The 2028 projection of $281 billion in wafer fabrication equipment spending—up from $150 billion in 2026 and $218 billion in 2027—represents a 37% compound annual growth rate that deserves forensic scrutiny rather than celebratory coverage. Goldman Sachs' revised forecast embeds three optimistic assumptions: sustained AI capital expenditure through 2028, rationalized export controls, and supply chain delivery capacity that has historically failed to meet comparable targets. As someone who has spent a decade auditing technology claims against on-chain and operational realities, I approach this forecast with the same skepticism I apply to crypto protocols promising 20% yields.
The forecast arrives at a peculiar inflection point. Global foundries operate at 80-85% utilization, with advanced nodes (5nm and below) effectively saturated at 95%. The 2024 announcement cycle already committed over $300 billion in cumulative capex across TSMC, Samsung, SK Hynix, and Micron. The equipment supply chain—ASML, Applied Materials, Lam Research, Tokyo Electron, KLA—maintains gross margins between 45-61%, reflecting an oligopoly that extracts pricing power from every wafer start. The question is not whether the demand exists. It does. The question is whether the forecast's assumptions survive contact with geopolitical and operational reality.
The equipment market is bifurcating into two distinct growth engines. The logic path runs through advanced foundry expansion—TSMC's Arizona fab complex ($65 billion, three fabs), Samsung's Taylor facility ($37 billion), and the 2nm GAA transition. The memory path runs through HBM and DRAM, with SK Hynix's $90 billion Yongin cluster and Micron's $100 billion-plus New York/Idaho DRAM builds. The HBM trajectory is particularly revealing: from HBM3E at 12 layers to HBM4 at 16 layers, each stacking generation demands TSV etching, electroplating, and bonding equipment that does not overlap with logic fab tooling. This bifurcation creates what the Goldman forecast implicitly recognizes: a two-engine growth model where memory equipment spending becomes a parallel, not derivative, driver of WFE expansion.
But the forecast's most fragile assumption is export control rationalization. The 2028 WFE figure requires China to sustain $40-50 billion in annual equipment procurement. Yet the current trajectory moves in the opposite direction. The December 2024 HBM export controls, Japan's July 2023 restrictions covering 23 equipment categories, and the Netherlands' licensing requirements for immersion DUV tools point toward fragmentation, not rationalization. My estimate is that export controls introduce a 10-15% downward variance to Goldman's forecast if current trajectories hold. The China market—roughly 30% of global equipment demand—cannot be replaced by U.S. or European fab expansion within this timeframe.
The geopolitical overlay introduces a 25% probability of a pessimistic scenario. A comprehensive decoupling scenario—including mature-node equipment—would reduce global WFE by more than 20% as the industry splits into redundant regional clusters. The efficiency loss, estimated at 10-20% from duplicated investments, is a tax on the entire value chain. What the forecast does not price: the Chinese equipment domestic substitution program. The National Fund Phase III, with $48 billion in allocated capital, targets exactly the 28nm etching and deposition segments where domestic champions like Naura and AMEC already demonstrate production viability. If domestic substitution reaches 40% by 2028—a plausible trajectory given the extended fab construction timelines—China's import demand contracts beyond what export controls alone would achieve.
The equipment cycle's structural extension is the underappreciated bull case. The conventional semiconductor cycle lasts 2-3 years. The AI-driven cycle, by contrast, carries characteristics of a structural repricing. Cloud hyperscalers—Microsoft, Google, Amazon, Meta—have committed to $300 billion-plus combined capex for 2025. These are not cyclical commitments; they are strategic positions for AI workloads that have no mature alternative. The HBM effect is equally structural. Traditional DRAM was a commodity with violent cycles. HBM is tied directly to AI accelerator demand, with price per bit at 5-8x conventional DRAM. This shifts memory revenue from cyclical to semi-structural, potentially compressing the amplitude of the historical equipment cycle.
The counter-argument deserves scrutiny. The equipment industry's average gross margins of 45-60% reflect a seller's market that presupposes the supply-demand gap persists. My assessment of delivery capacity suggests the constraint is real: ASML EUV lead times run 12-18 months, high-NA EUV extends to 24 months, and advanced fab qualification cycles span 2-3 years. But equipment makers are not capacity constrained in the way they were in 2020-2021. The question is whether their delivery capability matches the forecast's implied volume—approximately 80-100 EUV units per year, up from roughly 50 in 2024. ASML's production expansion plans support this, but the company's own guidance reflects a more conservative trajectory.
The risk that the forecast is overly optimistic is not symmetrical with the risk it is overly pessimistic. The industry is at the beginning of a multi-year expansion cycle. But the forecast's CAGR of 37% requires not just AI demand persistence, but also the absence of a major geopolitical shock, an export control reversal, and the successful execution of an unprecedented capacity expansion. The probability that all three conditions hold through 2028 is low. My assessment places the realistic WFE range at $220-240 billion for 2028—10-15% below the forecast. The direction is correct; the magnitude is aggressive.
The implication for investors and supply chain participants is clear: the equipment sector will outperform, but the multiples—currently 30-35x PE for the majors—already price in the optimistic scenario. The risk-reward has shifted from asymmetric upside to asymmetric downside. For those tracking this cycle, the metrics to monitor are not the headline WFE numbers but the leading indicators: DRAM contract price movements, TSMC CoWoS capacity utilization, ASML EUV order book duration, and the pace of Chinese equipment qualification. Those will tell you when the forecast's assumptions are breaking—before the guidance revision makes it public.
The forecast is a directional signal with an embedded magnitude error. The equipment super-cycle is real. The $281 billion number is less certain. The market will eventually reconcile the two; the question is whether investors will be positioned on the right side of that reconciliation.