TSMC $265B Arizona, Nvidia Screening & Kimi K3's Chip Selloff

**Jamie Call** (0:00)
AI Hardware & Chips Daily News.
I'm Jamie Call, thanks for joining me.
Today TSMC Arizona 2NN Timeline, Credibility Test for US Foundry Ambitions.
TSMC just raised its Arizona commitment to $265 billion, announcing four additional fabs targeting two nanometer and below. That's the headline, but the credibility test isn't the money, it's the timeline. The company has now outlined 12 fabs in total the Arizona site. FAB 21, the first phase, has reached 92% yield on 4nm production. That's a real number, and it matters. Yield at that level removes the Greenfield execution doubt that hung over US Foundry ambitions for the past three years. The question now shifts to what comes next, and how fast. The four new 2nm fabs don't have a published construction schedule. TSMC's CEO qualified the rollout as conditional on customer demand.
That's a standard commercial hedge, but it also means the 30% of global 2nm capacity the US is targeting by 2030 is still a projection, not a plan. The expansion also includes COWOS advanced packaging capacity in Arizona. That's the detail most coverage underweights. COWOS is the bottleneck in AI accelerator supply right now. It's the process that bonds chips together into the high-bandwidth packages that make Nvidia's GPUs work at data center scale. Moving that capacity out of Taiwan changes the geographic concentration of AI hardware supply in a way that fab announcements alone don't. The signal here is that TSMC isn't just building wafer production in the US. It's building the full supply chain segment that AI hardware actually depends on. TSMC's financial position backs the ambition. The company posted 77% year-on-year net income growth in the most recent quarter, revised full-year 2026 revenue guidance above 40%, and now derived 66% of revenue from high-performance computing. That's up 20 points sequentially.
This is a company with the cash to run the buildout. The execution risk is institutional, not financial.
While TSMC is expanding supply, Nvidia is tightening access.
Compliance vetting in Malaysia, Singapore and Japan has cut the authorized buyer list for AI Chips by more than half. Enterprises are now facing 8-12 week approval delays just to get through screening. Here's the thing, the important distinction here is that this isn't a policy change, it's enforcement becoming operational. Export controls have been escalating since 2022 What's new is that the compliance friction is now material to procurement timelines. Hyperscalers with direct supply relationships are insulated.
Smaller cloud providers and co-location operators aren't, and that gap is widening. No published criteria, no public appeals process. For a company that can't remediate a rejection it doesn't understand, that's not a temporary obstacle. It's a structural disadvantage. I think we may feel. The third major development connects back to both of those. Moonshot's Kimi K3 model, the largest open-weight release to date, is priced at $15 per million tokens against comparable US models running at 50 Benchmarks show competitive or better performance versus the current US frontier. The market's response was immediate. TSMC fell 7%.
AMD fell merely 8 ASML dropped 4.6. Nvidia fell 3.7. The investor logic is straightforward. If Chinese AI labs can close the performance gap at a fraction of the training and inference cost, the hardware demand thesis starts to look shakier. That's the foundational question Kimi K3 puts on the table. US semiconductor revenue growth forecasts are built on the assumption that USAI dominance drives hardware consumption. If those assumptions are wrong, the models need adjusting. Consider this. Two other developments round out the picture. Cadence's agentic EDA tool is now integrated into Rapidus' customer-facing design platform for its 2nm GAA process, targeting 2027 tape-outs. Zero yield data, zero volume commitments. But agentic design orchestration is now live in the production flow. It's an interesting bet. Use software maturity as a substitute for process maturity. India's revised semiconductor mission cuts Fabcapx subsidies from 50% to 40%, removes tech transfer support, and shifts the focus toward chip design and packaging.
The 28mm Tata Fab signals a realistic acceptance of non-cutting edge positioning. That's not a failure. It's a recalibration toward what India can actually execute.
The metrics worth tracking from here are specific. TSMC's construction schedule for the four new Arizona Fabs. Whether Fab 21 Phase 2 meets its 2027 volume ramp. Rapidus yield data if it appears. And how quickly analyst models adjust their US. Seneconductor TAM forecasts to reflect the cost compression coming out of Chinese AI labs. The US. Foundry buildout is real. The financial commitment is credible. But the 2 nanometer timeline is still a projection backed by conditional intent, not confirmed milestones. That gap between commitment and delivery is exactly what the next 12 months will start to resolve. Thanks for listening. This podcast was built using AI technology. A YesOui production.