The number is staggering. Goldman Sachs is now projecting Wafer Fab Equipment (WFE) spending to hit $218 billion in 2027 and a breathtaking $281 billion by 2028. The market's immediate reaction was a collective exhale. Semiconductor equipment stocks ticked up. Analysts nodded. The AI trade, it seems, has a new floor of support.
But as someone who spent 2017 auditing ICO smart contracts for reentrancy vulnerabilities, I've learned that the most dangerous narratives are the ones that feel most comfortable. The ones that align perfectly with prevailing sentiment. The ones that tell you exactly what you want to hear. And Goldman's forecast, while built on rigorous data, contains hidden assumptions that could unravel faster than a poorly written ERC-20 token.
Let's dissect this. Not as a cheerleader. As an auditor.
The Context: A Cycle Built on Borrowed Time
The semiconductor industry has always been cyclical. Boom, bust, consolidation, repeat. The 2018-2019 downturn saw WFE spending drop over 10%. The 2022 crash was worse. But this cycle feels different because it's being driven by something that looks structurally new: AI.
The narrative goes like this: AI training and inference demand is insatiable. NVIDIA's H100 and B200 GPUs are sold out. CoWoS packaging capacity is the bottleneck. HBM memory is the new gold. Therefore, the logic goes, we need more fabs. We need more equipment. We need more capacity. And we need it yesterday.
Goldman's forecast quantifies this narrative. It assumes AI demand persists through 2028. It assumes 2nm GAA yields ramp quickly. It assumes HBM4 enters production on schedule. It assumes DRAM supply remains tight through the end of the decade.
Each of these is a testable hypothesis. And each has a failure mode.
The Core: Where the Forecast Gets Fragile
Let me walk you through the three structural assumptions that keep me up at night.
Assumption One: The Yield Ramp
The entire 2026-2028 spending peak is predicated on 2nm GAA and HBM4 yields improving rapidly. This is a bold assumption. TSMC's 3nm process took over a year to reach 80% yields. Samsung's 3nm GAA is still struggling around 60-70%. HBM3E yields at SK Hynix are estimated at 70-80%.
Now apply that history to the future. If 2nm yields take longer to ramp—and GAA is a fundamentally different transistor architecture—then equipment purchases get delayed. Not cancelled. But delayed. And in a cyclical industry, delay can be devastating to a forecast. The difference between a 2027 peak and a 2028 peak is the difference between a bull market and a correction.
Assumption Two: The EUV Bottleneck
Here's the number that nobody is talking about. ASML's annual EUV production capacity is roughly 50-60 units. High-NA EUV, which is required for 2nm and below, costs over $300 million per unit. The delivery lead time is 12-18 months.
Goldman's forecast implies that ASML, AMAT, Lam Research, and TEL all dramatically expand their own production capacity. But equipment manufacturers have their own supply chains. Their own bottlenecks. ASML depends on Carl Zeiss for optics. AMAT depends on specialized components. You can't just turn a dial and double production.
Based on my audit experience, when a forecast requires a 20%+ CAGR in a supply-constrained market, the risk isn't in the demand side. It's in the supply side. The equipment makers simply may not be able to deliver.
Assumption Three: The Packaging Shift
Here's what most analysts are missing. The WFE spending structure is shifting. Traditionally, front-end equipment accounted for about 80% of WFE. But HBM and CoWoS expansion is driving a massive increase in advanced packaging equipment spending. TSV etching. Hybrid bonding. Thin wafer processing.
This is a structural shift that benefits different players. Besi, ASM International, and other back-end equipment makers are gaining share. But the market hasn't fully priced this in. The traditional WFE leaders—ASML, AMAT, Lam—are not the primary beneficiaries of this packaging boom.
This is where I see the narrative trap. Investors are buying the semiconductor equipment complex as a monolith. But the composition of spending is changing. Some will win. Some will lose. The index doesn't tell you which.
The Contrarian Angle: The Overcapacity Time Bomb
Here's the counter-intuitive angle that nobody wants to hear. What if the forecast is right? What if WFE spending does hit $281 billion in 2028? Then we're building for an overcapacity crisis in 2029-2030.
The history of semiconductor cycles is a history of herd behavior. In the 2010s, the memory industry overbuilt. DRAM prices crashed. Samsung, SK Hynix, and Micron all suffered. The same dynamics are now playing out globally. The CHIPS Act in the US. The European Chip Act. Japan's semiconductor revival plan. China's Big Fund III. Every major economy is building fabs simultaneously.
This is the onshoring paradox. More geopolitical fragmentation means more redundant capacity. More redundant capacity means a deeper downturn when the cycle turns.
And it will turn. It always does.
Let me give you a specific example from my own experience. In 2020, during DeFi Summer, I built a framework to analyze yield optimization strategies across Uniswap and Compound. I saw the same pattern. Everyone was piling into the same trade. Everyone was assuming the yields would last forever. And then the music stopped. The yields didn't just normalize. They collapsed.
