Bernstein flagged CPUs as a fresh growth driver for Taiwan Semiconductor Manufacturing (TSM.US) three months ago, and its latest report raises that call another notch: the real shift is not in GPUs or ASIC XPUs, but in the fact that custom CPUs developed in-house by Amazon and Google are beginning to scale up in deployment, driving demand beyond prior expectations.

The incremental volume comes from "ASIC CPUs," not AI accelerators. The CPUs custom-built by Amazon and Google for their own use are deployed only internally and are not sold externally, making them ASICs in nature; Bernstein calls them "ASIC CPUs" to distinguish them from ASIC XPUs aimed at AI training and inference. Both categories are seeing rising demand, but the recent surprise is that agentic workloads are accelerating CPU demand, especially for the in-house CPUs the two cloud providers are now deploying.

The report highlights another structural difference: on ASIC CPUs, Amazon and Google are far more self-sufficient than on XPUs, relying much less on ASIC design service providers. Amazon's Graviton5 is already its fifth-generation in-house CPU, built on a 3-nanometer process with 192 Arm Neoverse V3 cores and 172 billion transistors, with everything from design to production arrangements with suppliers held in-house. Google had prior experience with two generations of CPU designs (Maple and the first version of Axion), and the second-generation Axion was designed in-house with only production arrangements handed to Global Unichip (GUC); that chip is benefiting from agentic workloads and is ramping rapidly from late 2026 into 2027.

The key point is that this is additive rather than substitutive: these in-house CPUs do not displace general-purpose CPUs from Intel and AMD, but add demand on top of them, and ultimately all of it must land on TSMC wafers 鈥?Graviton5 uses 3-nanometer, while the second-generation Axion lands in the N4 family.

CPU Revenue: From $24 Billion to $50 Billion, Share Rising from 7% to 16%

Bernstein projects TSMC's CPU revenue rising from about $5 billion in 2023, roughly $10 billion in 2024, about $14 billion in 2025, and around $24 billion in 2026, extrapolating all the way to about $40 billion in 2027 and roughly $50 billion in 2028; as a share of total revenue, that rises from 7% to 16% accordingly. To clarify the methodology, this estimate excludes Apple's in-house CPU used in PCs (for example, the A18 Pro shared by the iPhone 16 Pro/Pro Max and the MacBook Neo), which is also the reason for the small differences versus the historical data in the previous version of the report.

On customer structure, AMD remains the largest single CPU customer from 2026 to 2028, holding a 50% to 60% share; Intel's share falls from above 30% in 2025 to about 10% in 2028; Google and Amazon were still grouped under "others" in 2025, after which their contribution rises markedly.

Node Divergence: N3 and N2 Revised Up, N5 and N7 Under Pressure

Demand is not improving across the board. PCs and low-end smartphones are weaker than expected, with the low end especially weak, softening demand slightly for N5 and N7; N3 and N2, by contrast, are shared by XPUs, CPUs, and high-end smartphone SoCs, making demand firmer, and Bernstein accordingly raised their revenue contribution. A14 is set to succeed N2 on schedule, contributing meaningful revenue starting in the second half of 2028, with the cadence from N2 to A14 only slightly longer than two years. The first adopting customers are Apple, Qualcomm, MediaTek, and AMD; Nvidia will also adopt A14, but not until 2029 鈥?consistent with its failure that year to migrate to N3 as early as other XPU suppliers, and thanks to its advantages in CUDA and system-level integration, this pace has not affected its leadership in the AI XPU market. Overall, A14 will still deliver more revenue dollars to TSMC at the equivalent stage than N2 and all prior nodes.

Capex: $77 Billion in 2027, but Capex Intensity Still Declining

Reflecting the upward demand revision, Bernstein raised its capex forecast from $63 billion in 2026 to $77 billion in 2027 and $83 billion in 2028. At the same time, it maintained its prior view that capex as a percentage of revenue will gradually decline, consistent with TSMC's comments in its most recent earnings report. Advanced packaging capacity was also revised slightly higher. TSMC's own CoWoS-equivalent capacity rises from about 60,000 wafers per month at the end of 2025 to about 135,000 in 2026, about 175,000 in 2027, and about 180,000 in 2028; including converted equivalent capacity from outsourced assembly and test providers (OSAT) such as ASE and Amkor as well as Intel's EMIB-T, total market capacity climbs from about 65,000 wafers per month at the end of 2025 to about 380,000 per month by the end of 2028. EMIB-T substrate yields still need improvement, but guidance from MediaTek last week indicates improvement is on track, so Bernstein maintained its capacity forecast.

Earnings and Valuation: 35% EPS CAGR Paired with a 20x Forward P/E

Under the revised model, TSMC's US dollar revenue grows 44%, 40%, and 27% in 2026, 2027, and 2028 respectively, corresponding to roughly $176 billion, $247 billion, and $315 billion; factoring in the drag from N2 and overseas capacity, gross margin is roughly flat (66.6% in 2026, 65.7% in 2027), with earnings per share growing at the same pace as revenue, a compound annual growth rate of about 35% from 2026 to 2028. Revised 2026 EPS is NT$113.40 (previously NT$110.70) and 2027 EPS is NT$157.79 (previously NT$144.85); relative to market consensus, fourth-quarter 2026 EPS is 5% higher, 2027 is 6% higher, and 2028 is 10% higher. The target price of NT$3,828 is based on a 20x forward P/E multiplied by fourth-quarter 2027 to third-quarter 2028 EPS of NT$191; the US ADR target price of $600 is a direct conversion at the exchange rate.

On the near term, the report forecasts fourth-quarter 2026 revenue of NT$1.699 trillion, up 62.4% year over year, 0.5% below market consensus; EPS, however, is 5% above consensus 鈥?revenue is not the advantage, and the outperformance comes from margin and expense assumptions. It is worth noting where valuation stands: TSMC's current forward P/E is about 16x, versus a seven-year average of 17.5x and a plus-one-standard-deviation level of 21.1x. The 20x used in the target price is already well above the historical average and approaching the upper edge of plus one standard deviation, meaning that of the 60% upside, valuation re-rating and earnings revisions each account for a portion.