Google TPU shipment estimates for 2027 now range from 6.7 million to 8.8 million chips. At the high end, one Taiwan sell-side model translates that into more than 100,000 racks. We hear the same uncertainty in our own supply-chain interviews. Nearly everyone expects volumes to surge. Almost no one can pin down a precise number. The two major rack assemblers today are Inventec (2356 TT) and Celestica $CLS. Supply-chain checks suggest Foxconn (2317 TT) and Compal (2324 TT) could join in 2027 because existing capacity may not be enough.

Hanmi Semiconductor is showing a big AI packaging lineup at SEMICON Taiwan. The company is already No.1 in HBM TC bonders and now pushing deeper into 2.5D equipment for AI chips, plus HBM4 tools and a next-gen Wide TC bonder for larger DRAM dies. Hanmi Semiconductor is showcasing five models of 2.5D packaging equipment: FC Bonder 3.5 FC Bonder 75 2.5D TC Bonder 40 C2S 2.5D TC Bonder 40 C2W 2.5D TC Bonder 120 Among these, the FC Bonder 3.5, FC Bonder 75, and 2.5D TC Bonder 40 C2S and C2W have been released sequentially since the end of last year and are being supplied to global foundry and OSAT companies for mass production, while the 2.5D TC Bonder 120 is scheduled for release soon. Hanmi Semiconductor official stated, "Semicon Taiwan is a global event where AI semiconductor technologies converge," adding, "Hanmi Semiconductor will lead the growth of the AI ​​semiconductor market by expanding the supply of advanced packaging equipment not only for its globally leading HBM TC Bonder but also for the AI ​​system semiconductor market." #HanmiSemiconductor# #HBM# #AI# #Semiconductors#

MLCC spot prices are up 3-5x on scarce AI-server specs since April. Some individual models up 8-10x. Murata, Samsung Electro-Mechanics, and Taiyo Yuden all raising prices. Nearly every major MLCC manufacturer trades in Japan, South Korea, Taiwan, or China. @roundhill launched $CCML today giving US investors direct access to those names.

$TSM $MU $SKHY $NVDA TSMC Senior VP and Co-COO Hou Yongqing (Cliff Hou) at SEMICON Taiwan 2026 On the speed of AI demand Compared with the past 30 years, he has “never seen demand grow at such a rapid pace and at such high frequency.” If equipment demand at the end of last year was 1x, it rose to 1.5x after one quarter and reached 1.9x by July—a 90% increase in just over half a year, nearly doubling. On the ecosystem challenge “How the entire Taiwan semiconductor ecosystem can cope with such rapid demand growth in just six months is a major challenge we currently face.” On technology and collaboration AI’s need for computing power and energy efficiency is driving packaging and system-level performance “much faster than before.” The supply-chain model must change: “We must ensure that the performance achieved at one stage can still be realized at the next partner’s stage.” This requires “a brand-new model of cooperation.” On fab expansion TSMC is simultaneously building 13 fabs in Taiwan—“something we have never done before”—plus another 5–6 overseas, for a total of nearly 20. Previously the company typically built only 4–5 at a time. “Even so, it is still not enough to meet demand.”

AI rack battery backup units are moving from 3.5kW to as much as 25kW in two years. Taiwan battery module suppliers say 3.5kW was the mainstream BBU output in 2025. That moved to 5.5kW in 2026. Dynapack (3211 TT) has started sampling units above 15kW and targets mass production in 2027. Celxpert (4931 TT) plans 15kW and 25kW models in 2027. Dynapack also plans to triple its BBU capacity this year. $NVDA’s MGX-compatible 800 VDC power rack arrives in 2H26. A Delta Electronics (2308 TT) design shown at GTC pairs a 110kW power rack with five BBUs. The suppliers say higher-output modules require more battery cells, extending backup time and raising the selling price per BBU.

This is serious competition for NVDA. The startup says it took just 44 days after getting its test chips back from Taiwan Semiconductor Manufacturing to have them up and running inference workloads—the computing processes that allow AI models to respond to user queries—a process that usually takes six months or more. When Etched needed help designing servers and racks to integrate its chips into one big system, the company in 2024 recruited Brian Loiler, a top systems engineer who spent nearly 23 years at Nvidia. Loiler has since recruited to Etched about a dozen engineers from Nvidia who, in some cases, turned down attractive counteroffers, he said.

