSFP, at ECOC 2026. Initial availability is targeted for H1 2027. The architecture keeps lasers outside hot compute and switch packages, allowing the light source to be replaced independently. Multiple wavelengths also help increase bandwidth density where fiber routing and package space are constrained. For investors, the interesting angle is $LITE’s opportunity to expand from laser chips into complete external light-source modules. Commercializing CPO requires designs that address heat and maintenance alongside bandwidth. This announcement does not establish that heat-related laser failures or replacements have been common. $LITE disclosed no historical failure rate; it described separating lasers from the hot CPO/NPO environment. - Thermal management: Moving lasers away from hot ASICs makes cooling easier; the objective is broader than minimizing optical loss. - Serviceability: If a laser fails, the light source can be replaced without disturbing the expensive compute or switch package. - CPO: Optical engines share a package with the ASIC; the lasers do not necessarily have to be integrated there. - “Pluggable” in ELSFP refers to a removable laser source, rather than a conventional removable optical transceiver. > This architecture combines CPO/NPO optical engines with external pluggable lasers. CPO and pluggable light sources work together. > The investment opportunity is supplying external light sources for new CPO/NPO deployments, rather than replacement demand driven by frequent failures.