Consider a market axiom: cycles mean-revert. What do you call a cycle that extends for four years with a compound growth rate of 36%? That's not a cycle, that's a structural shift. Goldman Sachs made that distinction on August 25, 2025, when it raised its wafer fabrication equipment (WFE) spending forecast to 2028. The trajectory is jarring: $150 billion in 2026, $215 billion in 2027, $281 billion in 2028. Growth rates of 36%, 45%, and 29% respectively. Most analysts read this as bullish. I read it as a stress test.
My first job in this industry was auditing Solidity contracts in 2017, at the height of the ICO bubble. I spent 120 hours manually tracing the Uniswap V1 core, hunting for integer overflows. I found one in the price calculation logic — a silent killer that could have drained the liquidity pools if mainnet launched without the fix. That experience taught me a rule I still apply to every forecast I analyze: smooth curves hide assumptions. Goldman's forecast is a smooth curve. The question is whether the assumptions underneath it are solid or sand.
The market is pricing a 2.5x expansion in semiconductor capital equipment in four years. The consensus calls it the AI supercycle. But supercycles have a bad habit of turning into super crashes. Let me check the math.
The Context: What WFE Actually Measures
WFE is the aggregate capital deployed on the machines that fabricate silicon wafers — lithography, etching, deposition, metrology, cleaning. It is the "sell shovels" layer of the semiconductor industry, and it's a beautiful business: high barriers, high concentration, and high margins. In 2024, global WFE spending was approximately $110 billion. Goldman's forecast projects $281 billion by 2028 — a 2.5x increase in four years.
Three structural drivers underpin the forecast. First, DRAM scaling is moving from 1α/1β nodes to 1γ/1δ nodes — roughly 10-15nm equivalents. Second, HBM (High Bandwidth Memory) is iterating from HBM3E to HBM4, with stack layers rising from 8-12 to 16. Third, advanced foundry is converting from 3nm to 2nm GAA nodes, with EUV double patterning transitioning to high-NA EUV.
The HBM story is the critical piece of the entire forecast. HBM3E consumes three to four times the equivalent wafer capacity of a standard DDR5 module. The three suppliers — SK Hynix, Samsung, and Micron — control the HBM market, with SK Hynix holding over 50% share. Storage capacity utilization is running at 85-95%, and DRAM inventory is at 4-6 weeks — well below the normal 8-10 weeks. HBM is priced at a 3-5x premium over standard DRAM. The supply is that tight.
But here's what Goldman's forecast implicitly assumes: that this tightness persists through 2028. The 2017-2018 storage supercycle lasted two years, and the crash was brutal. The current cycle is different — AI demand is structural — but the question remains: is it a shift or a spike?
The Core: Deconstructing the Capital Math
Let me break down the assumptions, one by one, because that's what I do. I structure my analysis as a system map, not a list.
The first bottleneck is ASML. $281 billion in WFE by 2028 requires roughly 300+ EUV lithography systems installed. ASML's current annual capacity is 50-60 EUV machines. High-NA EUV — the next-gen systems priced above $300 million — has a lead time of 24+ months. The math doesn't close: if ASML cannot scale production, the WFE forecast is a theoretical ceiling, not a practical floor. The equipment supply chain is the most concentrated in the world. ASML has a 100% monopoly on EUV lithography. Lam Research, TEL, and AMAT control 80%+ of etch equipment. Film deposition is locked down by AMAT, TEL, and ASM International.
The second bottleneck is the depreciation curve. Memory makers use 5-year accelerated depreciation; foundries use 7-year straight-line. The 2025-2027 expansion will release depreciation expense in 2027-2029. My analysis shows storage vendors will face a 5-10 percentage point gross margin headwind; foundries 3-5 points. The break-even utilization rates climb to 85%+ for memory and 80%+ for foundries. If AI demand misses those thresholds, you're looking at margin compression that the market is not currently pricing in.
