Goldman Sachs is projecting a 36% CAGR in wafer fab equipment spending through 2028. That's a $281 billion peak. The market is treating this like a linear extrapolation of AI demand. It's not. It's a complex options chain where the underlying is geopolitical tension, the implied volatility is delivery bottlenecks, and the biggest risk isn't the demand side—it's the supply of a single machine that costs more than a commercial airliner.
Let's cut through the noise. The forecast hinges on a few key assumptions that most retail traders are glossing over. The first is the seamless integration of High-NA EUV lithography. The second is a historic shift in spending toward memory. And the third is that AI capex, the fuel for this entire fire, doesn't stall. I've spent the last decade watching these cycles snap. The code bleeds, but the liquidity stays cold. This time is no different.
The High-NA EUV Bottleneck
The entire 2027-2028 projection of $218B and $281B rests on a single point of failure: ASML's ability to deliver its EXE:5200 High-NA EUV systems. We're not talking about incremental upgrades here. These are €300-400 million machines with optics so precise they'd be considered a physics experiment if they weren't already in production. ASML's annual EUV capacity is roughly 50-60 units. High-NA is a fraction of that.
Based on my audit experience, when a system has a single point of failure, you don't plan for the happy path. You plan for the latency. The current delivery timeline has first units shipping in 2025-2026, with volume ramp in 2027. But the supply chain for these machines—the Zeiss optics, the Cymer light sources—isn't scaling at the same rate as the demand curve Goldman is modeling. If High-NA slips by six months, the entire 2028 peak shifts right, and the CAGR drops faster than a leveraged position in a flash crash.
The market is pricing in perfection. The physical world rarely delivers it.
The Memory Shift Nobody's Hedging
Goldman's report correctly identifies DRAM/HBM as the primary growth driver. But the implication is deeper than just "AI needs memory." It signals a structural shift in WFE spend composition. For the next three years, memory will outspend logic. That's a reversal of the last decade's trend.
Here's the math that matters. HBM3E consumes 3-4x the DRAM die area of standard DDR5. HBM4, entering production in 2025-2026, requires hybrid bonding—a process that demands entirely new equipment precision. SK Hynix, Samsung, and Micron are looking at a combined $50B+ in HBM-related capex through 2027. This isn't just about adding cleanroom space. It's about retrofitting fabs for a fundamentally different packaging paradigm.
The hidden assumption here is that DRAM supply remains tight through 2028. That implies HBM demand will absorb a massive chunk of the world's DRAM output. If AI inference demand hits a speed bump—if the models don't get cheaper to run, or the monetization doesn't materialize—the DRAM glut returns faster than a bear market rally. The supply-demand balance is a knife's edge, and Goldman is betting the blade stays sharp.
The Delivery Trap
Let's talk about the physical constraints that spreadsheets ignore. WFE spending is not a pure demand function. It's a function of what equipment vendors can actually ship. Applied Materials and Lam Research are sitting on 12-18 month lead times. ASML's EUV queue is backed up. The industry's capacity to expand its own capacity is limited by the same supply chain it's trying to serve.
This creates a fascinating dynamic. If demand stays hot, equipment vendors gain pricing power. They can raise prices 5-10% annually, and customers will pay because the alternative is falling behind in the AI arms race. This is a seller's market, and the "shovel sellers" are going to extract maximum rent. But it also means the actual WFE spend might be capped by delivery capacity, not demand. The forecast could be 10-15% too high simply because the machines can't be built fast enough.
Incentives align only when the risk is priced in. Right now, the risk of delivery failure isn't priced into equipment stocks. The market sees the revenue projection and assumes it's a lock. It's not. It's a constraint problem dressed up as a demand story.
The China Variable
Goldman's model is built primarily on non-China demand. That's a critical caveat. China accounts for 20-25% of global WFE spending, and it's being systematically cut off from advanced equipment. The US export controls are tightening, not loosening. If Washington extends restrictions to mature-node equipment—a real possibility in the next round of rules—China's WFE spend could drop by half.
That's a 10-12% hit to the global number that Goldman isn't fully modeling. The offset is supposed to come from the US, Europe, and Japan building out their own fabs. But those projects are plagued by cost overruns and labor shortages. The CHIPS Act money is flowing, but the construction timelines are slipping. The assumption that the rest of the world can simply absorb China's loss is optimistic.
Meanwhile, China's response is the wildcard. The Big Fund III, with ¥344 billion, is pouring money into domestic equipment makers. Companies like Naura, AMEC, and ACM Research are targeting 30-50% annual growth. They're not going to crack the advanced node barrier in three years—the EUV moat is too deep. But in mature nodes, they're becoming legitimate competitors. This isn't a near-term threat to ASML or Applied Materials, but it's a structural change in the competitive landscape that the market is underpricing.
The Cycle Peak
Here's the part that keeps me up at night. The forecast shows growth decelerating from 45% in 2027 to 29% in 2028. That's the classic signature of a cycle top. WFE spending is brutally cyclical. We saw it in 2018, we saw it in 2022. Three years of hyper-growth is usually followed by a sharp correction as capacity catches up with demand.
Goldman is essentially predicting the peak of this cycle in 2028. If that's the case, the smart money should be positioning for the downcycle now, not chasing the last leg up. The equipment stocks are trading at 20-35x earnings, which is reasonable if the growth materializes. But if the cycle turns in 2029, those multiples will compress violently. Volatility is the only constant truth.
The market is treating this as a new paradigm—a "growth cycle" where AI demand has permanently smoothed out the cyclicality. I've heard that story before. It was called the "supercycle" in 2021, and it ended with inventory write-downs and cancelled orders. The physics of semiconductor manufacturing haven't changed. The demand might be different, but the supply response is the same. It takes 2-3 years to build a fab, and when all the fabs come online at once, the oversupply hits like a brick wall.
The Contrarian Play
So where does that leave us? The consensus is long equipment stocks, long AI, long the future. The contrarian angle is to recognize that the biggest risk isn't demand—it's execution. The High-NA EUV ramp, the HBM transition, the delivery bottlenecks, the geopolitical shocks. Any one of these can break the chain.
I'm not saying the Goldman forecast is wrong. I'm saying it's a scenario, not a certainty. The market is pricing it as a certainty. That's the opportunity. When the leverage snaps, the silence is loud. The question is whether you're positioned for the noise or the silence.
Liquidity is a mirror, not a floor. The order flow is telling you what the big money is doing, not what the fundamentals are. Right now, the order flow is chasing the equipment trade. The smart money is likely hedging against the delivery risk, the China risk, and the cycle risk. The question is whether you're on the right side of that trade.
The Takeaway
Watch the ASML order book. Watch the lead times. Watch the DRAM spot prices. These are the leading indicators that will tell you if the Goldman forecast is on track or if it's a house of cards. The 2028 peak is a target, not a guarantee. The path to that peak is full of potholes, and the market is only pricing in the smooth road.
I don't trade on forecasts. I trade on the divergence between the forecast and the physical reality. Right now, the physical reality is that the machines can't be built fast enough, the geopolitical environment is deteriorating, and the cycle is aging. The code bleeds, but the liquidity stays cold. Position accordingly.