Bernstein: ASIC CPUs further bring in additional demand, raising TSM.US's target price to $600

The Zhitong Finance App learned that three months ago, Bernstein listed CPUs as the new growth driver for TSM.US (TSM.US), and the latest report took this judgment one step higher: the real change was not in GPUs or ASIC XPUs, and the CPUs developed by Amazon and Google began large-scale deployment, driving demand beyond previous expectations.

The new volume comes from “ASIC CPUs”, not AI accelerators

CPUs customized by Amazon and Google for personal use are only deployed in-house and are not sold externally. In nature, they are ASICs. Bernstein calls them “ASIC CPUs” to distinguish them from ASIC XPUs for AI training and inference. Demand for both is growing, but the recent surprise is that intelligent workloads are speeding up CPU demand, especially self-developed CPUs being deployed by two cloud vendors.

The report highlights another structural difference: Amazon and Google are far more self-sufficient than XPU on ASIC CPUs, and are far less dependent on ASIC design service providers. Amazon's Graviton 5 is already a fifth-generation self-developed CPU with a 3-nanometer manufacturing process, 192-core Arm Neoverse V3, and 172 billion transistors; Google has had two previous CPU design experiences (Maple and Axion first edition), and the second-generation Axion was designed by itself and only handed over production arrangements to Creative Electronics (GUC). The chip is benefiting from intelligent workloads and has climbed rapidly from the end of 2026 to 2027.

The key point is superposition rather than substitution: these self-developed CPUs don't crowd out Intel and AMD's general-purpose CPUs, but they add demand on top of both, and eventually fall on TSMC's wafer — the Graviton 5 uses 3 nm, and the second-generation Axion falls on the N4 family.

CPU revenue: $24 billion to $50 billion, up 7% to 16%

Bernstein pushed TSMC's CPU revenue from about 5 billion US dollars in 2023, 10 billion in 2024, 14 billion in 2025, and about 24 billion US dollars in 2026, all the way to about 40 billion in 2027 and 50 billion US dollars in 2028; the share of total revenue increased from 7% to 16%. In order to clarify the caliber, this estimate excludes self-developed CPUs used by Apple on PCs (for example, iPhone 16 Pro/Pro Max and MacBook Neo share A18 Pro), which is why there is a slight difference in historical data from the previous report.

In terms of customer structure, AMD remained the largest single CPU customer from 2026 to 2028, accounting for 50% to 60%; Intel's share fell from a high of over 30% in 2025 to about 10% in 2028; Google and Amazon were also classified as “other” in 2025, and their contributions increased markedly thereafter.

Process differentiation: N3 and N2 upgrade, N5 and N7 pressure

Demand is not improving across the board. PCs and low-end smartphones were weaker than expected. The low-end was particularly evident, weakening demand for N5 and N7; N3 and N2 were shared with XPU, CPU, and high-end phone SoC, making demand stronger as a result, and Bernstein raised the revenue contribution of the two accordingly.

As planned, A14 took over N2 and began contributing significant revenue in the second half of 2028. The pace from N2 to A14 was only slightly longer than two years. The first customers to adopt were Apple, Qualcomm, MediaTek, and AMD; Nvidia is also using the A14, but is expected to be in line with 2029 — in line with not migrating to N3 as early as other XPU vendors back then. With the advantages of CUDA and system-level integration, this pace has not affected its leading position in the AI XPU market. Overall, A14's revenue to TSMC at the same stage will still exceed N2 and all previous processes in US dollars.

Capital expenditure: $77 billion in 2027, but expenditure intensity is declining

In response to rising demand, Bernstein raised the capital expenditure forecast from $63 billion in 2026 to $77 billion in 2027 and $83 billion in 2028. At the same time, it maintains the previous judgment that capital expenditure as a share of revenue will gradually decline. This is consistent with TSMC's statement in its latest financial report.

Production capacity for advanced packaging has also increased slightly. TSMC's own CoWOS equivalent production capacity increased from about 60,000 tablets/month at the end of 2025 to about 135,000 pieces/month in 2026, 175,000 in 2027, and 180,000 pieces/month in 2028; if the equivalent production capacity of the Sun Moon Light and Anjiao Packaging and Testing Factory (OSAT) and Intel EMIB-T is converted, the total market has risen from about 65,000 tablets/month at the end of 2025 to about 380,000 pieces/month by the end of 2028. The EMIB-T substrate yield still needs to be improved, but MediaTek's guidance last week shows that improvements are on track, so Bernstein maintains its production capacity forecast.

Profit and valuation: EPS compound growth rate of 35%, with a forward price-earnings ratio of 20 times

According to the revised model, TSMC's revenue in US dollars increased by 44%, 40%, and 27% in 2026, 2027, and 2028 respectively, corresponding to about 176 billion, 247 billion and 315 billion US dollars; considering the drag on N2 and overseas production capacity, gross margin remained roughly the same (66.6% in 2026 and 65.7% in 2027), and earnings per share grew at the same rate as revenue, with a compound growth rate of about 35% from 2026 to 2028. Revised earnings per share of 2026 NT$113.40 (original 110.70), 2027 NT$157.79 (original 144.85); compared with market consensus, earnings per share for the fourth quarter of 2026 were 5% higher, 6% higher in 2027, and 10% higher in 2028.

The target price of NT$3828 is based on 20 times the forward price-earnings ratio multiplied by the earnings per share from the fourth quarter of 2027 to the third quarter of 2028 NT$191; the target ADR price for US stocks of $600 is directly converted at the exchange rate. Looking at the immediate end, the report's revenue forecast for the fourth quarter of 2026 is NT$1.699 trillion, up 62.4% year over year, 0.5% lower than market consensus; earnings per share are 5% higher than the consensus — the revenue side is not dominant; the excess comes from profit margin and expense assumptions.

The location of the valuation should be indicated: TSMC's current forward price-earnings ratio is about 16 times, a seven-year average of 17.5 times, and a positive standard deviation of 21.1 times. The target price used by 20 times is already significantly higher than the historical average and is close to the upper edge of the positive standard deviation. In other words, in 60% of the upward space, valuation repair and profit improvement each account for a portion.