The 2nm Pilgrimage: Lisa Su's Private Jet, Musk's Terafab, and the Week N2 Capacity Became a Scheduling Problem

A Private Jet, and Two Scenarios

4:30 PM, October 5, 2026. AMD CEO Lisa Su's private jet touched down at Taipei Songshan Airport on a two-day emergency itinerary with a single explicit objective: lock in additional TSMC 2nm capacity for AMD's next-generation 'Venice' EPYC server platform. It had been exactly five months since her last Taipei visit. The market quietly absorbed what it means for the CEO of the world's #2 datacenter chip designer to charter a private jet to Taipei twice in five months.

The same week, on a different front, Elon Musk publicly confirmed that Tesla's 'Terafab' — a Texas mega-fab initiative targeting chips capable of supporting ~1 trillion watts of AI compute — is in active discussions with TSMC. According to TechNews, cooperation is being framed under a 'two scenarios' structure: either TSMC establishes a dedicated fab for Terafab, or Tesla receives a carved-out allocation inside the existing Arizona/Taiwan capacity roadmap. Either way, the contours should surface before TSMC's October 15 analyst call.

Lisa Su's private jet and Musk's 'two scenarios' are not independent events. They are two faces of a single structural shift. TSMC's N2 capacity through the 2026–2028 window is effectively sold out, and the decision variable has moved from 'how much will you pay' to 'who gets a seat at the table.'

From Price Negotiation to Capacity Allocation

The old TSMC leading-node logic was simple: highest willingness-to-pay wins the biggest slice. The N2 of H2 2026 behaves differently. With two years of forward demand effectively pre-booked, TSMC is now acting as a central bank — designing which customer receives what percentage of allocation in which quarter. Three signals landed simultaneously this week.

First, GUC (創意電子), TSMC's ASIC design-services affiliate. GUC posted September 2026 revenue of NT$7.145B (~US$218M, +97.8% YoY, +22.3% MoM), a new all-time high. The headline number is impressive, but the real signal is one line buried in the release: HBM4E IP has been finalized on TSMC's N2P. HBM4E will be the memory backbone of 2027–2028 hyperscaler AI accelerators, and the fact that this IP 'runs first' on N2P means a substantial share of N2 capacity is pre-reserved for customers who come through the ASIC design-services funnel. GUC is no longer a design house — it is the VIP lane to N2 capacity access.

Second, TSMC's quiet backing of zero-revenue AI chip startups. Per TechNews and CTEE on October 5, TSMC has been selectively supporting early-stage AI chip startups with no product and no revenue. In the old regime, leading-node sample tokens were reserved for the top-10 customers with multi-trillion-NT$ revenue. Now TSMC is distributing them to startups it believes might become future winners. This is the classic central-bank reflex when capacity turns scarce — allocate on ecosystem contribution, not just purchasing power.

Third, the Terafab 'two scenarios' structure itself. In the old regime, a client of Tesla's scale would have received an immediate 'how much do you want to spend' response. The fact that TSMC is instead putting structural choices on the table — dedicated fab vs. existing-line carve-out — reveals capacity has reached the point where money alone cannot clear it. Musk's unusual decision to publicly confirm the discussions on his own account reinforces this: back-channel procurement is no longer possible at this volume.

The Numbers That Back the Scarcity

Evidence that the N2 scarcity narrative is not speculative shows up directly in September revenue across the Taiwan supply chain. Foxconn (2317) posted September revenue of NT$1.16T (+38.4% YoY, +25.7% MoM), its first-ever monthly print above NT$1T. Nanya Technology (2408) hit NT$45.1B (+576.6% YoY), a 10th consecutive monthly record. Winbond (2344) came in at +257% YoY. Aspeed (5274) doubled at +101% YoY. UniEpitaxy (3081) hit +190% YoY on silicon photonics. When every layer of the semiconductor stack sets records simultaneously, demand is not a single-SKU price rebound — it is a forward pre-order on the entire AI accelerator roadmap.

DDR5 16Gb spot trading at US$58.17 on October 6, and Micron's decision to roughly double FY2027 capex to US$50B+ (+80% YoY), are the memory-side responses. But the real top-stack bottleneck is unchanged: the N2 capacity those memories plug into.

Positioning: October 15 and Beyond

TSMC (2330) closed October 5 at a record NT$2,580 with market cap at NT$66T, extending its lead as the world's largest semiconductor company. Foreign brokers already revised EPS estimates and target prices upward ahead of the call. The question for investors watching the October 15 analyst call is not traditional guidance. It is whether TSMC publicly reveals the allocation function — how N2 capacity will be distributed across Terafab, AMD, Apple, and hyperscaler ASICs through 2028.

Lisa Su's private jet is not a tactical signal. It is a structural one. In the age of capacity scarcity, the designers who have 'a seat at the table' for N2 access become the winners of the 2027–2028 AI cycle. Those left standing fall two generations behind, regardless of how much cash they wave. This was the week that inflection point was declared out loud.

Key Sources: - Lisa Su is expected to arrive in Taipei by private plane at 4:30 pm (TechNews, 2026-10-05) - TSMC and Tesla's Terafab are expected to have two options (TechNews, 2026-10-05) - GUC September revenue hits record, HBM4E IP finalized on TSMC N2P (TechNews, 2026-10-05) - Why TSMC bets on zero-revenue AI chip startups (CTEE, 2026-10-05) - TSMC market cap hits record NT$66T ahead of earnings call (UDN, 2026-10-05) - plus 7 more