Goldman Sachs just extended its semiconductor equipment (WFE) cycle forecast through 2028, projecting spending to climb from $150 billion in 2026 to $281 billion by 2028 โ a trajectory implying 36%, 45%, and 29% year-over-year expansion across three consecutive years. On the surface, this reads as an unambiguous bullish signal for every chip-adjacent narrative, including the AI-compute tokens that have been riding the infrastructure wave. But let me run the pre-mortem before we uncork the champagne. Every equipment cycle I've audited over the past decade โ from the 2017 DRAM supercycle to the 2021 GPU shortage โ carried the same structural flaw: it assumed demand was linear while supply response was elastic. The WFE forecast is no different. The real question isn't whether spending grows. It's which link in the chain breaks first.
WFE spending represents the capital deployed by chipmakers on fabrication equipment โ lithography, etching, deposition, metrology โ the physical machinery that converts silicon wafers into compute. Goldman's revised forecast rests on three expansion vectors: DRAM process scaling from 1ฮฑ/1ฮฒ toward 1ฮณ/1ฮด nodes, HBM technology iteration from HBM3E to HBM4 (with stack layers moving from 8/12 to 16), and advanced foundry transitions from 3nm to 2nm GAA architectures requiring high-NA EUV lithography. The capex commitments behind these vectors are staggering. SK Hynix, Samsung, and Micron are collectively earmarking over $80 billion in 2025 capital expenditure, with HBM capacity expansion as the primary driver. TSMC is spending $38โ42 billion โ roughly 35โ40% of revenue โ on advanced node capacity.
The equipment supply chain remains extraordinarily concentrated. ASML holds a 100% monopoly on EUV lithography, while Lam Research, Tokyo Electron, and Applied Materials control over 80% of the etching and deposition market. China's equipment localization rate sits at roughly 20โ25% by value, targeting 50% by 2030 under policy mandates. This is the macro-institutional frame that matters. We're not just observing a semiconductor cycle; we're witnessing the physical infrastructure layer for the entire AI narrative โ including the crypto projects that have staked their token economics on decentralized compute networks.
Insight One: The HBM Wafer Multiplier is the Real Story.
The conventional reading of this cycle is that AI drives GPU demand, which drives foundry demand, which drives equipment spending. True, but incomplete. The hidden engine is HBM's wafer consumption profile. An HBM3E 8-layer stack consumes three to four times the wafer capacity of standard DDR5. This is the mechanical reason DRAM supply stays tight: every gigabyte of HBM allocated represents four gigabytes of conventional DRAM not produced. Goldman's forecast implicitly accepts that this cannibalization persists through 2028. Based on my experience modeling the 2021 memory cycle, this is the most defensible assumption in the entire report โ the physics of wafer allocation don't lie, even when market narratives do.
Insight Two: ASML is the Binding Constraint.
The equipment supply chain itself is the bottleneck. ASML produces roughly 50โ60 EUV units annually, with high-NA EUV priced above $300 million per unit and delivery cycles extending beyond 24 months. The forecast's growth trajectory assumes ASML can scale output to meet demand โ an assumption that has historically proven optimistic. When I audited equipment supply chains during the 2021 chip shortage, the pattern was identical: the monopoly supplier's capacity became the de facto cap on industry expansion. The regulatory moat here is absolute โ no competitor can replicate EUV technology within a decade, and export controls only reinforce ASML's position by limiting alternative market access.
Insight Three: The Growth Curve Shape is the Tell.
Look at the year-over-year trajectory: 36% in 2026, 45% in 2027, then a drop to 29% in 2028. That's a peak-and-decline pattern. Goldman is signaling that AI infrastructure investment hits its first saturation point around 2028. This matters enormously for anyone building on the AI narrative โ including Web3 compute projects whose token valuations are priced on perpetual infrastructure expansion. The equipment makers' gross margins โ ASML at 50โ55%, AMAT and Lam at 45โ48% โ will hold through 2027, but the deceleration signal is already embedded in the numbers.
Here's the counterintuitive angle. The equipment forecast assumes AI demand is structural, not cyclical. But the equipment makers are also the ones selling that narrative. ASML, AMAT, and Lam have every incentive to maintain the optics of a multi-year supercycle โ their stock prices depend on it. The 2017โ2018 DRAM cycle lasted roughly two years before supply caught up and prices collapsed. What's different this time? The answer is supposed to be "AI is structural." But structural demand has a way of becoming cyclical when capital expenditure overshoots. DRAM inventory sits at 4โ6 weeks, well below the normal 8โ10 weeks โ but that's precisely the setup that preceded the 2018 crash.
There's also the geopolitical blind spot. The forecast implicitly assumes US-China tech decoupling remains contained. But China's Big Fund Phase 3 โ 344 billion yuan deployed toward equipment localization โ is a structural headwind to the Western equipment oligopoly. If Chinese localization accelerates, a meaningful portion of projected WFE growth shifts away from ASML and AMAT toward domestic suppliers like Naura and AMEC. The equipment forecast also assumes no extreme geopolitical scenario โ a Taiwan Strait contingency would invalidate every number in the model. I'd assign that a 10โ15% probability over three years, which is not negligible.
Hunting for the story that defines the next cycle: the semiconductor equipment forecast is the physical layer of the AI narrative. When Goldman's WFE growth curve peaks in 2027 and decelerates in 2028, the AI-compute token narrative will face its first real stress test. Watch the ASML order book, watch DRAM contract prices, and watch whether HBM4 ramps on schedule. The silicon bottleneck will tell you when the AI story is nearing its first inflection โ before the token market does. The narrative always decouples from reality first; the equipment data is where reality lives.
