伯恩斯坦解讀：50GW算力重估設備股，AI設備超級週期來了？

TL;DR

• Bernstein estimates that under the 50GW scenario, cumulative related WFE from 2027-2029 will be approximately $736 billion.
• Each additional 1GW/year of AI computing power requires roughly 46K-50K WSPM of wafer capacity, with DRAM and HBM accounting for the largest share.
• Applied Materials, Lam Research, and KLA have higher earnings elasticity, but pipeline capacity does not equal firm orders.

A new report from Bernstein translates AI data center expansion into semiconductor manufacturing equipment orders: if AI data centers add 50GW of computing power annually by 2030, global related WFE spending from 2027-2029 could cumulatively reach approximately $736 billion, with around $291 billion in 2029 alone.

WFE refers to wafer fab equipment spending, a key demand driver for equipment companies like Applied Materials (AMAT), Lam Research (LRCX), KLA (KLAC), ASML (ASML), and Tokyo Electron. For investors, the most direct question from this calculation is: as AI computing power continues to expand, how much incremental orders and earnings elasticity can equipment makers gain?

At the time of the report's release, semiconductor equipment stocks had already experienced a significant rally, followed by a notable pullback from highs. Meanwhile, the US data center construction pipeline continues to grow. According to publicly available data, as of June 2026, the project pipeline capacity rose to 338GW, an increase of 217GW over the past 12 months, far exceeding current operational capacity.

Changes in US data center active capacity and project pipeline; pipeline capacity increased from 121GW to 338GW.

Every Additional 1GW of Computing Power Requires Approximately 50,000 Wafers Per Month of Capacity

The core calculation of this report is that each additional 1GW/year of AI computing power requires roughly 46K-50K WSPM of new wafer capacity. WSPM stands for wafer starts per month, a common metric for measuring fab capacity.

Data center capacity itself does not directly translate into equipment orders. Only when AI servers require more GPUs, HBM, DRAM, NAND, and advanced logic chips do fabs need to expand capacity, leading to incremental WFE spending for equipment companies.

Within the approximately 46K WSPM/GW of new demand, DRAM accounts for the largest share at about 53%; NAND approximately 20%; HBM around 16%; and advanced logic about 11%. This means AI data center expansion does not solely drive demand for advanced GPU processes but also significantly boosts memory capacity needs, especially DRAM and HBM.

This is why Applied Materials is most in focus under this calculation. Incremental wafer demand mainly comes from DRAM, HBM, and NAND, and Applied Materials has higher exposure to memory equipment segments like deposition and etching, making its earnings elasticity more direct.

Each additional 1GW of computing power requires approximately 46K WSPM: DRAM 53%, NAND 20%, HBM 16%, Logic 11%.

Annual WFE in 2029 Could Approach $291 Billion

In the base case scenario, AI data centers add 50GW of computing power annually by 2030 relative to the 2026 baseline. To support this target, related WFE needs to be deployed progressively from 2027-2029.

Scenario calculations show that AI-driven WFE spending alone would cumulatively reach approximately $376 billion from 2027-2029; including an annual non-AI baseline spending of about $120 billion, total three-year WFE spending would be around $736 billion. The annual path is roughly $200 billion in 2027, $245 billion in 2028, and $291 billion in 2029.

These figures are higher than equipment spending assumptions in more conservative current models. If the 50GW scenario materializes, the scale of WFE in 2029 will approach $300 billion; under higher GW scenarios, equipment spending has further upside potential.

The report also presents more aggressive scenarios. In the 75GW scenario, earnings upgrades for equipment companies could exceed 100%; in the 100GW scenario, WFE spending potential in 2029 expands further, and valuation multiples for some companies could compress below 10x.

However, these remain model calculations, not confirmed orders. They depend on whether the data center construction pipeline translates into actual operations, whether AI server shipments keep pace, whether fabs are willing to expand capacity preemptively, and whether the equipment supply chain has sufficient delivery capability.

Under the 50GW scenario, cumulative WFE from 2027-2029 is about $736 billion; 2029 WFE is about $291 billion. Under the 100GW scenario, 2029 WFE is about $542 billion.

Applied Materials Has the Highest Elasticity; Lam Research and KLA Also Benefit

The stock impact is primarily concentrated on Applied Materials, Lam Research, and KLA.

Under the 50GW scenario, the 2029 EPS for the three companies could rise 37%-60% above current Wall Street consensus. Corresponding forward P/E ratios could fall to the 15-20x range, while current equipment stocks are generally trading in higher ranges.

Among them, Applied Materials shows the greatest elasticity. Under the 50GW scenario, its 2029 EPS is nearly 60% above consensus, corresponding to a forward P/E of approximately 14.9x; Lam Research's EPS is about 54% above consensus, corresponding to about 18.4x; KLA's EPS is about 37% above consensus, corresponding to about 21.2x.

The reason lies in the composition of wafer demand. New capacity brought by AI expansion is mainly concentrated in DRAM, HBM, and NAND, not solely in advanced logic. Applied Materials has broader coverage in memory-related equipment, making it easier to capture the incremental demand than companies benefiting only from certain segments.

According to the report, Bernstein maintains outperform ratings on multiple equipment stocks including Applied Materials, Lam Research, KLA, ASML, and Tokyo Electron, with Applied Materials remaining the top pick. Screen is rated neutral.

Under the 50GW scenario, AMAT/LRCX/KLAC 2029 EPS is 59.5%/54.4%/37.1% above consensus; P/FE declines to 14.9x/18.4x/21.2x.

Pipeline Capacity Still Needs to Overcome Hurdles of Power, Financing, and Delivery

The most common misinterpretation of this calculation is to treat data center pipeline capacity directly as future equipment orders.

The significant expansion of the US data center pipeline over the past year indicates strong willingness for AI infrastructure investment. However, there are multiple hurdles between the pipeline and actual operation: power availability, land permitting, financing costs, GPU supply, customer demand, and network and cooling infrastructure all affect the eventual pace of deployment.

There are also constraints on the equipment side. If WFE is to surge toward the $300 billion level within a few years, equipment companies, component suppliers, and fabs all need to expand capacity in tandem. The semiconductor equipment industry is not one that can rapidly ramp up volume indefinitely; advanced equipment, critical components, installation and commissioning, and customer qualification all extend delivery lead times.

The model assumptions themselves have boundaries. The calculations are based on specific GPU architectures, power consumption, chip areas, and capital intensity assumptions, and assume that WFE must be in place by the end of 2029 to support incremental computing power in 2030. Changes in architecture, reductions in power consumption per unit of computing power, improvements in chip yield, and adjustments in capital intensity could all alter final equipment demand.

Deviation in the other direction also exists. If the demand for replacing older computing capacity before 2030 is not fully accounted for, equipment demand could have further upside; however, if the monetization of AI applications slows, or if large cloud vendors scale back capital expenditure, the 50GW, 75GW, or even 100GW scenarios might prove overly optimistic.

This report does not provide a definitive order book but rather a clearer conversion framework: every additional 1GW of AI data center computing power likely corresponds to approximately 50,000 wafers per month of capacity and around $8 billion in incremental WFE demand. Equipment stocks have already priced in some of the AI expectations; the divergence lies in whether data center construction can materialize to the extent that supports an annual WFE level of $300 billion.