MediaTek: TPU Upside Overlooked as AI Agent Phones Could Revive Smartphone Business in 2027

Core Conclusion

Rating Overweight with new price target NT$2,588 (17% upside from NT$2,215). MediaTek is now the top pick, replacing TSMC. The market underappreciates two overlapping drivers: Google TPU volume and ASP trajectory, and the potential for on-device AI agents to reignite smartphone SoC demand in 2027. Near-term smartphone weakness is well-known and priced in; the April 30 earnings call should serve as a clearing event. Upward EPS revisions are driven by TPU.

Evidence Chain

1. TPU: Confirmed 3nm Volume and 2nm ASP Upside

• Google’s Cloud Next announced TPU 8t (training) and 8i (inference). Supply chain reticle-size checks confirm MediaTek’s 3nm chip is TPU 8t, not the lower-end v7e as some assumed. This validates MediaTek’s SerDes IP capability and mass production readiness.
• MediaTek guided AI ASIC revenue >US$1bn in 2026. Base case assumes 400k units at US$4,000 ASP. For the 2nm Humufish (v10), recent checks indicate two compute dies and two I/O dies — double the Zebrafish design. Combined with SerDes upgrade and larger packaging, ASP is now estimated at US$12–15k, up from prior US$8–10k.
• TSMC’s 1Q26 earnings call indicated willingness to open computing dies to third-party packaging (e.g., Intel EMIB), supporting higher volume for MediaTek’s 2nm TPU without CoWoS capacity constraints. This reduces execution risk.

2. Smartphone: Weakness Expected; AI Agent Phones Add Upside

• 2Q26 revenue guidance implies ~5% Q/Q decline to NT$140.7bn, below consensus of NT$141.4bn. Foundry supply chain shows MediaTek and Qualcomm cutting 4nm wafer orders by 20–30k wafers (15–20mn units), affecting Dimensity 7000/8000 series (ASP US$30–60). This is a China-driven correction due to memory cost hikes and lack of compelling AI smartphone features.
• Android AI agents launched on Galaxy S26 and Pixel 10 in early 2026. Multi-step app tasks (ordering, scheduling) require token-intensive inference. On-device computing reduces cloud costs and could drive a replacement cycle. Google I/O on May 19 is a catalyst for wider ecosystem adoption. This underpins 2027–28 SoC shipment upside beyond base case.

3. Earnings Estimate Revisions

• 2026e EPS raised 17% to NT$68.81 (from NT$58.70), 2027e +10% to NT$129.50, 2028e +31% to NT$172.53. Revisions driven by modem share gains (Apple, Pixel) in 2026 and higher 2nm TPU ASP in 2027–28.

Key Risks

• Smartphone demand deterioration extends beyond 2Q26; if China recovery fails or memory costs persist, volume may undershoot. Bear case assumes 10x 2027e P/E (NT$1,320).
• TPU ramp delays from SerDes IP performance or mass production schedule. If 2nm Humufish slips or ASP is lower, EPS upside shrinks.
• Gross margin dilution from TPU mix (ASIC margins typically lower than smartphone SoC) could compress net margins faster than expected, especially if R&D spend stays elevated.
• Competition: Broadcom’s 3nm TPU (SunFish) or in-house Google designs could reduce MediaTek’s share beyond v8/v10.

Valuation and Trade Implication

• New PT of NT$2,588 is derived from residual income model (cost of equity 9.2%, beta 1.2, risk-free 2%, intermediate growth 9%, terminal 3%). Implies 20x 2027e EPS (NT$129.5) — one standard deviation above the 18x long-term average, justified by TPU-driven earnings acceleration.
• Bull case: NT$3,550 (27x 2027e) assumes stronger AI agent adoption and TPU volume; bear case: NT$1,320 (10x).
• Key catalysts: Apr 30 earnings call (remove smartphone overhang), May 19 Google I/O (AI agent strategy), Computex (MediaTek keynote). Relative to TSMC, MediaTek is less consensus-owned and offers more upside to the new PT.

Appendix Data Summary