UMC Just Ended a Two-Decade Price-Cut Habit — and Is Paying $3.2B to Join the AI Optics Race

A 10% blended ASP hike for 2027 is the real headline; the silicon photonics capex is the bet that the optical tape has stopped believing in

UMC · Earnings Call · 2026-10-28

The Year the Haircut Became a Raise

For UMC, the annual January price negotiation has been a slow bleed for years. On the January 2026 call, UBS's Sunny Lin framed the baseline plainly: “in 2024 and 2025, basically, you provided roughly mid-single-digit type of price reset for across the board” — Sunny Lin, Analyst · 2025-10-29. Management's answer then was tentative — “we do anticipate a more favorable ASP environment in 2026 versus 2025” — Jason Wang, President · 2026-01-28 — a better environment, not a reversal. Nine months later, Jason Wang is calling a regime change. “We expect 2027 to mark a turning point in our blended ASP trajectory, moving from an environment of ASP erosion toward a meaningful positive uplift beginning in the first quarter” — Jason Wang, Chief Executive Officer (CEO) · 2026-10-28, with a blended ASP increase of about 10% year over year. That is not a mix trick alone — management explicitly says it blends like-for-like price with a richer product mix, and the justification is honest about the two drivers: customer demand in selected technologies now exceeds available capacity, and raw material plus depreciation costs are rising. Customers, in other words, are being asked to share the freight. The analyst pushback was exactly right, and it is where the story lives. Haas Liu of BofA noted utilization has climbed from 85% in Q2 to above 90% in the second half — a 16-point swing from Q4 last year to guidance — while pricing stayed flat. Why would customers suddenly accept a double-digit increase? Wang's answer: “The product mix is a part of it, and part of it is really reflecting the inflationary cost increase” — Jason Wang, Chief Executive Officer (CEO) · 2026-10-28. That is a price increase with a cost excuse attached — which means it is more defensible than a pure demand grab, but also more reversible if the cost curve bends.

$3.2 Billion, Customer-Backed, and Late

The board approved roughly $3.2 billion of additional capex on top of the ~$2 billion 2026 cash budget. Management is insistent this is not a return to the broad-based mature-node capex wars of the last cycle:

This is not a broad-based capacity expansion. We are directing our capital toward customer-backed growth opportunities that will fuel UMC's next phase of growth.
Jason Wang, Chief Executive Officer (CEO) · 2026-10-28

The targets are specific: ~18,000 wafers a month for silicon photonics, 22,000 for DTC and silicon capacitor capacity, and 20,000 for TFLN modulators. Silicon photonics goes in Singapore, the capacitor work in Tainan, TFLN in the 8-inch fabs. The technical claim is that UMC skipped 8-inch PIC and went straight to 12-inch, and that the decision paid: better propagation loss than peers, a 1.6T platform already qualified by multiple hyperscalers, and a roadmap to 300G and 400G per lane for 3.2T systems. On TFLN, Wang is blunter than usual — “I think we much ahead about anyone else in the TFLN development” — Jason Wang, Chief Executive Officer (CEO) · 2026-10-28. The deep trench capacitor for mobile power delivery enters mass production this quarter; the substrate-embedded version for AI servers has taped out. But here is the arithmetic problem the call did not resolve cleanly. Almost all of this new capacity — the 60,000 wafers of monthly capacity across the three platforms — does not start installing until late 2028 and does not reach scale until 2030. And yet Wang told Gokul Hariharan of JPMorgan that AI-related revenue, roughly $300 million in 2026, “will probably exceed a quarter of our revenue” — Jason Wang, Chief Executive Officer (CEO) · 2026-10-28 within three years — i.e., by 2029. Those two statements can only be reconciled by reconfiguring existing fabs (converting an 8-inch line to TFLN, for instance) and by mix. Notably, the capacity configuration mismatch Wang now cites to explain why 92% utilization is not 100% loaded is exactly that reshuffling. 2027 AI revenue will be capacity-constrained, by his own admission.

The Margin Nobody Should Expect Back

The most revealing exchange was on gross margin. A 10% ASP increase plus full utilization would normally be a margin story. Instead, CFO Chitung Liu pre-empted it:

Depreciation is going to be certainly higher and maybe a couple of years in a row. We are talking about low single digit depreciation growth, at least for 2027, if not more for 2028... for gross margins for next two years at least, depreciation, along with the rising raw material cost, is going to put pressures on our overall financial gross margin.
Chitung Liu, Chief Financial Officer (CFO) · 2026-10-28

That is a quiet but material walk-back from prior framing. On the July 2025 call, the same CFO said the depreciation curve would peak and that “'26 and '27, the increased magnitude will be a lot less, could drop down to single digits” — Chi-Tung Liu, Chief Financial Officer · 2025-07-30 — with the implication that gross margin had a floor. The accelerated AI capex has now pushed that floor out by two years. Management's substitution is EBITDA: the structural profitability and EBITDA margin are "intact" and "in good shape" versus the prior cycle. Investors should understand the trade: UMC is selling its gross-margin recovery to buy a position in AI optics, and it is telling you so.

The Tape Disagrees With the Timing

Here is the uncomfortable confluence. The AI data center complex, up 53% over 360 days across 45 tickers, is down 5.4% over the last 90 days on a 6-up/50-down breadth. The optical interconnect cohort is worse: co package optics is up 85% over a year but down 5.2% in the last 90 days across all thirteen names, and per-lane bandwidth stories like G-per-lane are down ~7% in the same window. The market has spent the last quarter de-rating exactly the silicon-photonics supply chain UMC is now accelerating into. Wang's answer to the oversupply question — “Overcapacity is an industry question, not just a UMC question” — Jason Wang, Chief Executive Officer (CEO) · 2026-10-28 — is reasonable, and the commitment-backed framing (capacity tied to customer product ramps, built in phases) is the right mitigation. But the peer set is now crowded with the same intent: Amkor is adding advanced packaging capacity, ASM is talking high-performance DRAM for HBM, Teradyne is shipping AI accelerators, Celestica is scaling integrated optics. Every one of them reported in the past week. UMC's differentiating claim is that it owns the front-end photonics process rather than assembling someone else's — a real moat, if the demand curve holds at 3.2T. The call, in sum, is a confident company making a genuine strategic pivot at the moment its end-market tape turned cautious. The pricing turn is the part that is already real; the AI revenue mix is the part that requires a capex cycle to land on schedule and an optical market that re-accelerates. Two years of gross-margin pressure is the price of admission.