The third bottleneck is the capacity trade-off. HBM's wafer-side consumption is so heavy that the industry is cannibalizing DRAM supply to produce HBM. This is the supply-side logic of DRAM tightness. SK Hynix, Samsung, and Micron are all expanding HBM capacity, but each HBM wafer costs three to four standard DRAM wafers. The supply of standard DRAM is being squeezed — that's why contract prices rose 15-25% in Q2-Q3 2025 and why Goldman expects DRAM tightness to persist through 2028.
Now the fourth layer: the equipment vendors' pricing power. ASML's gross margin is 50-55%, AMAT 45-48%, Lam Research 45-48%, TEL 40-45%. The forecast implicitly assumes these margins hold through 2028. The evidence is on their side — there is no meaningful competition in EUV, and the oligopoly in etch and deposition is stable. But here's the catch: the equipment vendors' pricing power is a function of their own supply chain concentration. The upstream suppliers of high-purity quartz, specialty alloys, and precision components are themselves concentrated. Any disruption in the upstream chain — a fire, a trade restriction, a natural disaster — cascades downstream.
The Chinese localization angle deserves its own note. China's equipment localization rate is 20-25% by value (2024 estimate), with a policy target of 50% by 2030. The reality is that Chinese fabs can substitute locally made tools at mature nodes (28nm and above), but the advanced nodes (7nm and below) still require imported equipment — which is now restricted by US export controls. The forecast depends on China's expansion, but the expansion is mostly mature-node. The US controls, as of 2025, cover advanced logic (14nm and below), advanced DRAM (18nm and below), and 128-layer-plus NAND. China's memory expansion — CXMT and YMTC — is effectively capped at mature nodes.
The final piece of the capital math is the "front-loaded" pattern of the forecast. The growth rates are 36% in 2026, 45% in 2027, then 29% in 2028. The peak is 2027. Goldman is implicitly saying that AI infrastructure capex reaches its first wave of saturation in 2027-2028. After that, the growth rate decays. This is the most interesting signal in the forecast, and most people are ignoring it. A 45% growth rate is a peak, not a plateau. The question is what follows.
The Contrarian Angle: The Blind Spots
The first blind spot is the AI bubble. If AI capex returns on investment — if the revenue from AI applications doesn't justify the capital deployed — cloud providers will cut their infrastructure budgets. The trigger points are NVIDIA's earnings guidance, AWS/Azure/GCP capex announcements, and AI application revenue numbers. The probability of a 30-50% downside in WFE spending over the next 12-24 months is 20-30%. That's not a tail risk; that's a live scenario.
The second blind spot is the storage cycle. DRAM/HBM tightness could ease faster than expected if new capacity comes online quickly, or if HBM yields improve rapidly. The probability of a premature storage cycle peak is 25-35%. If that happens, storage capex peaks a year early and the WFE forecast is off.
The third blind spot is geopolitics. The forecast assumes a contained US-China tech conflict, no Taiwan Strait crisis, and a stable ASML EUV supply. A Taiwan Strait conflict or a full US-China decoupling would collapse the forecast — WFE spending could drop 50%+. The probability is 10-15% over three years, but the impact is catastrophic.
The fourth blind spot — the one I find most interesting — is the depreciation wall. The 2027-2029 depreciation wave is a known unknown. Storage vendors' gross margins will compress by 5-10 points. Current valuations assume margins at 2025-2026 levels, not 2028 levels. When the depreciation hits, earnings expectations will be revised down. The market is not pricing this.
The Takeaway: Watch the 2027 Inflection
Goldman's forecast is a framework, not a prediction. The $281 billion 2028 target is the ceiling of three assumptions: AI demand holds, geopolitics remain contained, and ASML scales. Each assumption is plausible; together they are fragile. My reading is that the cycle peaks in 2027, and the 2028-2029 depreciation wave will compress margins and stock multiples. Watch the forward indicators: NVIDIA's guidance, ASML's order book, SK Hynix's HBM yields, and the DRAM contract price. When those numbers change, the forecast changes. When the forecast changes, the market will follow. Patterns emerge from chaos, not noise. The signal is the 2027 peak.