Hook
Goldman Sachs is calling for global wafer fabrication equipment spending to hit $281 billion by 2028. That's a 36% compound annual growth rate from current levels — a number so aggressive it reads like a leveraged yield farm's projected APY during a bull run. The last time I saw a forecast this confident, it was a DeFi protocol promising 400% returns on a liquidity pool that drained within six weeks.
Here's what catches my eye: the forecast implies DRAM and HBM memory will consume roughly 40% of that spending. Memory. The most cyclical, boom-and-bust corner of the semiconductor world. Goldman is essentially betting that AI demand will flatten the memory cycle into a growth curve. Code doesn't care about your feelings, and neither does the memory market's historical tendency to punish overconfidence.
Context
The semiconductor equipment industry sits at the top of the manufacturing value chain. Equipment represents 70-80% of a wafer fab's total capital expenditure. ASML holds a 100% monopoly on EUV lithography — the single most critical tool for advanced nodes below 5nm. Applied Materials, Lam Research, and Tokyo Electron dominate etching and deposition. KLA owns over half the metrology market.
These companies are the "shovel sellers" of the AI gold rush. Every GPU, every HBM stack, every advanced node requires their machines. The customer list is brutally concentrated — TSMC, Samsung, Intel, SK Hynix, and Micron account for 50-70% of equipment revenue. But in a seller's market, concentration cuts both ways. When TSMC's N5 and N3 fabs run above 95% utilization and DRAM suppliers are scrambling for capacity, equipment vendors hold the leverage.
The current cycle is unambiguous. AI training chips are sold out. CoWoS advanced packaging capacity is the bottleneck for NVIDIA's entire GPU pipeline. DRAM contract prices have climbed 50-80% over the past year. HBM prices run five to eight times higher than standard DDR5. This is not a demand problem — it's a supply problem, and supply problems are exactly what equipment vendors monetize.
Core
Let me break down what Goldman's numbers actually imply, because the surface-level read misses the structural mechanics.
First, the High-NA EUV assumption. Goldman's 2027 figure of $218 billion and 2028 figure of $281 billion cannot materialize without ASML shipping its EXE:5200 High-NA EUV systems at scale. Each unit costs €300-400 million. ASML's current EUV production capacity is roughly 50-60 units per year. The transition to High-NA — required for 2nm GAA nodes from TSMC, Samsung, and Intel — demands a complete tool refresh. If High-NA delivery slips, the entire WFE forecast slips with it. This is a single-point-of-failure risk hiding inside a diversified-looking number.
Second, the memory tilt. Goldman ranks DRAM/HBM as the primary growth driver. This implies memory capex-to-revenue ratios will hit 40% — far above the historical 25-30% range. For that to work, HBM demand must absorb massive DRAM wafer capacity. HBM3E consumes three to four times the DRAM die area of standard DDR5. HBM4, entering production in 2025-2026, requires hybrid bonding — a technology that demands entirely new equipment precision. SK Hynix, Samsung, and Micron are each committing $15 billion or more to HBM-related expansion. The equipment intensity per unit of memory capacity is structurally higher than logic, which means memory's share of WFE spending will keep climbing.
Third, the delivery bottleneck. Equipment lead times run 12-18 months for EUV and 6-12 months for etch and deposition tools. The equipment vendors themselves need 2-3 years to expand their own manufacturing capacity. You cannot compress a supply chain that depends on Zeiss optics and precision German engineering. The 2026-2028 spending window assumes vendors can physically deliver the tools. My read: actual spending will undershoot the forecast by 10-15% purely on delivery constraints — but the pricing power that scarcity creates will push equipment margins higher, offsetting the volume shortfall.
Fourth, the China variable. Goldman's forecast is built primarily on non-China demand. But China still represents 20-25% of global WFE spending. The Big Fund III — 344 billion yuan — is pouring into domestic equipment makers like Naura, AMEC, and Piotech. These companies are targeting 30-50% annual growth through 2028. If US export controls tighten further — potentially extending to mature-node equipment — China's WFE spending could drop by half. That's a 10-12% haircut on the global number. Conversely, if China's domestic substitution accelerates faster than expected, the competitive landscape for Applied Materials and Tokyo Electron in mature-node segments gets squeezed.
Contrarian
Here's where the consensus narrative breaks down. Everyone is treating this as a pure AI demand story. But the real signal is the cyclicality that Goldman is implicitly dismissing.
WFE spending has always been brutally cyclical. The 2017-2018 cycle peaked and corrected. The 2021-2022 cycle did the same. Three consecutive years of 36-45% growth — which is what Goldman projects — has never happened without a subsequent correction. The 2028 figure of $281 billion with growth decelerating to 29% looks like a top. The question isn't whether the cycle turns; it's whether AI demand genuinely flattens the cycle into a secular growth curve, or whether we're looking at the most crowded trade of 2026.
The second contrarian angle: the equipment vendors themselves are the safest position in this trade, but the market is pricing them like growth stocks when they're still cyclical businesses. ASML trades at 30-35x earnings. KLA at 25-30x. These multiples assume the AI-driven demand persists through 2028 and beyond. If AI capital expenditure hits any speed bump — a major model fails to monetize, cloud providers trim capex guidance, or GPU competition triggers a price war — these multiples compress violently. Panic sells, liquidity buys. The question is whether you're positioned to buy the panic or caught in it.
The third angle: the "sell the shovel" thesis is correct but crowded. Every institutional investor knows equipment vendors benefit from AI expansion. That's not an edge. The edge is in the second-order effects — the materials suppliers, the hybrid bonding equipment makers, the metrology specialists that get overlooked because they don't have the brand recognition of ASML. And the overlooked risk is the delivery bottleneck itself. If equipment can't ship, the spending gets deferred, not cancelled — but deferred spending in a cyclical industry often becomes cancelled spending when the cycle turns.
Takeaway
Goldman's forecast is a map, not a guarantee. The structural logic is sound — AI demand is real, memory tightness is real, advanced node utilization is real. But the forecast embeds assumptions about High-NA EUV delivery, HBM demand persistence, and geopolitical stability that could each break the thesis. The equipment cycle will peak — the only question is whether 2028 marks the top or just another waypoint. Yield is the bait, rug is the hook. In this market, the rug is a cyclical correction wearing an AI costume. Position accordingly, verify everything, and never confuse a forecast with a fact.