TSMC Preparing for Full-Scale Mass Production of "Panel-Level Packaging (PLP)" Semiconductors TSMC is set to go head-to-head with Samsung Electronics using "panel-level packaging (PLP)," a next-generation semiconductor packaging technology. PLP can significantly improve the productivity of AI chip manufacturing, and with TSMC hurrying to ready mass production, a contest for leadership with Samsung Electronics—which entered the market first—looks inevitable. According to industry sources on the 15th, TSMC is building out a materials, components, and equipment (MCE) supply chain to establish its PLP mass-production system. It is currently in discussions with domestic and overseas MCE companies on equipment investment. TSMC is reported to be planning to begin PLP mass production as early as next year, and this is read as a move in earnest toward that goal. PLP is a technology in which a semiconductor wafer with circuitry already formed is cut into individual chips (dies) and then packaged on a rectangular panel to produce the finished product. It contrasts with "wafer-level packaging (WLP)," which is performed on a round wafer. When chips are packaged on a circular wafer, the edge regions cannot be completed into chips and must be discarded—meaning lower productivity. Running the process on a rectangular panel instead allows chips to be produced with no wasted area. Based on a standard 600×600 mm rectangular panel, roughly five to six times as many chips can be produced compared with the mainstream 300 mm (12-inch) wafer. The company holding the upper hand in PLP technology is Samsung Electronics. After acquiring the PLP business from Samsung Electro-Mechanics in 2019, it has built up technical capability by applying PLP to mobile application processors (APs) and power management ICs (PMICs). TSMC, by contrast, had been passive on PLP, given that it had secured its foundry competitive edge with conventional WLP. But the situation reversed as the AI chip market grew explosively—PLP can increase AI chip output and is also advantageous for realizing large-area AI chips. Accordingly, TSMC began pursuing the PLP business from 2024. It is expected to build and run a pilot production line this year and, following performance evaluation, enter large-scale production around next year. It is also reported to have already secured a global AI chip customer. As TSMC accelerates toward PLP mass production, competition with Samsung Electronics is expected to intensify further. Samsung, too, plans to expand PLP application beyond its existing APs and PMICs to high-performance computing (HPC) chips such as AI semiconductors. Glass substrates, which are drawing attention as a substrate for AI chips, are also likely to be adopted in this PLP process—a point that suggests a leadership battle between Samsung Electronics and TSMC in the next-generation substrate market as well. An industry official said, "Not only Samsung Electronics and TSMC, but also global outsourced semiconductor assembly and test (OSAT) companies have jumped into the PLP process market in large numbers," adding that "along with fierce competition, market growth is expected."

PicoJool announced 200G VCSELs today. And it’s compelling. The company is sampling quad 100G, quad 200G, and 32x50G NRZ uVCSELs next quarter, with high-volume ramp targeted for early 2027. The 200G parts clock in at 37GHz bandwidth, a meaningful threshold for 200G-per-lane operation. Importantly, PicoJool is building on one of the largest existing optical manufacturing ecosystems. The company’s thesis is that GaAs VCSEL capacity scales far more easily than the InP laser supply chains that many investors have focused on. PicoJool is building on GaAs with WIN Semiconductor, which has shipped over a billion VCSELs for 3D sensing applications. This decision to stay on GaAs instead of chasing indium phosphide or silicon photonics is important. GaAs VCSEL manufacturing capacity is vastly larger than merchant InP laser capacity. I'm also interested in the "slow and wide" option. The 32x50G uVCSEL isn't trying to push a single lane to extreme speeds. It runs 32 lanes at 50G NRZ, a massively parallel approach that fits naturally into co-packaged optics and near-packaged optics architectures. The cost target is to compete with copper which makes sense for AI scale-up interconnect, where you're wiring hundreds of GPUs inside a rack. The founder has done this before, too. Al Yuen shipped the first GbE VCSELs in 1996. At Alvesta, his first startup, he invented the active optical cable in 2001, the now-standard format where the E/O conversion lives inside the cable assembly rather than a separate pluggable module. Alvesta was acquired by Emcore in 2002. He went on to develop vertical oxidation techniques for high-volume VCSEL fabrication in 2016. Roadmap goes to 3.2T. Worth watching!

