According to the Zhitong Finance app, Bernstein upgraded CPU to a new growth driver for TSMC (TSM.US) three months ago, and its latest report has further elevated this assessment: the real change is not in GPU or
ASIC XPU, while demand driven by the large-scale deployment of custom CPUs by Amazon and Google has exceeded previous expectations.
The new volume comes from "ASIC CPU," not AI accelerators.
Amazon and Google customize CPUs for their own use, deploying them internally only and not selling them externally; these are essentially ASICs, which Bernstein refers to as “ASIC CPUs” to distinguish them from those designed for AI training and inference.
ASIC and XPU. Demand for both is growing, but the recent surprise has been that agent-type workloads are accelerating CPU demand, particularly the custom CPUs being deployed by two cloud providers.
The report highlights another structural difference: on ASIC CPUs, Amazon and Google are far more self-sufficient than XPU and rely much less on ASIC design service providers. Amazon's
Graviton5 is the fifth-generation in-house CPU, built on a 3-nanometer process with 192 Arm Neoverse V3 cores and 1720
One billion transistors, with all design and production coordination with suppliers handled in-house; Google previously had two generations of CPU design experience (Maple and the first version of Axion), second-generation Axion
Designed in-house and outsourced solely for production to GUC, this chip is benefiting from agent workloads and ramping up rapidly from late 2026 through 2027.
The key point is augmentation, not replacement: these in-house CPUs do not displace Intel and AMD's general-purpose
not CPU, but adds new demands on top of both, and ultimately both land on TSMC wafers—Graviton5 uses 3nm, and the second-generation Axion lands on the N4 family.
CPU revenue: $24 billion to $50 billion, increasing from 7% to 16%
Bernstein estimates TSMC's CPU revenue at approximately $5 billion in 2023, $10 billion in 2024, $14 billion in 2025, and $24 billion in 2026.
billion USD, projecting to approximately $40 billion in 2027 and $50 billion in 2028; the share of total revenue accordingly rising from 7% to 16%. For clarity, this calculation excludes Apple’s PC
The custom-built CPU used on it (e.g., the A18 shared by iPhone 16 Pro/Pro Max and MacBook Neo)
Pro), which is also the reason for the slight differences from the historical data in the previous report.
On the customer side, AMD will remain the largest single CPU customer from 2026 to 2028, maintaining a share of 50% to 60%; Intel's share will decline from 30% in 2025
The high above has dropped to approximately 10% by 2028; Google and Amazon were still categorized under "other" in 2025, but their contributions rose significantly thereafter.
Process differentiation: N3 and N2 upgraded, N5 and N7 under pressure
Demand is not uniformly strong. PC and low-end smartphone demand fell short of expectations, particularly at the low end, causing slightly softer demand for N5 and N7; N3 and N2 are being impacted by XPU, CPU, and high-end smartphone SoCs
Shared, demand thus stronger, Bernstein accordingly raised revenue contribution for both.
A14 is scheduled to succeed N2, beginning to generate significant revenue in the second half of 2028, with the transition from N2 to A14 slightly longer than two years. The initial adopters include Apple, Qualcomm, MediaTek, and
AMD; NVIDIA will also adopt A14, but not until 2029—consistent with its decision not to migrate to N3 as early as other XPU suppliers, leveraging CUDA
The pace has not affected its leadership in the AI XPU market, thanks to the advantages of system-level integration. Overall, A14 will still generate more revenue for TSMC at this stage than N2 and all previous processes.
CapitalExpenditure: $77 billion in 2027, but spending intensity continues to decline
Demand revisions have led Bernstein to raise its capital expenditure forecast to $77 billion in 2027 and $83 billion in 2028, up from $63 billion in 2026.
billion U.S. dollars. We maintain our previous assessment: capital expenditures as a percentage of revenue will gradually decline, consistent with TSMC’s recent earnings commentary.
Capacity for advanced packaging has also been slightly increased. TSMC’s own CoWoS equivalent capacity is expected to rise from approximately 60,000 wafers per month by the end of 2025 to about 135,000 in 2026 and approximately
175,000, approximately 180,000 per month by 2028; if including the packaging and testing capacity of OSATs such as ASE and Amkor, as well as Intel’s EMIB-T equivalent capacity, the total market capacity will reach approximately
From 65,000 units/month to approximately 380,000 units/month by the end of 2028. EMIB-T
Base yield still needs improvement, but MediaTek's guidance last week indicated that improvements are on track, so Bernstein has maintained its capacity forecast.
Earnings and Valuation: 35% EPS CAGR supported by a 20x forward P/E ratio
Under the revised model, TSMC's USD revenue is projected to increase by 44%, 40%, and 27% in 2026, 2027, and 2028, respectively, reaching approximately $176 billion, $247 billion, and $315 billion.
billion U.S. dollars; considering the drag from N2 and overseas capacity, gross margin remains roughly flat (66.6% in 2026, 65.7% in 2027), with earnings per share growing at the same pace as revenue, 2026 to 2028
Annual compound growth rate of approximately 35%. Revised EPS for 2026: NT$113.40 (previously NT$110.70), for 2027: NT$157.79 (previously
144.85); 5% higher than market consensus for Q4 2026, 6% higher for 2027, and 10% higher for 2028.
Target price of NT$3,828 based on a 20x forward P/E multiple applied to EPS of NT$191 for Q4 2027 to Q3 2028; US ADR target price of $600
Dollars are converted directly at the exchange rate. Near-term, the report forecasts revenue of NT$1.699 trillion for Q4 2026, a 62.4% year-over-year increase, 0.5% below market consensus; earnings per share exceed consensus.
5% — Revenue side is not the main advantage; the excess comes from profit margins and fee assumptions.
The valuation multiple should be noted: TSMC's current forward P/E is approximately 16x, compared to a seven-year average of 17.5x and one standard deviation above the mean at 21.1x. The target price uses a 20x
It has significantly exceeded the historical average and is approaching the upper bound of one standard deviation above the mean, meaning that within the 60% upside potential, valuation recovery and earnings upgrades each contribute a portion.