Prices of silicon wafers used in semiconductor production are set to rise in 2027 as more firms ink long term agreements (LTAs), media report, noting capacity expansion for both memory and logic chips is fueling demand. 12-inch Epitaxial wafers: estimated up 15%-25% 12-inch Polished Wafers: seen up over 40% for most, 50% for 2nd tier clients GlobalWafers, Taiwan’s biggest maker, is seen at $122 average selling prices (ASPs) in 2027, up from $92 for 12-inch silicon wafers this year, and up to $166 in 2028. Shipments are seen rising to 16.174 million wafers per quarter next year, from 11.907 million per quarter this year.

AMD has followed up its Advancing AI data release with a new blog and a white paper on EPYC Venice that attempts to put real detail behind the claims first made on stage in July. The 2.24x platform and 1.2x per-core advantages over NVIDIA Vera in SPECrate 2026 Integer now come with estimated scores, compiler versions, memory configurations, and all 14 subtests broken out. The generational data is impressive, with the 256-core EPYC 9996 landing between roughly 1.5x and 1.8x over the 192-core Turin flagship across Redis, NGINX, MySQL, MongoDB, and server-side Java, all measured by AMD on Venice silicon. @AMD also rebuilt its 100 kW rack model around SPEC CPU 2026, which moved Venice to 3.4x over Vera and nudged its own Turin figure down slightly. The core story is the same one AMD told at Advancing AI, now with some supporting detail. The agents-per-watt and performance-per-watt claims from the keynote have no backing data in this paper, interestingly. All of it is AMD internal testing or estimates, so independent results on shipping systems from both vendors are what will settle the debate. Blog: White paper:

Microsoft’s self-developed AI chips will be in mass production in 2027 as all work on design and capacity allocation remains smooth, media report, citing the head of Microsoft Taiwan, Sean Pien. Prior leaks had put production in the 2nd half of 2026. Pien said in the near term, the chips will power Microsoft’s internal needs, and won’t be open for third party use. He said the chips will drive down overall computing costs. $MSFT #semiconductors#

Microsoft has doubled its Taiwan AI data center plan to four from 2 due to strong demand, media report, adding they will be located in the northern Taiwan towns of Gueishan and Luzhu. The top data center clients for Microsoft in Taiwan are semiconductor firms and other high tech companies, the report says, citing the head of Microsoft Taiwan. $MSFT

China’s 3 top display screen panel makers, BOE, TCL CSOT and HKC, issued price increases to clients in a coordinated move, and said they will cut production during China’s Golden Week holiday (Oct. 1-7), media report, calling it a clear attempt to drive up prices. The 3 companies hold over 60% of the global LCD panel market. The price hikes will take effect as early as this month.

Micron faces a strike at one Taiwan location, Taoyuan, after mediation talks with the union there collapsed, media report. The union said it will move forward with a strike vote since Micron failed to address the issue employees care most about, a profit-sharing system, which the union remains open to discuss. Micron's talks with another union in Taichung will resume Oct. 22 (10/22). $MU #semiconductors#

TSMC will begin 1.4nm (A14) trial production in the 1st quarter next year, media report, and equipment is already being moved-in to Hsinchu’s Fab 20 (P3). TSMC’s Fab 25 in Taichung, central Taiwan, will be the second 1.4nm fab, with phase 1 construction to finish in April, and trial production to begin as early as the 3rd quarter-2027. Taichung will be the main A14 cluster. The A14 timeline is ahead of schedule, something TSMC had noted during its 2nd quarter earnings call in July. $TSM #semiconductors#

Jason Chen, chairman of Taiwanese PC and IT hardware maker Acer, dismissed the memory Big 3's claim that memory prices will keep rising through 202X, saying it simply isn't true. Specifically, he said the reason memory makers keep coming out and saying "the price uptrend will continue into 2027" is that antitrust rules prevent them from coordinating prices directly, so they use public statements to send a kind of "signal" to one another. He said a veteran like him can tell just by looking. He also pushed back on the market narrative of a memory supply shortage: "No. How could it stay short forever? Chinese capacity keeps coming onto the market. The supply shortage problem has already completely disappeared." Chen also expects the cost of the SSDs and memory Acer holds in 2027 to be lower than this year. If that gets reflected in end product prices, he sees a possibility that prices at least stop rising and stabilize in the second half of next year. Chen said prices could be raised further in 1Q27 as well. But since this is a period of pricing chaos, he expects the size of the increases to gradually shrink. He estimates component prices could peak around the middle of 2027, when the capacity that chipmakers have added is expected to start running in earnest. After that, prices would enter a stable phase, and whether they can actually come down in the second half remains to be seen. He admitted, though, that nobody can know the exact timing of the reversal. Chen said PC selling prices coming back down is strictly "a hope." He criticized component suppliers for treating price hikes like a fad. Countless suppliers are showing up to notify him of price increases, he said, and a whole series of components, from PCBs to glass fiber cloth to memory, are going up one after another. He wrapped up by saying this chain of price hikes ultimately creates inflation, which in turn pushes central banks to raise rates, and that "this is not a normal phenomenon." *This quote does not reflect my views and is shared solely for informational purposes.