MS on storage. I actually think its much worse than many realize in terms of what agentic storage demand will cause. Detailed in our full report on agentic AI impact on storage. thediligencestack.com/p/the-…

하이닉스 HR 분과 지속적으로 교류하며 미팅을 합니다. 로드맵을 들으니 전망이 밝아보였습니다. 누라두님께서 언급한 포인트도, HR을 통해 들었기에 중요한 트윗이라 생각하여 공유드립니다 :)

Stacy has created a new process for creating cuboidal silicon ingots called the Rasgon process It's a huge upgrade over the traditional Czochralski (CZ) process

Warren Buffett: "If I do something brilliant — don't count on it — with $5 billion, it's 1% of [Berkshire Hathaway's] net worth. And 1% doesn't do much. So I have to think about big things — and they are limited." "I made money in an entirely different way when I was working with, originally, $9,800 ... Everything I did could double our net worth. There were thousands of possibilities — and I wanted to know every one of them." (Charlie Rose || 2022)

Rumor: TSMC will partner with Innolux on FOPLP (Fan-out Panel-Level-Packaging) advanced packaging for AI chips, media report. TSMC will work with Innolux at its Longtan plant, aiming to use FOPLP on future AI and HPC chips. Innolux is already said to be working with SpaceX on FOPLP, and its annual report says monthly FOPLP output has already 10x’ed to 40 million units, at high yield. Capacity is full. $TSM #Innolux $3481TW $2303TW #semiconductors #semiconductor #FOPLP money.udn.com/money/story/56…

There is no stock market bubble. Large Cap Fwd P/E's (2027/28) & Analysis: $NVDA: 22.9x -> 16.1x $GOOGL: 24.8x -> 20.6x $MSFT: 20.2x -> 16.8x $AMZN: 24.2x -> 17.9x $META: 15.7x -> 13.8x $AVGO: 19.7x -> 16.2x Analysis: $NVDA: 16.1x (2028) on >40% growth is the market pricing a sharp deceleration. But Vera Rubin in full priduction ramping in H2. Feel like the market is pricing in ASIC share erosion heading into 2028, but then you've got CPO coming soon. $GOOGL: The most priced to perfection on this list on cloud backlog Search re-acceleration. But you've got things like TPUs, OCS Jupiter fabric as a structural cost edge. Gemini too. $MSFT: Stupid cheap. Maia live in DCs for Copilot/foundry inference. Targeting ~50% of AI compute on own silicon by 2028. Rights to OAI's chip designs. $AMZN: Tough company to model margin trajectory on tbh. AWS acceleration retail/advertising operating leverage make the earnings base the most optionality-driven & hardest to pin. And Trainium shipping, anchored by Anthropic. Plus at what point do robotics/automation start to ramp? $META: Very cheap due to a "trust discount" on their AI capex, not an earnings problem. If the capex visibly converts to revenue/margin, you get both EPS upside multiple re-rating. $AVGO: Steep de-rating after their earnings report last week mainly because of ASIC ramp. On CPO, there's arguments to be made that Broadcom wins over Nvidia longer-term. Personally, I'll be buying Nvidia, Microsoft, Meta and Broadcom a little more heavily through Q3-Q4. I highly doubt we'll see such compressed valuations in quality tech names for a very long time. Caveat: 2028 PE's are via my own EPS modelling.

I share concerns about China’s access to advanced AI models, but if the admin feels so strongly about this, I have a series of questions it should answer: - Why did it loosen export controls to allow AI chip sales to China, which allow China to build its own Mythos? - Why is it not enforcing existing export controls that would prevent China from smuggling AI chips from Southeast Asia and other countries? - Why is it not enforcing existing export controls that would prohibit Chinese companies from training advanced AI models on remotely accessed AI chips? Or imposing tighter controls on remote access? - Why has it still not closed a loophole it created that allows Chinese front companies outside China from making AI chips at TSMC or Samsung? - Why has it not tightened controls on China’s access to semiconductor manufacturing equipment (which have not been updated in over 18 months - the longest the US has ever gone without updating them)? - Why has it not imposed equivalent controls on all advanced AI models being served to China/Chinese companies? - Why did it restrict access to all countries and foreign nationals accessing Mythos/Fable, not just China? If the admin was serious about addressing the challenges posed by China in AI, it would be using export controls to address all of these questions and build a comprehensive strategy to prevent China from building or obtaining advanced models. But over the last 1.5 years, it has loosened or ignored controls on China, and only opened new loopholes in controls it inherited. If the admin truly has deep concerns about China’s access to advanced models, it has to act accordingly. It isn’t.

Open source isn't just good for developers, it's one of America's strongest tools for AI security. More models means more defenders and more front doors protected. Earlier this week at the @MilkenInstitute Global Conference, Jensen sat down with @beckyquick to explain why 👇

Anthropic Deputy CISO Jason Clinton earlier this week: "I suspect that we have ~7 to 10 months before open weights models have this capability. And when that happens, we will have the roughest, most difficult time in our careers that we've had as cybersecurity folks. That will happen for approximately I would guess 18 months."

BRUH I'm talking about trust Why are u inserting AI demand in it?? AI demand will obviously grow (including demand/token processing for closed-source models), but a backup option where an enterprise controls the model layer is a must now