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

Jamie Call (0:00)

AI Hardware & Chips Daily News.

I'm Jamie Call. Thanks for joining me. Today, TSMC Texas Terafab, Strategic Pivot or Defensive Hedge.

TSMC and TSMC is now in early confirmed discussions with Elon Musk's Terafab project for a potential Texas FAB partnership. That's not a rumor anymore. Both sides have acknowledged it publicly, which changes the weight of this story considerably. The background matters here briefly. Musk launched Terafab in 2025, driven by Tesla's long-term chip demand. Intel joined in April 2026, bringing its 14A process to the table. Sansum was in earlier conversations that didn't progress. Now TSMC is the name in the room, and that tells you something about how seriously the US domestic Sani conductor agenda is being pressed. The important distinction is what kind of involvement TSMC might take. Multiple structures are being explored. TSMC could hold an only ship's stake and operate the fab, or SpaceX and Terafab could hold majority control, with TSMC playing a more technical role. That difference matters enormously for CAPEX allocation, technology transfer risk, and how TSMC's own margin profile absorbs the commitment. The structure is unresolved, and until it isn't, Terafab remains more strategic signal than confirmed buildout. Set against that uncertainty, TSMC's broader US commitment is not ambiguous at all. The company raised its 2026 CAPEX guidance to $60-64 billion, and separately announced an additional $100 billion commitment to Arizona. That's a multi-year spending signal that significantly exceeds the prior three-year period. Texas, if it proceeds, would be a second US hub with potentially six advanced fabs. The honest question is whether this expansion is strategic ambition or defensive positioning. TSMC is clearly responding to geopolitical pressure to diversify outside Taiwan. The Terafab talks fit that pattern. What's less clear is whether the economics justify the pace. Depreciation from this CAPEX wave will pressure margins. Gross margins above 60% are the benchmark investors' track. That's not guaranteed through a ramp cycle of this scale. While TSMC expands westward, China has been building in a different direction. Chinese fabs accumulated 343 immersion DUV lithography machines by early 2026

270 of those are ASML NXT, 1980i units. $13 billion was spent on 90 additional units in 2024 alone. Here's the thing. The signal here is scale and timing. China moved aggressively before tighter US export restrictions could close the window.

EUV access remains blocked, but 343 immersion DUV machines represent a meaningful 7 manometer production capability without it. China announced mass production capability at that node in July 2026

Yield rates and commercial viability haven't been independently verified, so the actual manufacturing output from this stockpile is still an open question.

What isn't open is whether the accumulation was deliberate. It clearly was.

The other major development worth tracking is AMD's demonstration of a 320 billion parameter model running locally on the Ryzen AI Max Plus PRO 495 The model ran at 20 tokens per second on device.

That's not a research to know. That's a production platform benchmark at a scale that until recently required cloud infrastructure. The implication for the broader AI infrastructure thesis is worth sitting with. If capable local inference at this parameter scale is commercially viable, some portion of demand that was assumed to flow toward data center capacity may not. It doesn't collapse the cloud buildout thesis, but it does complicate the assumption that all AI workloads route through a hyperscaler. Enterprise buyers now have a real on-device option. Two shorter signals round out today's picture. Tesla half the Lpddr5 memory in its AI5 chip and cut Ai6 Lpddr6 by one third. Whether that reflects Tesla's specific supply pressure or a broader constraint in the memory market isn't confirmed. If other chip makers follow, DRAM and HBM Denon projections could soften faster than current CAPEX cycles anticipate. Consider this. On the infrastructure side, the Americas data center co-location market is projected to reach $53.66 billion by 2031, growing at 12.13% annually.

Liquid cooling is expanding at 17.84% annually, which is the faster number. Power and cooling are becoming the harder constraints, more acute in some respects than wafer supply. That's the shift worth watching, as hyperscalers chase AI-ready capacity. Two things to keep close. First, the Terafab ownership structure. When that resolves, it'll clarify how much of TSMC's US expansion is driven by genuine commercial logic versus geopolitical obligation. Second, TSMC's advanced node pricing. A 5-10% increase is expected for 2027

Whether customers absorb it without pushback will tell you a great deal about how much pricing power TSMC actually holds, as Intel's 14a enters risk production on the same timeline. The race to leading edge is getting crowded. The economics of getting there are getting harder. That's the frame for everything else in this space right now.

Thanks for listening. This podcast was built using AI technology. A YesWee production.