The data shows a 36% year-over-year increase in wafer fab equipment spending for 2026, followed by 45% in 2027, before settling at 29% in 2028. Goldman Sachs did not publish this forecast to be polite. They published it because the order books for ASML, Applied Materials, and Lam Research are filling up faster than the fabs can pour concrete. The revised cycle extension to 2028 is not a narrative. It is a supply chain reality check.
Contrary to the hype that AI is a software story, the hard capital is flowing into lithography systems, deposition tools, and etch chambers. The WFE forecast, climbing from $150 billion in 2026 to $281 billion by 2028, tells me one thing: the industry is betting that AI compute demand is not a fad. It is a structural shift in how we allocate capital to silicon.
I have spent the last decade auditing smart contracts and yield strategies, but the same forensic principles apply here. The code does not lie, only the audits do. In semiconductors, the code is the process node, and the audit is the capital expenditure plan. Goldman's revised forecast is an audit of the AI supply chain, and it is passing with high marks.
Context: The Memory Supercycle and the Advanced Logic Squeeze
To understand why Goldman raised its forecast, you have to look at the memory market. DRAM contract prices rose 15-25% quarter-over-quarter in Q2 and Q3 of 2025. HBM, the high-bandwidth memory that sits next to NVIDIA's GPUs, is commanding a 3-5x premium over standard DDR5. The three memory giants—SK Hynix, Samsung, and Micron—are running at 85-95% capacity utilization. That is not a healthy operating level. That is a screaming signal for more capacity.
SK Hynix is spending tens of billions to double HBM output. Samsung is expanding DRAM and HBM lines. Micron is doing the same. On the logic side, TSMC is pouring hundreds of billions into 2nm and 3nm capacity, with CoWoS advanced packaging still sold out. The utilization rates tell the story: advanced logic is at 75-85%, mature logic is lower, but memory is the tightest it has been since the 2017 supercycle.
This is not 2017, though. The 2017 cycle was driven by smartphone replacement and cloud buildouts. This cycle is driven by AI training clusters that consume memory like a furnace consumes oxygen. HBM3E 8-layer stacking consumes 3-4x the wafer capacity of standard DDR5. That is the hidden squeeze. The memory makers are not just adding capacity; they are converting existing DRAM capacity to HBM, which reduces the supply of commodity DRAM. This is why Goldman sees DRAM tightness persisting into 2028.
Core: The Order Flow Analysis of WFE Spending
The WFE spending trajectory is not a smooth line. It is a series of discrete, high-value orders that reveal the underlying strategy of the major fabs. Let me break down the order flow.
The HBM Capacity Grab
SK Hynix is the market leader in HBM with over 50% share. Their expansion plans are not incremental. They are building new fabs in Cheongju, South Korea, specifically for HBM production. The equipment for these fabs is not off-the-shelf. It requires advanced TSV (through-silicon via) etching and deposition tools that only a handful of suppliers can provide. Lam Research and Tokyo Electron are the primary beneficiaries here. The etch and deposition steps for HBM are 2-3x more complex than standard DRAM, which means higher equipment value per wafer.
Samsung is playing catch-up, but their equipment orders are ramping. They are converting part of their Pyeongtaek fab to HBM production, which requires significant retooling. Micron is expanding in the US and Japan, with a focus on HBM3E and HBM4. The combined capital expenditure of these three memory makers is expected to exceed $80 billion in 2025, representing 30-40% of their revenue. That is a massive bet.
The Advanced Logic Arms Race
TSMC is the 800-pound gorilla in advanced logic, holding over 90% market share at 5nm and below. Their 2025 capex is projected at $38-42 billion, with a significant portion going to 2nm GAA (gate-all-around) technology. The transition from FinFET to GAA is not a simple node shrink. It requires new deposition and etch tools, and the transition to high-NA EUV lithography is a step-change in cost. Each high-NA EUV system costs over $300 million, and ASML is the only supplier. The delivery lead time for these systems is 24 months or more.
This is where the WFE forecast gets interesting. The equipment density per wafer—the amount of capital required to produce a single wafer—is increasing. For 2nm and below, the equipment cost per wafer is 20-30% higher than at 3nm. This is not just about adding capacity; it is about the increasing complexity of the manufacturing process. Goldman's forecast implicitly assumes that the yield ramp for 2nm and HBM4 will be smooth. If yields are poor, the fabs will need to buy more equipment to compensate, which would actually increase WFE spending. If yields are good, they will need less equipment per wafer, but the demand for AI chips is so high that they will still need to add capacity.