Over the past three decades, the global semiconductor industry was built on a highly efficient model of specialization. The United States controlled chip design, EDA, IP, and end-platform ecosystems. Taiwan became the center of advanced foundry manufacturing and packaging. South Korea dominated DRAM, NAND, and memory supply. Japan retained key positions in materials and equipment. China provided massive electronics manufacturing capacity and end-market demand.

Steam Deck OLED prices have increased again across several Asian markets, with South Korea seeing the biggest jump. The 1TB model now costs ₩1,578,000, up 51% from its previous price. Updated prices: - Japan – 512GB OLED: ¥137,980 | 1TB OLED: ¥167,980 - South Korea – 512GB OLED: ₩1,298,000 | 1TB OLED: ₩1,578,000 - Taiwan – 512GB OLED: NT$26,280 | 1TB OLED: NT$31,800 - Hong Kong – 512GB OLED: HK$6,488 | 1TB OLED: HK$7,788 The latest increases range from 38% to 51%, depending on the model and region. This follows an earlier round of price hikes in March 2026 linked to exchange rates.

🚨 POD UP! HAPPY FRIDAY! Bestie Guestie Marc Benioff (@Benioff) fills in for Sacks A LOT covered on this one: -- Trump-Xi Summit: Trade, Taiwan, midterms impact -- AI's impact on software: What thrives and what dies? -- OpenAI could sue Apple over failed ChatGPT integration -- Thinking Machines drops new model, future of AI is multi-sensory -- BIG Science Corner on a potentially devastating El Nino (0:00) Salesforce CEO Marc Benioff joins the show! (1:14) Trump-Xi summit, doing business in China as a US company, impact on Americans and the midterms (18:46) Taiwan, chips, AI models, and peace through trade (31:41) AI's impact on software: What SaaS thrives, what SaaS dies? (47:26) OpenAI is considering suing Apple over failed ChatGPT integration (56:54) Thinking Machines releases real-time model, future of consumer AI, multi-sensory models (1:02:24) Science Corner: Impacts of a historically strong El Nino in 2026 (1:11:40) Anthropic goes after "Dark SPVs"

- Intel ($INTC) CEO Lip-Bu Tan said AI-driven memory shortages could become even more severe in 2027, with some memory prices already rising 5x to 7x. - He said memory can account for as much as 70% to 80% of the BOM in some budget smartphones and laptops, meaning sustained price increases could pressure low-margin consumer device production first. - The article attributes part of the broader shortage to HBM expansion competing with commodity DRAM for manufacturing capacity, tightening supply beyond HBM itself. - $INTC views advanced packaging that integrates CPUs, AI accelerators, HBM, and high-speed I/O into increasingly large packages as a key strategy, highlighting EMIB as an alternative approach to CoWoS. - Tan also identified package substrates as a bottleneck, saying the AI substrate supply chain is heavily concentrated among two major suppliers in Japan and two in Taiwan. - He argued that reinforcement learning and AI agents are reviving CPU demand because CPUs increasingly handle orchestration across models and applications, while Intel is currently unable to satisfy all requested CPU volume. - Intel is pursuing a "co-opetition" strategy with $NVDA, combining Intel CPUs with Nvidia GPUs and NVLink while continuing to compete in other areas. - Tan also highlighted power and cooling as additional AI infrastructure constraints, with Intel investing in approaches including liquid cooling and microfluidic cooling. > The bigger message is that AI bottlenecks are spreading from memory → advanced packaging → substrates → CPUs → power → cooling. Looking at AI infrastructure only through GPUs increasingly misses large parts of the hardware stack. > On memory, the important point is the interaction between HBM and commodity DRAM capacity. If AI data-center demand remains elevated, the resulting capacity allocation could keep not only HBM but broader DRAM supply and pricing tight. > Packaging may be the most interesting part. As accelerators become larger and integrate more HBM, EMIB/CoWoS-class integration, substrates, thermal management, and testing all become more critical. AI scaling is increasingly shifting from simply designing a better GPU to being able to manufacture and support an increasingly complex system.