Goldman's 2028 WFE Forecast: The Memory Supercycle Has a Depreciation Cliff
Goldman Sachs just extended its semiconductor equipment cycle forecast to 2028. The headline numbers are aggressive: wafer fab equipment spending climbing from $150 billion in 2026 to $281 billion by 2028, with year-over-year growth of 36%, 45%, then 29%. The market will read this as a green light for equipment makers and memory suppliers alike. I read it as a warning label wrapped in a growth story. The trajectory is real — AI demand is not a mirage — but the shape of that curve tells a story Goldman isn't spelling out. The 2027 peak, the 2028 deceleration, the implied assumption that AI infrastructure investment matures in waves. Data does not negotiate; it only confirms. And the data here confirms something about this cycle that most analysts are missing: the equipment spending surge is front-loaded, and the back half carries risks that aren't in the model.
The forecast rests on three pillars: DRAM process migration from 1-alpha/1-beta to 1-gamma/1-delta nodes, HBM technology iteration from HBM3E to HBM4, and advanced foundry expansion at 3nm and 2nm GAA nodes. Goldman's implicit bet is that AI-driven memory demand isn't a short pulse — it's a structural shift that extends through 2028. That's the right call, but the reasoning needs scrutiny. The core mechanism is HBM's wafer consumption. An HBM3E 8-layer stack consumes three to four times the wafer capacity of standard DDR5. That's the real supply squeeze. It's not that fabs aren't building enough DRAM; it's that HBM production is cannibalizing the wafer supply that would otherwise go to commodity memory. DRAM inventory sits at 4-6 weeks against a normal 8-10 week baseline. That's not a healthy market. That's a market running on fumes.
The storage trio — SK Hynix, Samsung, Micron — is spending over $80 billion combined in 2025, roughly 30-40% of revenue. TSMC is adding another $38-42 billion, about 35-40% of revenue. These are unprecedented capital intensity levels. The equipment delivery cycle — 12-18 months for EUV, 24 months for high-NA — means today's orders don't translate to capacity until 2026-2027. That timing aligns with Goldman's forecast window, but it also means the industry is committing capital now based on demand projections that extend three years out. That's a long leash.
Let me break down what this forecast actually implies at the equipment level. The investment density per wafer is rising. From $150 billion in 2026 to $281 billion in 2028, the implied capex per wafer of capacity is climbing — not flat. That reflects the transition to 2nm GAA nodes and high-NA EUV lithography, where a single tool costs over $300 million. ASML's annual EUV capacity is roughly 50-60 units. If demand outstrips that, delivery lead times stretch from 12-18 months to 24 months or more. The forecast assumes the supply chain can keep pace. That's an assumption worth stress-testing.
The 2027 peak is the tell. Growth hits 45% in 2027, then drops to 29% in 2028. That's not a plateau — that's a signal. Goldman is telling you the first wave of AI infrastructure investment matures around 2027-2028. The second wave depends on applications that don't exist yet: embodied intelligence, large-scale AI agent commercialization. Silence in the ledger speaks louder than hype. The deceleration is the ledger speaking.
From my experience auditing infrastructure projects during the 2017 ICO boom, I learned to read the gap between announced spending and actual deployment. The same discipline applies here. The WFE forecast is an announcement. The actual deployment depends on utilization rates, yield ramps, and the ability of fabs to absorb new equipment without operational disruption. Memory makers are running at 85-95% utilization. That's the kind of tightness that drives equipment purchases, but it's also the kind of tightness that breaks when demand softens.
The HBM competitive landscape adds another layer. SK Hynix holds over 50% of the HBM market, with Samsung at 30% and Micron at 15%. The technology lead is 6-12 months. HBM4 is slated for 2025-2026 mass production, with stack layers moving from 8/12 to 16. Each generation requires new equipment — TSV etching, advanced bonding, thermal management. That's a multi-year equipment upgrade cycle that supports Goldman's forecast. But it also concentrates risk: if HBM4 yield ramps slip, the equipment spending tied to that transition gets delayed, not cancelled.
The financial math is equally telling. Equipment makers are running gross margins of 45-55% — ASML at 50-55%, AMAT and Lam at 45-48%. Those margins are sustainable only if pricing power holds. And pricing power holds only if demand exceeds supply. The WFE forecast implies exactly that, but it also implies something else: that the equipment oligopoly — ASML with 100% of EUV, Lam/TEL/AMAT controlling 80%+ of etch and deposition — can maintain that pricing power through 2028. That's a reasonable bet, but it's a bet on the persistence of scarcity. Scarcity has a way of resolving itself.
Here's what nobody's talking about: the depreciation cliff. Equipment depreciation runs 5-7 years. Memory makers use accelerated 5-year schedules. The massive expansion starting in 2025 means depreciation expenses hit peak concentration in 2027-2029. I've seen this movie before — in 2017, when the ICO boom's infrastructure spending created a similar lag effect. The yield that looks attractive today is not income; it is risk repackaged. Memory makers will need 85%+ utilization just to cover new depreciation. That's a fragile threshold in a cyclical industry. SK Hynix's net profit may hit record levels in 2026, but the depreciation wave that follows will compress margins by 5-10 percentage points. The market will be looking at peak earnings and missing the peak of the cost curve.
The second blind spot is China. Goldman's forecast partially depends on Chinese fab expansion. But export controls mean Chinese fabs are buying domestic equipment — localization is at 20-25% and climbing. If China's domestic equipment makers accelerate, global WFE spending faces structural downward pressure. The forecast doesn't price that in. The audit trail never lies, only the auditor can. And the auditor here is assuming a supply chain that remains intact through 2028. Geopolitical risk — Taiwan Strait tensions, further export control escalation — is the largest unmodeled variable in this forecast. A 10-15% probability of a supply chain shock that invalidates the entire projection is not a tail risk. It's a structural risk.
The 2028 extension is real, but the cycle has a shape. Watch the leading indicators: NVIDIA's guidance, cloud capex commitments, DRAM contract prices. When the 2027 peak arrives, the market will be late to recognize it. Speed without structure is just noise. Structure says: position for the upcycle, but keep one eye on the depreciation cliff and the other on Beijing. The question isn't whether Goldman's forecast is right — it's whether the market is pricing the back half of that curve correctly. It isn't.