UBS Raises TSMC's 2027 Capex Forecast to $90B as NVIDIA Revenue Share Surpasses Apple for the First Time
nashnova research
UBS raised its TSMC 2027 capex forecast from $80 billion to $90 billion, driven by 3nm capacity maxed out by AI chip orders at utilization rates above 100% — TSMC is shifting from a smartphone-chip foundry to an AI compute infrastructure company.
What exactly did UBS raise?
2027 capex forecast: $80 bn → $90 bn. 2028: $95 bn → $105 bn. 2026 stays at $63 bn.
This means → by 2028, TSMC's annual spend will be roughly 2.6× its 2025 level ($40.7 bn) — a pace of investment with no precedent in the foundry industry.
Three drivers behind the raise: faster capacity buildout (3nm and 2nm lines accelerated), equipment price hikes (vendors also demanding prepayments), and higher advanced-packaging spend.
How does the Q3 preview look?
UBS expects Q3 USD revenue up 14% quarter-on-quarter to about $45.7 bn, with gross margin at 67.1% — both at the high end of TSMC's guidance range.
Q4 revenue is forecast to rise another 10% to roughly $50.3 bn; full-year USD revenue growth is pegged at 41%, or $172.2 bn, above TSMC's own guidance of "slightly above 40%."
For 2027 UBS sees revenue growing another 40% to $240.7 bn, versus the Street consensus of 36%. This means → UBS is more bullish than most, betting AI chip demand will not decelerate.
Why is 3nm capacity so tight?
The customer mix on 3nm lines is transforming: cloud AI chips' share of N3 demand jumps from 36% in 2026 to 73% in 2027 — doubling in a single year.
In plain terms = a year ago the 3nm lines mostly made smartphone chips for Apple and Qualcomm; a year later, three-quarters of that capacity goes to AI chips from Nvidia, Google, and Amazon.
Supply cannot keep up: N3 utilization runs at 107% for full-year 2026 and 106% for 2027. Above 100% means the fabs are at full load and still short — they cope only by overtime and compressed maintenance windows. This is UBS's core justification for accelerated expansion.
Who is grabbing capacity — and who is giving it up?
Grabbing: Nvidia's Rubin platform needs roughly 384k N3 wafers in 2027; add Rubin Ultra and the companion Vera CPU and the total reaches about 731k wafers, or roughly 30% of total N3 demand. Google's three TPU lines rise from 191k to 464k wafers; Amazon's Trainium 3 from 115k to 181k.
Giving up: Apple iPhone processors drop from 442k to 128k wafers; Qualcomm from 129k to 51k; Intel outsourced PC processors from 137k to 70k.
This means → flagship phone and PC chips are migrating to 2nm, and the 3nm slots they vacate are immediately backfilled by AI chips — not a demand decline, but a generational handover.
How did Nvidia overtake Apple?
By UBS's estimates, Nvidia contributed about $23 bn to TSMC revenue in 2025 (18.8% share), already surpassing Apple's $20.9 bn (17.1%). By 2027 the gap widens: Nvidia $56 bn (23.3%) vs Apple $35.1 bn (14.6%).
Broadcom's contribution jumps from $5.28 bn in 2025 to $24.47 bn in 2027 — roughly 4.6×. This reflects the volume ramp of Google TPUs, Meta custom chips, and OpenAI custom silicon.
High-performance computing rises from 58% of TSMC's total revenue in 2025 to 74% in 2027; within that, AI-accelerator revenue doubles each year, reaching $104.7 bn in 2027. In plain terms = TSMC's revenue engine has switched from "selling smartphone chips" to "selling AI chips."
Can margins absorb this much spending?
Wafer ASPs are climbing: from $3,112 in 2025 to $4,290 in 2027 (+19.1%), then $5,124 in 2028 (+19.4%). The increase comes from TSMC's active price hikes and a product-mix shift toward higher-priced 3nm and 2nm nodes.
But the capex raise means more depreciation, trimming gross margin slightly: 2027 moves from 67.5% to 67.2%, 2028 from 68.1% to 67.8% — each down 0.3 percentage points. 2027 EPS dips from NT$149.14 to NT$148.98.
This means → higher depreciation compresses margins first; the payoff — more revenue — only materializes in 2028. UBS's logic is "spend now, earn it back later," but that hinges on 2028 revenue forecasts actually being met.
市场有风险，内容仅供研究参考，不构成投资建议。