Bernstein Bullish on Semiconductors: TSMC Price Hikes, Memory Upcycle, and ASIC Expansion
nashnova research
Can you still buy semiconductors here?
Bernstein's verdict: prior corrections have priced in most risks; the sector remains a conviction long.
Key headwinds include rising 10-year and 30-year U.S. Treasury yields pushing up financing costs, plus uncertainty around AI safety and large-model iteration.
This means → macro drag is real, but major AI customers sustain strong demand even as costs rise. Cost pressure falls harder on non-AI segments.
In plain terms = the broader environment is tougher, but the companies actually spending big on AI chips haven't pulled back — everyone else has.
What gives TSMC its pricing power?
TSMC's preliminary Q3 revenue came in at roughly NT$1.494 trillion, up 51% year-on-year. Bernstein forecasts 41% full-year 2026 revenue growth, with Q4 sequential growth of ~10% or higher — and notes the forecast still has room for upward revision.
The pricing engine: sub-3 nm advanced process and 2.5D/3D advanced packaging — stacking multiple chips together in a single package — remain in persistent shortage. This means → TSMC can keep raising prices through a "process upgrade + advanced packaging" combination, converting AI compute intensity directly into revenue growth.
Bernstein expects 2027 advanced-node price hikes of high-single to low-double digits, mature-node hikes of mid-single digits, starting to flow through in the first half of next year.
Projected capex for 2026–2028: $64 billion, $75 billion, $82 billion — a year-on-year ramp that signals confidence in capacity expansion.
Who gets TSMC's capacity — CPUs or ASICs?
Bernstein projects TSMC's 2027 CPU wafer revenue at close to $40 billion, roughly matching custom ASIC/XPU — purpose-built AI chips designed by hyperscalers — wafer revenue, each accounting for about 15% of total sales.
Incremental capacity tilts toward CPU and ASIC/XPU. Google and Amazon's in-house CPUs benefit directly. XPU also contributes the strongest packaging revenue stream.
This means → TSMC's capacity allocation is shifting from "open to all" to "AI hyperscalers first" — non-AI businesses will find it increasingly hard to secure slots.
Can memory price hikes last?
Third-party forecasts see Q4 DRAM and NAND prices rising 15–20% quarter-on-quarter, slightly above Bernstein's own estimate. Samsung's preliminary operating profit hit KRW 107.4 trillion, up 782.5% year-on-year — price hikes are converting into profit.
But HBM4 — the next-generation high-bandwidth memory chip critical for AI training — faces schedule delays; 2027 HBM price gains may fall short of expectations, and rising employee bonuses cap the upside surprise.
Bernstein places more weight on long-term supply agreements boosting earnings visibility, plus shareholder returns driving valuation re-rating. In plain terms = short-term pricing has a ceiling, but long-term locked-price contracts make profits more predictable — that is the real foundation for a higher multiple.
Preferred picks: Samsung (building HBM4 leadership + low valuation) and CXMT (China domestic substitution + execution + relative valuation edge). China's expanding memory supply also poses long-term competitive pressure on the global Big Three.
Why did Bernstein sharply raise its ASIC market forecast?
The 2028 global data-center XPU/ASIC market estimate was raised from $140–160 billion to $250–300 billion; 2026 sits at $50–70 billion, 2027 at $130–150 billion.
ASIC revenue share climbs from 11% of sales in 2026 to 65% in 2028; profit share rises from 15% to 68%. This means → within two years, ASIC will shift from a fringe business to the dominant profit driver.
Google's TPU v8 has entered volume shipment; v9 substrate yields are improving, setting the stage for 2027 delivery and earnings upgrades.
MediaTek joins the preferred list for the first time — the core thesis: partnering with Google to build TPU and broader custom ASIC business, potentially reshaping the company's growth and margin profile.
What does Bernstein's preferred list signal on valuations?
Preferred names: Samsung, TSMC, Kioxia, CXMT, MediaTek; Micron is the only U.S. semiconductor company on the list.
Largest implied upside: Samsung preferred shares at roughly 92%, SK Hynix target price implying over 60% potential gain.
Mature-node companies are treated very differently: UMC stays at "underperform" with an implied 30.2% downside; Vanguard International Semi at "market perform," just 4.0% upside; Novatek at "market perform," implied 7.7% downside.
This reflects another core thread in Bernstein's report: the valuation divergence between advanced-node and mature-node / consumer-electronics names is widening — TSMC's quarterly guidance this season will be the key checkpoint for whether the pricing thesis delivers.
市场有风险，内容仅供研究参考，不构成投资建议。