TSMC Terafab Talks, AMD On-Device AI & China's DUV Stockpile

Failed to add items

Sorry, we are unable to add the item because your shopping basket is already at capacity.

Add to cart failed.

Please try again later

Add to wishlist failed.

Please try again later

Remove from wishlist failed.

Please try again later

Follow podcast failed

Unfollow podcast failed

TSMC Terafab Talks, AMD On-Device AI & China's DUV Stockpile

Listen for free

View show details

(00:00:00) TSMC Terafab Talks, AMD On-Device AI & China's DUV Stockpile
(00:01:08) TSMC Capex and US Expansion Scale
(00:02:01) China DUV Stockpile Quantified
(00:02:57) AMD Ryzen AI 320B Local Inference
(00:03:44) Tesla Memory Cuts and Colocation Growth
(00:04:37) Key Watchpoints This Cycle

TSMC has confirmed early discussions with Elon Musk's Terafab project for a Texas fab partnership — no longer a rumour, and the structure under consideration (ownership stake vs. technical role) will define how much capex risk TSMC absorbs and how seriously US domestic semiconductor policy is being executed.

Set against that, TSMC's broader US commitment is already concrete: $60–64B in 2026 capex guidance plus a separate $100B Arizona commitment, with Texas potentially adding six advanced fabs as a second US hub. The open question is whether the economics justify the pace, or whether geopolitical pressure is driving a capex cycle that will compress gross margins through the ramp.

Meanwhile, China has quietly built a formidable DUV lithography stockpile — 343 immersion machines, 270 of them ASML NXT:1980i units, acquired at $13B in 2024 alone. Mass production at 7nm was announced in July 2026. Yield verification remains outstanding, but the strategic intent is unambiguous.

AMD's demonstration of a 320B parameter model running at 20 tokens per second on the Ryzen AI Max Plus PRO 495 is the most disruptive data point of the cycle: capable local inference at scale now exists on a production platform, complicating the assumption that all AI workloads route through hyperscalers.

Rounding out today's briefing: Tesla halved LPDDR5 memory in its AI5 chip and cut AI6 LPDDR6 by a third — a potential signal for softening DRAM demand — and the Americas colocation market is projected at $53.66B by 2031, with liquid cooling growing fastest at 17.84% annually.

A YesWee production. Built using AI technology.

This episode includes AI-generated content.

No reviews yet