Hello world, it’s Wednesday, October 7th.
Elon Musk ruled out any operating role for TSMC at Terafab and Intel stays in the venture, with Musk’s companies building and running the fab on licensed TSMC process technology. Elsewhere, SpaceX is in talks to borrow $40 billion through Apollo to buy Nvidia chips, Applied Materials and Intel formalized an R&D pact on next-generation transistors and packaging, and Samsung Electro-Mechanics landed a $216 million AI server MLCC order.
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 AI infrastructure from the ground up — power, data centers, and the cloud on top.
If you’re running open models — GLM, DeepSeek, Kimi, Qwen — Crusoe’s managed inference is fast: up to 9.9x faster time-to-first-token and 5x the throughput of vLLM, thanks to their MemoryAlloy caching layer and tokenizer fastokens. And it’s the easy kind of fast: pick a model, grab an API key and start inference. No GPUs to provision, no infrastructure or clusters to manage, and it’s pay per million tokens.
$5 in free credits to try it. Send your first request today →
Musk Confirms Intel Stays in Terafab as TSMC Operating Role Ends
Elon Musk has ruled out TSMC taking any operational role at Terafab, the Texas AI chip complex he is building, saying his own companies will construct and run the facility while TSMC contributes process technology under license. The clarification came after Intel shares initially slid on fears Musk might hand TSMC a broader position, then rebounded once the structure became clearer. Bloomberg confirmed that Intel will keep working on the Terafab venture, preserving its place in a project that had briefly looked like it might sideline the company entirely. Terafab thus takes shape as a Musk-owned fab running TSMC process nodes with Intel in a supporting capacity, according to Bloomberg.
Austin: Note to readers, this whole convo started recently with a scoop from Tim Culpan:
Vik: TSMC licensing its process for someone else to run? That seems like news in itself.
SpaceX Pursues $40B Apollo Financing to Stockpile Nvidia Chips
SpaceX is in talks to borrow $40 billion, led by Apollo, specifically to purchase Nvidia chips for its AI buildout. The scale is hard to ignore: forty billion dollars in debt, raised not for rockets or satellites but for GPUs. Doosan Enerbility has separately completed production of a 380-megawatt gas turbine destined directly for SpaceXAI, holding a shipment ceremony at its Changwon headquarters with roughly 70 attendees. SpaceX is sourcing chips through private credit and commissioning bespoke power infrastructure without routing either through a hyperscaler. The financing round, if closed, would rank among the largest single AI hardware purchases on record.
Vik: They’re building the whole stack themselves. Chips through private credit, and a 380MW Doosan turbine shipping straight to SpaceXAI, no hyperscaler in between. Compute’s still short, so owning your own supply makes sense right now.
Austin: At Jensen’s $40B per GW for Vera Rubin, this loan buys roughly 1 GW of compute. One 380MW turbine is a start, but they’ll need a lot more power than that.
Penguin Solutions lifts fiscal 2027 revenue target to $2.43 billion
Penguin Solutions raised its fiscal 2027 revenue outlook to $2.43 billion, up from a prior target of $2.2 billion, after reporting a Q4 earnings and revenue beat driven by demand for AI systems integration and memory subsystems. The company now guides for fiscal Q1 2026 revenue between $480 million and $520 million, with full-year adjusted operating margins expected to expand alongside the revenue ramp. The stock climbed on the news. Analyst reactions split: Stifel raised its price target on the strength of the revised guidance, while Barclays cut its rating, citing valuation concerns as the revenue trajectory steepens. The divergence puts investors in the familiar spot of deciding whether the growth is worth the multiple.
Vik: Memory is doing the heavy lifting here. Integrated Memory was over nine-tenths of the year over year gain, so this is a memory subsystem story dressed up as AI systems.
Samsung Electro-Mechanics Cements AI Server MLCC Lead With Fresh $216M Order
Samsung Electro-Mechanics has landed a $216 million MLCC contract for AI server supply, adding to a position that already covers more than 40% of the AI server MLCC market, according to Seoul Economic Daily. The company now holds over 40% share in AI server MLCCs, a component that has become a chokepoint in next-generation server builds as GPU-dense configurations demand far higher capacitor counts per board than conventional rack servers. Digitimes reports that the new deal locks in 2027 delivery commitments as supply pressure in the segment continues to build, with lead times tightening across the MLCC tier. The $216 million figure represents a discrete order on top of existing volume, not a cumulative tally, reinforcing how concentrated procurement has become around a single supplier.
Vik: 600K MLCCs in a single rack. That’s where the demand is coming from.
