Goldman Sachs is telling you to buy the shovel. They point to a 36% CAGR in wafer fab equipment spending through 2028. A $281 billion peak. AI demand, HBM shortages, a new golden age for equipment makers. But I have read the fine print. This forecast is not a projection. It is a wish list. It prices in perfection: flawless High-NA EUV ramp-ups, a HBM hunger that never fades, and an AI capex cycle that has never once in history behaved as the model predicts. The code of this market is written in order books, not spreadsheets. Let me follow the trace.
Context: The Forecast and Its Unstated Assumptions
The report is a macro-level teardown of the global semiconductor equipment cycle. Goldman's core thesis is a simple, powerful loop: AI demand drives HBM and advanced node expansion, which in turn forces a historic surge in WFE spending. The numbers are extraordinary. A 36% CAGR from 2025 to 2028. To put that in perspective, the industry has not sustained a compound growth rate like that since the mid-1990s. The forecast bankrolls the entire sector: ASML, Applied Materials, Lam Research, KLA. And the market is already pricing this in, with the sector's PEG ratios at 1.5 to 2.0. But a forecast is a claim. The burden of proof is on the data.
Core: The Contradictions in the Forecast
The first critical juncture is the High-NA EUV dependency. The 2027 and 2028 numbers do not work without the successful mass deployment of ASML's EXE:5200 series. These machines, priced at 300 to 400 million euros per unit, are the backbone of the 2nm and sub-2nm ramp. There is a supply constraint that the forecast simply sidesteps. ASML's annual EUV output is capped at 50 to 60 units. High-NA output is even more limited. The delivery lead time is 12 to 18 months. If you pencil in the production capacity of these machines versus the expected volume of leading-edge wafer starts needed to justify $200 billion-plus of spending, the equations do not balance. The forecast implicitly assumes a flawless ramp-up. There is no evidence of that. In the industry, every single advanced node ramp has been a fight. N3 took longer than TSMC's original schedule. Intel's 18A is also a work in progress. The forecast bakes in a perfect 2026-2027, which is not how the physics works. Smart contracts do not lie; the yield curves do. The second contradiction is the HBM bubble. The report leans heavily on HBM as the primary growth driver. It assumes DRAM supply tightness through 2028. It also states that HBM4 will require hybrid bonding, a technology that is still in early commercial adoption. But look at the footprint. An HBM3E unit consumes three to four times the DRAM die area of a standard DDR5. This is a capacity hog. The forecast for sustained HBM demand implies a demand for AI compute that has no historical precedent. The demand for the end product must remain perfect. I have tracked the money flows through the Terra collapse and the DeFi, but this is a new kind of fragile. The forecast assumes the AI infrastructure buildout will maintain a 40%+ growth rate through 2027. If you trace the actual capital flows from the hyperscalers, you see a boom, but you also see a history of cyclical digestion. The current inventory cycle is already showing signs of the trough. The report treats a supply shortage as a permanent state. Shortages are not permanent. They are an invitation to overbuild.
There is also the matter of supply chain latency. The report's model has an elasticity that does not exist in the physical world. If you have a $218 billion WFE spending year in 2027, the equipment makers themselves must have scaled their manufacturing capacity. Applied Materials and Lam Research are not idle. But their lead times for the advanced deposition and etch tools remain high. The report's forecast implies a delivery bottleneck that will inevitably push out the spending curve. This is a supply chain that is already at the limit. In a 2027 environment with a 45% year-on-year increase, the system will not be able to move that product. The equipment will be delivered late, pushing the revenue recognition to 2029, breaking the model's internal logic.
Contrarian: What the Bulls Get Right
I have to give credit where it is due. The report does not present a load of empty narratives. The core demand signal is real. AI training and inference demand is not a synthetic trend. The HPC/AI application segment is growing at 40-50% CAGR. This is the biggest single secular shift in the industry since the personal computer. The reasoning is simple. This AI infrastructure requires the leading edge, and the leading edge requires the best equipment. The high capital intensity is a moat. The industry is a supplier's market. The customer has no leverage. The price inelasticity of the EUV machine is a fact. And the financial metrics of the equipment makers are robust. KLA's gross margin is over 60%. ASML is at 50% plus. The demand for the technology is real. The value of the shovel is not in question. The question is whether the forecast's pace is the actual pace. The forecast is a direction, not a map. The industry will expand. But it will not expand at a constant, unbroken, linear rate. The floor is a mirror reflecting greed, not value.
Takeaway: The Signal in the Noise
Hype burns out, but the ledger remains cold. Goldman's forecast is not a fraud. It is a benchmark. It is a description of a world that works perfectly. But the market is not perfect. The capacity constraints, the geopolitical fragmentation, and the AI capex cycle are variables. If you are a founder or an investor in this space, do not buy the full 36% CAGR. Watch the yield of the High-NA machines. Watch the quarterly capex guidance from the three big cloud providers. If the AI spending dips, the WFE forecast will look like a discount. In the blockchain, truth is coded, not claimed. In the fabs, truth is booked, not forecasted. The first data points will come when the numbers for the year are reported. The contract is in the purchase orders. The bond is the delivery. Follow the lead. Trust the data. The cycle will turn. It always does.