Jensen's banquet, Young Liu's 2x-capital EPS pledge, and a 9.64% GDP forecast — Taiwan is no longer pricing one cycle
TL;DR Foxconn didn't hit a 19-year high and break NT$4 trillion in market cap because of a quarter. It hit it because Chairman Young Liu told shareholders that **CSP capex is tracking above $700B this year and could reach $1 trillion next year** — and that single sentence reset the discount rate the market applies to Taiwan's entire AI supply chain. TAIEX closed at a record 44,732.94 on NT$1.81T turnover with NT$80.3B of foreign net buying. The Taiwan stats bureau lifted its 2026 GDP forecast to **9.64%, the highest since 2008**. What the market is pricing is no longer one quarter or one product cycle — it is **the floor of the hyperscaler decade**.
The pattern — regime, not cycle The message from Taipei over the last 72 hours is unusually coherent. Foxconn (2317-TW) closed limit-up at NT$289, market cap above NT$4 trillion, with a record NT$7.2 dividend approved. Young Liu set a long-term EPS target of **2x capital per share annually**, the next rung after five consecutive years above 1x. The same day, MediaTek (2454-TW) guided data center ASIC shipments to $2B in 2026 and "several billion dollars" in 2027. Quanta's (2382-TW) Barry Lam said AI server demand is robust into 2030. Compal (2324-TW) is doubling server revenue to 10% of sales. The single number that ties them all together is **$1 trillion** — the figure Liu repeated twice.
Jensen's banquet, priced into the curve May 28 in Taipei. Jensen Huang convened more than 30 Taiwanese tech CEOs and reiterated that AI demand is "10-100x the prior chip cycle," with another 10x in the next decade. Wistron's (3231-TW) chairman quoted Huang at his AGM the next day to dismiss bubble concerns; Wistron, Quanta, and Inventec all closed limit-up into Computex. Reporting points to Nvidia's cumulative Taiwan spend approaching **$150B**.
What matters is that these remarks are now the default assumption inside the supply chain, not aspirational headlines. Within hours of Liu citing $1T twice, TrendForce raised its 2027 global memory market forecast to $1.28T (+44% YoY). Marvell guided FY28 revenue to ~$15B, prompting Goldman to re-rate Taiwan ASIC, CCL, and OSAT names. Synopsys lifted FY26 guidance to $9.67B and stated that "the hyperscaler custom silicon trend is set."
The bottleneck is migrating from memory to wire The sharpest single warning came from Wiwynn (6669-TW) on May 28: AI data center shortages are spreading **beyond memory to networking ICs and other key components, with tightness extending through 2028**. The same week, Foxconn declared a push to become the **global No.1 in CPO switches**, targeting 10,000 units in 2026 and "multi-fold" growth from 2027. Airoha (6526-TW) taped out its 800G PAM4 DSP in Q4 2025. Wiwynn will showcase CPO at Computex with Ayar Labs, GUC, and Browave. The bottleneck is climbing the stack — memory → networking → optical interconnect — and Taiwan EMS/ODM names are positioning to own the **next** bottleneck before it arrives.
Memory repricing aligns the picture The memory cycle has lined up in the same direction. Morgan Stanley turned bullish on legacy DRAM on May 29, upgrading Winbond and Nanya to Overweight on the thesis that Micron's Taiwan-to-US equipment shift and SK Hynix's Wuxi DDR4/LPDDR4 cuts widen the 2H26 supply-demand gap. Samsung shipped the industry-first 12-layer HBM4E samples to AMD, Nvidia, and Google. Dell's COO warned of DRAM, NAND, CPU and HDD inflation through 2H — its AI server revenue was up **757% YoY**. DDR5 16Gb spot has stabilized at $41.9. Memory is no longer a bottom-fishing trade; it is a **seller's pricing regime**.
Lee Jae-yong's secret visit, MediaTek's refusal Reports surfaced this week that Samsung Chairman 이재용 made a secret Taiwan visit, offering MediaTek a bundled HBM/memory + foundry package. MediaTek Vice Chairman and CEO Rick Tsai publicly rejected it: "TSMC is our most important and longest-term partner." That same week TSMC confirmed it will **upgrade Kumamoto Fab 2 from 6-12nm to 3nm**, expanding capex to $17B — agreed directly between C.C. Wei and Japan's PM Takaichi. Morgan Stanley downgraded GUC to Equal-Weight (NT$4,888 target) on the same Samsung Foundry threat. The **frictional cost of the regime shift** is being levied first on boundary names like GUC.
Positioning - **High-conviction long:** Foxconn (2317) — the most direct beta to the $1T capex regime. Even at NT$4T market cap, its 19-year-high multiple still sits at cycle-average levels. - **Momentum long:** Wiwynn (6669), Quanta (2382), Wistron (3231) — claiming the next bottleneck in CPO and L10 AI servers. - **Structural long:** TSMC — SVP Kevin Zhang's framing of "energy efficiency, not raw compute, as the binding constraint" telegraphs pricing power on N2/A14. - **Caution:** GUC — exposed to Samsung Foundry design-service incursion; no momentum after the Morgan Stanley downgrade. - **Watch:** MediaTek (2454) — the first earnings print that has to validate the $2B → multi-$B ASIC transition.
Bottom line Foxconn's 19-year high is not a function of one quarter's print. It is the market accepting that the **$700B → $1T capex floor has hardened into a regime**. The next six months' risk is not a cycle peak but **validation of that floor**. Computex "sell-the-news" is noise; the signal is Young Liu's 2x-capital EPS pledge.
Key Sources: - Foxconn Chair Sees CSP Capex Hitting $1T in 2027; H2 Outlook 'Very Good' (Anue, 2026-05-29) - Foxconn Hits Limit-Up at 19-Year High, Market Cap Tops NT$4T on Record NT$7.2 Dividend (Anue, 2026-05-29) - Huang's Taipei 'Trillion-Dollar Banquet' sends Foxconn, Quanta, Wistron limit-up on AI tailwind (TechNews, 2026-05-29) - Wiwynn warns AI server supply bottlenecks may persist through 2028 (Anue, 2026-05-28) - Taiwan hikes 2026 GDP forecast to 9.64%, a 16-year high, on AI export boom (Anue, 2026-05-29) - MediaTek Targets $2B ASIC Shipments in 2026, Multi-Billion in 2027 (Anue, 2026-05-29) - plus 54 more
If this analysis was helpful · ☕ Support Us · ✈️ Telegram