Hello world, it’s Friday, October 16th.
ASML is reportedly developing a hybrid bonding tool, according to SemiAnalysis, a direct move onto Besi’s turf in advanced packaging equipment. Elsewhere, credit-default swaps on CoreWeave now imply a better-than-60% chance of default, TSMC has finished construction on its second Arizona fab as brokerages project its 2028 CapEx as high as $105 billion, and Schneider Electric raised its 2026 guidance for the second time on surging data-center orders.
Let’s get into it. — Austin & Vik
Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!
Presented by Crusoe
Crusoe builds the full stack for AI, from power and data centers to the cloud. It can also help you tune a small model on your own data for just a few bucks.
For repeatable work like reports, an open model trained on your terminology and formats can match a frontier model, and it’s cheaper and faster to run.
Crusoe’s Serverless Fine-Tuning handles the training: upload your dataset, pick an open model like Llama, DeepSeek, or GLM, and pay per million tokens used, not per GPU hour.
About $2 buys roughly 5M training tokens on smaller models.
Deploy on Crusoe Cloud in one click or download the weights and deploy on your own.
ASML Enters Hybrid Bonding, Pressing Directly on Besi’s Territory
According to SemiAnalysis, ASML is developing a hybrid bonding tool, a direct move into advanced packaging equipment that sits at the core of Besi’s business. Hybrid bonding, which enables die-stacking at pitches too fine for conventional interconnects, has been one of the few high-growth equipment markets where Besi operated without a credible threat from the lithography giants. That changes if ASML qualifies customers on its own tooling. Besi built its position over years of process development with chipmakers racing toward 3D integration; ASML arrives with an installed-base relationship at every major fab on the planet, which is not nothing. The competitive pressure lands precisely as AI-driven demand for stacked memory and logic accelerates volume commitments across the industry.
Vik: This is a logic play first, not the HBM wave the framing suggests. SemiAnalysis says it starts as a die-to-wafer tool, and D2W only runs in volume in logic today - exactly where Besi’s hybrid bonding business sits, at TSMC and Intel. HBM-grade bonding is still too expensive and too slow at high stacks.
Austin: Besi isn’t on its own though. Applied Materials owns 9% of Besi, and the two launched Kinex together last October, an integrated die-to-wafer hybrid bonder. So ASML is lining up against Applied here too.
CoreWeave Default Odds Breach 60 Percent as Bonds Keep Falling
Credit-default swaps on CoreWeave now imply a better-than-60% chance of default, an all-time high reached Thursday as its high-yield bonds extended a slide triggered by the company’s plans to raise capital from insurance pools. The selloff has stripped more than $10 billion in market value across just four sessions, per Deutsche Bank, which called the drop a buying opportunity. Jensen Huang stepped in publicly, telling investors that CoreWeave is improving and requires no Nvidia backstop, a defense that steadied the stock without doing much for the bonds. The core anxiety is structural: CoreWeave, a loss-making neocloud, finances GPU infrastructure almost entirely against hyperscaler contracts, and credit markets aren’t convinced that’s enough collateral.
Vik: If CoreWeaave has to play lenders 15% interest, then there is not a lot of trust left in the public eye, especially when CoreWeave is trying to dip into insurance pools. smh.
Austin: Technically Nvidia already signed a backstop in Sept 2025, a $6.3B order to buy any CoreWeave capacity its customers don’t use, through April 2032. So you could interpret this as not needing to carry out that backstop, or no NEW backstop.
TSMC Completes Arizona Fab 2, Pursues Texas Site and TeraFab Talks
TSMC confirmed during its third-quarter earnings call that construction on its second Arizona fab is complete, with the site now moving toward equipment installation and production ramp. The company is also evaluating Texas as a second U.S. location, with discussions active around Elon Musk’s TeraFab complex in the state. TSMC has made one condition clear in those talks: it won’t operate any fab it doesn’t control. Brokerages tracking the earnings call now project TSMC’s annual CapEx reaching $105 billion by 2028, up from current levels. Quarterly profit rose 68.7 percent, and CEO C.C. Wei said AI demand continues to exceed supply. Texas would mark a third U.S. state for TSMC’s expanding footprint.
Vik: TSMC is very clear that they will no leave any lunch on the table, so their involvement in TeraFab will absolutely involve their business.
Austin: $105B is the UBS forecast, the most bullish of the brokerages. Still, that’s about 1.7x this year’s ~$62B guide in two years. And some of it is just pricier tools, with ASML confirming its first EUV price increases this week.
