Bernstein’s interpretation: Revaluing computing-power equipment stocks at 50GW—has the AI hardware supercycle arrived?

TL;DR

• Bernstein estimates that in a 50GW scenario, total related WFE over 2027-2029 could be about $736 billion.
• For every additional 1GW/year of AI compute capacity, about 46K-50K WSPM of wafer-start capacity is needed; DRAM and HBM account for the largest shares.
• Applied Materials, Lam Research, and KLA have higher earnings upside, but pipeline capacity is not the same as actual orders.

Bernstein’s latest report translates AI data center expansion into semiconductor manufacturing equipment order demand: if by 2030 AI data centers add 50GW of computing capacity every year, global related WFE spending in 2027-2029 could total about $736 billion, with 2029 alone at about $291 billion.

WFE refers to wafer-fab equipment spending. It is a key source of demand for equipment companies such as Applied Materials (AMAT), Lam Research Group (LRCX), KLA (KLAC), ASML, Tokyo Electron, and others. For investors, the most direct question in this estimate is: as AI compute capacity continues to expand, how much incremental orders and earnings leverage can it bring to equipment vendors?

At the time the report was released, semiconductor equipment stocks had already gone through a round of strong rallies, and more recently had pulled back clearly from their highs. At the same time, the U.S. data center construction pipeline is still expanding. Public excerpts show that as of June 2026, project pipeline capacity rose to 338GW, adding 217GW over the past 12 months—far higher than the current scale already in operation.

U.S. data center active capacity and project pipeline changes, with pipeline capacity rising from 121GW to 338GW.

About 50K wafers/month of wafer-start capacity per additional 1GW of compute

The core conversion in this report is: for every additional 1GW/year of AI computing capacity, about 46K-50K of new WSPM wafer-start capacity is needed. WSPM refers to the number of wafer starts per month, a commonly used metric for measuring fab capacity.

Data center capacity itself does not directly turn into equipment orders. Only when AI servers need more GPUs, HBM, DRAM, NAND, and advanced logic chips will fabs need to expand production, and equipment companies will then see increased WFE spending.

Within the incremental demand of about 46K WSPM/GW, DRAM has the highest share at about 53%; NAND about 20%; HBM about 16%; advanced logic about 11%. In other words, expanding AI data centers does not only drive demand for advanced-process GPUs—it also increases the need for storage production capacity, especially DRAM and HBM.

This is also why Applied Materials is the most watched name in these estimates. Incremental wafer demand mainly comes from DRAM, HBM, and NAND; Applied Materials has broader exposure across storage-related equipment, including deposition, etching, and other steps, which makes earnings leverage more direct.

For every additional 1GW of compute capacity, about 46K WSPM is needed—DRAM 53%, NAND 20%, HBM 16%, Logic 11%.

2029 annual WFE could approach $291 billion

In the baseline scenario, by 2030 compared with the 2026 baseline, AI data centers achieve an additional 50GW of computing capacity per year. To support this target, related WFE needs to be gradually lined up during 2027-2029.

Scenario calculations show that WFE spending driven solely by AI totals about $376 billion over 2027-2029; if you add a non-AI baseline spending of about $120 billion per year, the total WFE spending over three years would be about $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 the equipment spending assumptions in the currently more conservative model. If the 50GW scenario is realized, the WFE scale in 2029 would be close to $300 billion; under higher GW scenarios, there is room for equipment spending to move further upward.

The report also provides a more aggressive scenario. In a 75GW scenario, the potential upward revision to equipment companies’ earnings could exceed 100%. In a 100GW scenario, potential WFE spending in 2029 is further amplified, and for some companies valuation multiples could be pushed below 10x.

But these are still model estimates, not orders that have already materialized. They depend on whether the data center construction pipeline can be converted into actual deployments, whether shipments of AI servers can keep up, whether fabs are willing to expand capacity ahead of time, and whether the equipment supply chain has sufficient delivery capacity.

In the 50GW scenario, total WFE over 2027-2029 is about $736 billion, with 2029 at about $291 billion; in the 100GW scenario, 2029 is about $542 billion.

Applied Materials has the most upside; Lam Research and KLA also benefit

The stock impact is mainly concentrated in Applied Materials, Lam Research Group, and KLA.

In the 50GW scenario, the three companies’ EPS for 2029 could rise by about 37%-60% versus current Wall Street consensus. This would imply forward P/E potentially dropping into the 15-20x range, while today’s equipment stocks are still generally trading at higher multiples.

Among them, Applied Materials has the biggest upside. In the 50GW scenario, its 2029 EPS would be up close to 60% versus consensus, implying a forward P/E of about 14.9x. Lam Research’s EPS would be up about 54%, implying about 18.4x. KLA’s EPS would be up about 37%, implying about 21.2x.

The reason still lies in the composition of wafer demand. The incremental capacity brought by AI expansion is mainly concentrated in DRAM, HBM, and NAND—not a single advanced logic category. Applied Materials has broader coverage in storage-related equipment, making it more likely to capture incremental demand than companies that benefit only from selected steps.

According to the report’s wording, Bernstein maintains “outperform the market” ratings on multiple equipment stocks including Applied Materials, Lam Research, KLA, ASML, and Tokyo Electron, with Applied Materials still the top pick. Screen has a neutral rating.

In the 50GW scenario, AMAT/LRCX/KLAC 2029 EPS is up 59.5%/54.4%/37.1% versus consensus, and P/FE falls to 14.9x/18.4x/21.2x.

Pipeline capacity still has to clear power, financing, and delivery hurdles

The easiest place to misread these estimates is to treat data center pipeline capacity as future equipment orders directly.

U.S. data center pipeline capacity has expanded significantly over the past year, indicating that willingness to invest in AI infrastructure remains strong. But from pipeline to actual operations, there are multiple hurdles: power interconnection, land approvals, financing costs, GPU supply, customer demand, as well as network and cooling infrastructure—all of which affect the speed at which projects ultimately take off.

There are also constraints on the equipment side. If the WFE scale needs to ramp toward the $300 billion level within a few years, equipment companies, component suppliers, and fabs all need to expand capacity in sync. Semiconductor equipment is not an industry that can scale up infinitely and rapidly; advanced tools, key components, installation and commissioning, and customer qualification can all extend delivery cycles.

The model assumptions themselves also have boundaries. The related estimates are based on specific assumptions for GPU architectures, power consumption, chip area, and capital intensity, and assume that WFE will be in place by the end of 2029 to support the incremental compute capacity added in 2030. Changes in architecture, lower energy consumption per unit of compute, improved chip yields, and adjustments to capital intensity could all change the final equipment demand.

There can also be bias in another direction. If the need to replace legacy compute before 2030 is not adequately captured, equipment demand could still have upside. But if AI applications monetize more slowly than expected, or large cloud providers slow down capital expenditures, the 50GW, 75GW, and even 100GW scenarios could be overly optimistic.

This report does not provide a guaranteed orders list; rather, it offers a clearer set of conversions: for each additional 1GW of AI data center capacity, it may correspond to about 50K wafers/month of wafer-start capacity and about $8 billion worth of incremental WFE demand. Equipment stocks have already priced in part of the AI outlook in advance; the key disagreement is whether data center construction can be executed strongly enough to support annual WFE on the order of $300 billion.