TSMC could see a monthly record high NT$520 billion in September revenue, a catalyst to propel the Taiwan Stock Exchange in early October, media report, citing institutional analysts. TSMC reports September revenue on Oct. 8 (10/8). The current all-time-high is August’s NT$514.81B. $TSM #semiconductors#

It was my honor to keynote this forum in Taiwan. "If Taiwan can reach a durable shared view of its own future . . . Beijing will reach the conclusion that many nations have reached many times in history: This is a war that may be winnable, but it is not worth winning."

While oil & equity mkts were flattish last wk, bond ylds hit multi-decade highs in different countries as central banks raised rates. While AI frontier model pacing was initially a concern, that faded by Friday. I expect the upcoming wk to be challenging. I have some simple portfolio rules including: 1) Don’t Fight the Fed: hiking cycle seems to have begun 2) Don’t Fight the Bond Mkt: Multi-decade highs in ylds 3) Don’t Fight Seasonality: Sept is worst month for hit rate & returns with mid-term yrs even worse These rules influence how I think about leverage, the ratio of long positions versus short positions, and individual position sizing. Investing is hard enough without fighting headwinds. As Warren Buffett (a remarkable 61 year run came to an end last week) has said, the market has to keep pitching but you do not need to swing. For AI, the battle lines seem to be drawn. OpenAI and Anthropic are pushing for regulatory capture while $NVDA & $MSFT are pushing for better testing before models are released. I’m in the latter camp. No product from any company should be released before it is safe. AI related companies focused on the potential bullish implications in case there was model innovation pacing including: 1) Focus on security 2) Resources deployed for testing 3) Potential for any training slowdown offset by limited hardware capacity being deployed to the infrastructure layer The SOX index, a representation of the infrastructure layer, rallied 0.8% last week with $IGV (the software ETF) also up 2.8%. Software was led by the security names, the $HACK ETF was up 8.0%, despite the 0.1% decline in the S&P. Longer-term, my view is that companies do not need the most advanced models for 90% of their workflow and usage will increasingly go towards open weight models. I believe that ultimately the LLM layer is likely to become commoditized. As an example, the ASP per token peaked in late May and has gone down ~50% since then while the number of tokens produced has gone up by ~4x. This is Jevons paradox in action. This should also be supportive of the infrastructure layer. As the model layer becomes increasingly commoditized, I believe the winners will be those that have the following attributes: 1) strong open weight models 2) distribution capability 3) training data 4) base business that is highly cash flow generative I believe the following companies have the attributes above to varying degrees: 1) $META (Muse from Meta is the #1# free app on the Apple app store with ChatGPT #2#), 2) $GOOGL (I expect a new frontier level model soon that should help the stock) 3) $MSFT (Co-pilot might be the "safe" way ~450M M365 enterprise users choose to access AI. But their 27% ownership of OpenAI does concern me and they do not have the wealth of training data that Meta and Google have.) Earlier this year, the advent of Agentic AI increased token production by 10-100x. This should keep demand strong for the semiconductor portion of the infrastructure layer. Given the escalating geopolitical tensions over the weekend and its potential impact on oil/bond yields, this could prove to be a challenging week: 1) Ukraine fired over 1,000 drones at Russia, including "largest ever" attack on Moscow 2) There was a missile attack on Riyadh for the first time since July On a positive geopolitical note, hopefully something constructive comes from the meeting between Presidents Trump, Xi and the attending business leaders on Wednesday. All the best in the week ahead.

Intel is in talks with Taiwan’s AUO for advanced packaging aimed at CPO (co-packaged optics), due to AUO’s work on Micro LED technology, media report, a project that also uses glass substrates. The two may also work together on advanced packaging for high-density compute. $INTC #AUO# #semiconductors#

Nexperia, the Dutch automotive chipmaker, has tapped Tata Electronics to manufacture and package its power control chips (MOSFETs) in India, Reuters reports, the first major publicly announced foundry + packaging client for Tata, which has also been in talks with Analog Devices, NXP, others. Taiwan’s PSMC teamed up with Tata on its 300mm fab, including design, construction, training engineers, and licensing production technology. #India# #Tata# #semiconductors#

Winbond agreed to buy Infineon’s NOR Flash and F-RAM business in an all-cash deal valued at US$1.12 billion (NT$35.6B), Infineon said, adding the business serves customers across automotive, industrial and infrastructure markets. The deal is targeted to close in the 2nd half of 2027. Winbond plans to bring back the “Spansion” brand name for NOR Flash. The deal is set to make Winbond the biggest NOR Flash maker in the world, with around 30% market share. $IFFNY 2344