Schneider Electric Raises 2026 Guidance Twice as Data-Center Orders Surge
Schneider Electric lifted its 2026 financial targets for the second time this cycle after data-center demand drove sales well above expectations, the French industrial group announced this week. Order visibility has now extended into 2027 and 2028, according to Bloomberg, easing concern that AI infrastructure spending might be cooling. The raises come as power and cooling infrastructure tightens into the binding constraint on new capacity, a position that benefits Schneider directly. Shares slipped 1.53 percent on the day, a muted reaction that tends to greet good news already priced into a stock trading near cycle highs. The company did not name specific hyperscaler counterparties in its disclosure.
Vik: Power and cooling are the binding constraint on new capacity now, and Schneider’s order book proves it - triple digit data center demand in a single quarter, another record backlog.
SK Hynix Brings AI Inference Inside the Flash Cell at OCP Summit
SK hynix presented its 3D FeNAND analog computing-in-memory architecture at OCP Global Summit 2026, held October 12–15 in San Jose, California, pitching a design that runs AI inference calculations directly within NAND flash rather than shipping data out to a separate processor. The approach uses ferroelectric materials to perform multiply-accumulate operations at the storage layer, cutting the round-trip to DRAM or HBM that ordinarily consumes bandwidth and power. The company framed memory-centric server systems as the organizing principle for AI infrastructure, positioning FeNAND as a concrete step toward that architecture rather than a roadmap placeholder. SK hynix disclosed its A-CiM research in a newsroom post timed to coincide with the summit, detailing the ferroelectric 3D NAND stack behind the demonstration.Sector Watch
Compute & Edge
Volvo Cars expands its automated-driving partnership with Nvidia, deepening reliance on the Drive platform for next-generation safe-driving systems. (Nvidia News)
Arm publishes a new rack-level CPU benchmark for agentic AI, claiming twice AMD’s performance per rack in inference workloads. (The Elec)
Li Auto announces equity financing for its in-house chip business, accelerating its move toward vertically integrated automotive silicon. (GlobeNewswire)
Foundry & Packaging
Japan offers GlobalFoundries up to $21 billion to expand its U.S. semiconductor complex in an unprecedented public-private incentive package. (AP via ABC News)
LG Innotek targets 2027 chip-substrate revenue growth as ongoing supply talks with Big Tech customers near agreement, DigiTimes reports. (digitimes)
Unimicron halts construction at its Yangmei site after a fatal accident, adding execution risk ahead of its October 27 results. (Unimicron)
Memory
CXMT targets DDR5 RDIMM production by end-2026 while SiCarrier simultaneously showcases China’s advanced chipmaking equipment at a joint event. (digitimes)
Samsung Electronics highlights limits of AI memory standardization at a JEDEC forum, citing customer-specific requirements as an obstacle to common specs. (The Elec)
Kioxia sheds two-thirds of its peak market value as Chinese NAND rivals close the gap, with YMTC now ranked third globally in NAND shipments. (Nikkei Asia)
PCB & Components
Samsung Electro-Mechanics is on track to roughly double Q3 operating profit, driven by surging MLCC demand from AI server customers. (머니투데이)
Lite-On closes a $1.11 billion sustainability-linked loan oversubscribed 140 percent, funding expansion of its power and cooling product lines. (Dealroom)
Power
Infineon and Wolfspeed extend and expand their long-term SiC wafer supply agreement to include 200 mm wafers, securing Infineon’s EV and industrial power roadmap with next-generation substrate supply. (Business Wire)
Venture Global signs power-generation agreements with GE Vernova and Baker Hughes to build gas-fired plants supporting AI data-center loads. (Reuters)
Delta Electronics unveils a standalone microgrid system and warns that AI power gaps will drive adoption of distributed microgrid infrastructure. (Delta Electronics)
Optics & Networking
Nvidia unveils next-generation co-packaged optics technology placing optical engines directly on the interposer, advancing CPO integration toward production-grade AI switch designs. (The Elec)
KT rolls out a dedicated AI-only fiber line rated at 800 Gbps capacity to support the large data transfers required by AI data centers. (Light Reading)
Ericsson and Verizon complete the first outdoor 6G test call over the 7 GHz band, advancing early field trials for next-generation networks. (Light Reading)
Policy & Trade
Taiwan agrees to curb AI chip exports to China as part of a broader US trade deal framework, Bloomberg reports. (Bloomberg.com)
China proposes adding advanced AI models, chips, and agents to its export-control list, potentially restricting overseas technology transfers. (Bloomberg Tech)
Qualcomm signs a long-term patent cross-license agreement with Transsion, extending its IP licensing reach into fast-growing African and emerging markets. (Qualcomm)
Data Centers
Nscale eyes a $35 billion IPO valuation but its flagship Arctic data center is running six months late and has been hit by fires and labor disputes, Bloomberg reports. (Bloomberg)
OpenAI names Uday Ruddarraju, a former xAI executive, as its new head of infrastructure and compute, replacing departing CTO Sachin Katti. (The American Bazaar)
Tell us how to make Semi Doped better! Give us your feedback! ❤️