The China Factor
China is a wildcard in this forecast. Chinese fabs like SMIC, Hua Hong, and CXMT are expanding aggressively, but they are constrained by export controls. They cannot buy EUV lithography systems, and they are restricted from purchasing advanced etch and deposition tools. This means their expansion is focused on mature nodes (28nm and above), where domestic equipment suppliers like Naura and AMEC are gaining traction.
The equipment localization rate in China is currently 20-25% by value, with a policy target of 50% by 2030. This is a structural headwind for the global WFE forecast. If Chinese fabs can source more equipment domestically, the demand for imported equipment will be lower than Goldman's forecast implies. However, the Chinese market is still a significant contributor to global WFE spending, and the export controls are not stopping their expansion. They are just changing the mix of equipment they buy.
The Depreciation Cliff
The capex being deployed today will hit the income statement as depreciation in 2027-2029. Memory makers typically use 5-year accelerated depreciation, while foundries use 7-year straight-line. This means the massive capex of 2025-2027 will create a depreciation wall that will suppress gross margins by 5-10 percentage points for memory makers and 3-5 points for foundries.
This is a critical risk that the market is not pricing in. The equipment makers will see strong revenue growth, but their customers will face margin pressure. The question is whether the demand for AI chips is strong enough to offset the depreciation drag. Based on the current order flow, I believe it is, but it is a risk that needs to be monitored.
Contrarian: The Retail Blind Spot and the Smart Money Play
Retail investors are looking at the WFE forecast and seeing a simple story: AI is growing, so buy semiconductor equipment stocks. The smart money is looking at the same forecast and seeing a more nuanced picture. The WFE spending growth rate peaks in 2027 at 45% and then decelerates to 29% in 2028. This is not a sign of weakness; it is a sign of maturation. The first wave of AI infrastructure buildout will be largely complete by 2028, and the growth will shift from capacity expansion to technology upgrades.
The contrarian angle is that the memory makers, not the equipment makers, have the most earnings leverage. SK Hynix is projected to post record net income in 2026 as HBM4 ramps and DRAM prices remain elevated. The market is pricing equipment makers at 25-40x PE, but SK Hynix is trading at 15-20x PE. The earnings growth potential is not fully reflected in the memory stocks.
Another blind spot is the assumption that AI demand will remain strong through 2028. The hyperscalers—AWS, Azure, and Google Cloud—are spending heavily on AI infrastructure, but they are doing so on the expectation of future revenue. If AI applications fail to generate sufficient returns, the capex cycle could be cut short. This is the same mistake we saw in the 2022 Terra/Luna collapse, where circular liquidity was mistaken for real value. In this case, the circularity is between AI chip sales and cloud capex. If one leg breaks, the whole cycle unwinds.
I am not saying this will happen. I am saying that the risk is not zero, and the market is pricing in a smooth path to 2028. The smart money is hedging against the possibility of an AI capex pause in 2026-2027.
Takeaway: The Cycle Has Legs, But the Entry Point Matters
The Goldman Sachs forecast provides a clear three-year growth framework for the semiconductor industry. The WFE spending trajectory is supported by real order flow from memory makers and foundries. The HBM capacity grab is the most visible driver, but the advanced logic transition to 2nm and GAA is equally important.
However, the forecast is not without risks. The AI demand cycle could peak earlier than expected, and the depreciation cliff will pressure margins in 2027-2029. The geopolitical risk, particularly around Taiwan, is a tail risk that could invalidate the entire forecast.
My take is that the equipment makers are the safe play, but the memory makers offer the most upside. The key is to monitor the leading indicators: NVIDIA's earnings guidance, hyperscaler capex announcements, and DRAM contract prices. If these hold up, the cycle extends to 2028 and beyond. If they falter, the WFE forecast will be revised down faster than the fabs can cancel their orders.
Smart contracts execute logic, not intentions. The same applies to capital expenditure. The orders are placed, the equipment is being shipped, and the fabs are being built. The cycle is real. The question is whether the demand behind it is sustainable. I am cautiously optimistic, but I am also watching the order flow for signs of a slowdown. The data will tell us before the headlines do.