Applied Materials and Intel formalize next-generation transistor and packaging R&D pact
Applied Materials and Intel have announced a formal R&D collaboration targeting next-generation transistors, interconnects, and advanced packaging technologies, drawing on Applied’s EPIC Center in Silicon Valley and Intel’s R&D campus in Hillsboro, Oregon. The goal is to shorten time-to-market for semiconductor processes central to AI computing. The two companies will combine capabilities across both facilities to accelerate chipmaking innovations that span transistor scaling and advanced packaging in a single coordinated program. Applied and Intel describe the relationship as built on decades of prior work together, now structured to move faster at the leading edge. Besi and Applied have separately deepened their own packaging collaboration around the same period, extending Applied’s co-development reach across the packaging supply chain.
Vik: The Applied-Intel headline is Foveros and transistors, but the Besi extension is the part I’d watch - die-on-panel and photonics-enabled interconnect.
Austin: Intel joins Samsung, Micron, SK Hynix, TSMC, and Kioxia at EPIC. All three leading-edge foundries doing R&D at Applied’s site now. Good spot for Applied to be in.
AI-Powered Hack Hits Seven South Korean Banks, Exposes 68,000 Records
Hackers used a Chinese AI cybersecurity tool called Artex AI to breach at least seven South Korean financial firms and steal personal information belonging to 68,000 people, South Korean officials said. The attacks were initially detected last week, and the National Police Agency’s cyber terror unit opened a formal probe on Tuesday. Investigators said the intrusion is one of the first known AI-agent-assisted attacks on the global financial system.
Sources: wsj.com
Austin: AI-agent assisted you say? Crime even requires GPUs now…
Sector Watch
Memory
Micron union in Taiwan receives official authorization to strike, introducing potential production disruption risk at a critical DRAM manufacturing site. (Reuters)
Samsung Electronics Q3 profit seen jumping nine-fold year-on-year though analysts warn chip margins may remain flat amid HBM yield pressure. (Reuters)
Kingston retains 62% memory-module market share as B2B AI server demand drives record shipments, per DigiTimes analysis. (digitimes)
Foundry & Packaging
Towa announces a new Japan plant to reshore strategic chip-packaging equipment production as AI packaging demand accelerates. (digitimes)
NEXTIN wins a 25.44 billion won wafer inspection equipment order from SK Hynix, its largest single order disclosed this cycle. (The Elec)
Compute & Edge
AMD will report fiscal Q3 2026 financial results on an unspecified date, with no additional details provided in the announcement. (AMD)
Cirrus Logic acquires INVENTVM Semiconductor, expanding its mixed-signal and audio silicon portfolio with an undisclosed-value deal. (Business Wire)
Silicon Labs launches new IoT developer platform tools at its annual summit, adding edge-AI capabilities for connected hardware builders. (EE Times)
Optics & Networking
Openreach deploys Ciena’s Waveserver 1.6 Tbps platform across the UK national fibre network, the operator’s largest coherent upgrade to date. (Comms Business)
CIG and ml&s Group will establish high-speed optical-module production in Arteaga, Mexico, targeting AI data-center interconnect demand. (Business Wire)
AT&T forms a new 50%-owned fiber JV combining Gigapower and Forged Fiber 37 with Global Infrastructure Partners and CPP Investments to accelerate wholesale open-access fiber expansion. (Light Reading)
Power
Bloom Energy adds a Silicon Valley manufacturing facility as AI-era demand for solid-oxide fuel-cell power systems outpaces existing capacity. (CoStar)
Infineon and ZuriQ expand their partnership to develop scalable quantum-computing control electronics, deepening the chipmaker’s quantum hardware push. (Bisinfotech)
LS Power closes a $6 billion flagship infrastructure fund, citing AI-driven power-demand growth as the primary investment thesis. (Bloomberg.com)
Data Centers
LG Electronics signs a 5 GW cooling-solutions supply agreement for North American AI data centers, its largest cooling commitment to date. (The Elec)
Lambda targets a $4 billion fundraise ahead of a planned IPO as the Nvidia-backed GPU cloud expands capacity to meet hyperscaler overflow demand. (Reuters)
Huawei details global expansion plans for its Ascend 950 supernode cluster, outlining deployment roadmap for large-scale AI training infrastructure. (Huawei Central)
Policy & Trade
Amazon warns publicly that more than 100 proposed US data-center restriction measures risk ceding AI infrastructure leadership to rival nations. (Yahoo Tech)
Zhone acquires Tellabs in a telecom-equipment deal that consolidates two access-networking vendors amid carrier spending shifts. (Light Reading)
Microsoft and Nvidia CEOs are set to jointly unveil a new AI laptop at a San Francisco event, marking a high-profile co-launch for Copilot+ PC hardware. (Reuters)
Tell us how to make Semi Doped better! Give us your feedback! ❤️
The SpaceX/Apollo line is the one I'd watch. Capex funded from operating cash flow can be cut the moment demand slows; $40B of debt keeps the interest bill running regardless. If more of the GPU buildout shifts to borrowed money, a slowdown would hit the chip names harder than the current cycle math assumes.