聯發科

[audit 2026-05-24] drivers edited by aa9440307@gmail.com: pe_applied 40→50, pe_terminal 25→35 ## Notes (comps_pe — peer-median anchor, top-down EPS retained, multi-year path) ### Headline - Implied: **TWD 948 / 1,410 / 2,512** (low / mid / high) - Reference: TWD **3,860** → mid upside **-63.5%** (comps framework signals overvaluation) - Peer median Fwd P/E CY26E: **15.32x** (n=6 cohort; range 11.65x–46.27x) - 2454.TW CY26E EPS: **TWD 92.16** (top-down; market-implied at consensus 28x ≈ TWD 138; divergence -33.15% — top-down retained because BOM places TPU rev in CY27+, not CY26) ### Peer cohort (selected for end-market overlap) | Peer | Theme | Spot | Fwd P/E CY26E | |---------|--------------------------------|------------|---------------| | QCOM | handset SoC + modem | USD 180.50 | 14.98x | | AVGO | custom ASIC / hyperscaler | USD 414.14 | 46.27x | | MRVL | custom ASIC / hyperscaler | USD 78.20 | 25.64x | | NXPI | auto + industrial broadline | USD 232.50 | 15.66x | | 3034.TW | Taiwan fabless (Novatek) | TWD 545 | 11.65x | | STM | auto + industrial broadline | USD 38.20 | 13.40x | ### Multi-year forward (mean-reversion P/E: FY+1=15.32, FY+3=15.32 → flat at median) | Year | EPS mid | P/E | Implied raw mid | |--------|---------|--------|-----------------| | CY26E | 92.16 | 15.32x | TWD 1,412 | | CY27E | 108.92 | 15.32x | TWD 1,669 | | CY28E | 156.99 | 15.32x | TWD 2,405 | Note: `pe_terminal` deliberately = peer median (not premium) because by CY28E the TPU contribution is *in the EPS numerator*; applying a TPU premium to the multiple would double-count. To see the upside case, swap pe_terminal for P75 (23.14x) in the calculator → CY28E implied ≈ TWD 3,632, finally approaching spot. ### TPU royalty assumption (the dominant lever) Modeled at 10% of die ASP as design-partner royalty. This is the midpoint of a 5–15% band: - 5–8% for pure backend physical-design service (v8t I/O die) - 10–12% for full architectural co-design (v8i Compute die) - 12–15% for full design ownership with IP licensing (none of MediaTek's current TPU work meets this bar) Sensitivity: at 5% royalty, CY28 TPU rev contribution halves to ~TWD 181B; at 15%, ~TWD 544B. Single-lever sensitivity, linear flow-through. A dedicated TPU royalty sensitivity tab should be built before publishing any external target. ### EPS path drivers - **CY26E (TWD 92)**: pure handset SoC + WiFi/connectivity + auto/IoT; TPU = $0 per BOM. - **CY27E (TWD 109, +18%)**: ~½ handset/connectivity growth, ~½ first-year TPU royalty (~TWD 73B attributable rev). - **CY28E (TWD 157, +44%)**: dominated by v8t I/O die ramp 0.5M → 7M units (~TWD 363B attributable rev). Lowest confidence in the path — both the 7M unit forecast and 10% royalty are "low confidence" per BOM. ### Why the asymmetric range - Low (TWD 948) = P25 multiple × low EPS — comp-only floor with conservative ramp - Mid (TWD 1,410) = peer median × top-down EPS — central comp framework signal - High (TWD 2,512) = P75 multiple × high EPS — best case under AVGO-adjacent re-rating The full -64% gap vs spot reflects that **the market is paying for AVGO-like TPU optionality, not earnings**. Reader who believes that premium is justified should re-anchor headline to market-implied EPS (~TWD 138) at median multiple → mid implied lifts to ~TWD 2,114, still -45% vs spot. ### Caveats - 2454.TW spot TWD 3,860 sourced from weilab stock_quotes (fresh 2026-05-22). ✓ - Peer Fwd P/E values for QCOM / NXPI / STM / 3034.TW are sub-agent estimates from public IR (not fresh weilab quotes for peers); fresh peer prices would tighten the band slightly but not change the framework signal. - Consensus Fwd P/E (28x) is FactSet-style estimate, marked `stale_estimate` — Q1/Q2 2026 verified consensus would sharpen the reconciliation block. - TPU "ASP" anchors are Google capex per chip, not MediaTek revenue. Royalty assumption (10%) is the dominant single lever. - AVGO (46.27x) is a clear cohort outlier; its inclusion widens P75 to 23x, which is appropriate given MediaTek's structural TPU exposure deserves *some* AVGO-adjacent premium — but median (15.32x) excluding AVGO would only drop to ~15.0x, framework conclusion unchanged.