Semiconductor equipment spending is no different. The herd is always late. The smart money is always positioned for the turn before the turn happens.
The Geopolitical Wildcard
We also have to talk about geopolitics. Goldman's forecast is built on a global view. But the US export controls on China are tightening. The Dutch and Japanese restrictions are expanding. And China, which accounts for about 20-25% of global WFE, is being pushed toward domestic equipment.
The problem? Chinese equipment makers like Naura, AMEC, and ACM Research have made progress in mature nodes. But they're years away from competing in advanced processes. EUV lithography is completely off-limits. The gap is not closing quickly.
What does this mean for the forecast? It means that if export controls tighten further, ASML could lose a significant chunk of its revenue. China is ASML's third-largest customer. A complete export ban on mature DUV tools would hurt.
But here's the more subtle point. The onshoring trend—fabs being built in the US, Europe, Japan, and China simultaneously—is actually increasing global WFE demand. Each region needs its own equipment. This is a tailwind for the forecast. But it's also creating a future oversupply problem.
The Financial Engineering View
Let me put on my financial engineering hat for a moment. I have an MS in Financial Engineering, and I've spent years analyzing capital allocation in cyclical industries. The math here is straightforward.
If WFE spending hits $281 billion in 2028, then the depreciation burden on the industry will be enormous. New fabs take 12-24 months to ramp. Advanced nodes take longer. The depreciation peak will hit in 2028-2030, just as the market is likely turning down.
TSMC's gross margins are already projected to drop from ~55% to 50-52% due to depreciation. Memory makers will face even more pressure. The equipment makers will enjoy the boom. But the chipmakers—the ones actually buying the equipment—will see their margins compressed.
This is the classic capex trap. The more you invest, the more you need to run at high utilization to cover your fixed costs. And if demand softens, you're stuck with expensive idle capacity.
The Hidden Assumption No One Is Discussing
Here's the most important hidden assumption in Goldman's forecast: the persistence of AI capital expenditure. The forecast requires that Microsoft, Google, Amazon, and Meta continue to invest heavily in AI infrastructure through 2028.
But what happens if AI monetization disappoints? What if large language models hit a plateau? What if CSPs start building their own ASICs and reduce their dependence on NVIDIA? Any of these scenarios would compress the AI capex cycle.
And let's be honest: we've seen this movie before. In 2000, everyone was building fiber optic networks. The narrative was that internet traffic would double every 100 days forever. It did grow. But not fast enough to justify the capex. The overbuild destroyed the telecom equipment industry.
AI is following a similar trajectory. Not in terms of the technology—the technology is real. But in terms of the capex cycle. The buildout is happening faster than the revenue is being generated. At some point, the market will demand evidence of ROI. And if that evidence is slow to materialize, the capex gets cut.
Goldman's forecast has no room for this scenario. It's a linear extrapolation of current trends. And linear extrapolations always fail.
The Signal in the Noise
So what should investors actually do with this information? The forecast is a useful data point. But it's not a guarantee. It's a scenario. And scenarios have probabilities.
Here's my take. The equipment cycle is real. The AI demand is real. But the magnitude and duration of the cycle are uncertain. The market is pricing in the bullish case. The risk is asymmetric.
I'd rather be positioned for the contrarian outcome. The one where yields ramp slower than expected. Where ASML can't deliver enough EUV tools. Where AI capex gets rationalized in 2027. Where the overcapacity time bomb explodes in 2029.
That's not a bearish view on semiconductors. It's a bullish view on analysis. The people who get hurt in this market are the ones who treat forecasts as certainties. The ones who don't audit the assumptions. The ones who don't ask what could go wrong.
I've spent 23 years in this industry. I've seen cycles come and go. I've seen narratives build and collapse. The one constant is that the crowd is always late.
The question isn't whether Goldman is right. It's whether the market has already priced in the forecast. And if it has, the upside is limited. The downside is not.
The Takeaway: What Comes Next
The next narrative shift is already forming. It's not about AI. It's not about equipment. It's about the transition from the buildout phase to the optimization phase. The companies that will win in 2028 are the ones that are building efficiency into their operations today. The ones that are positioning for the inevitable downturn.
I'm watching the equipment backlog ratios. ASML, AMAT, and Lam have backlogs of 1.5-2.0x revenue. That's historically high. It tells me the current demand is real. But it also tells me that the market is front-loading orders. When the backlog normalizes—and it will—the equipment stocks will correct.
The smart play is not to chase the equipment complex at these valuations. It's to identify the companies that will benefit from the next cycle. The ones that are building the infrastructure for the post-AI era. The ones that are solving the overcapacity problem. The ones that are creating efficiency through software and analytics.
The narrative is always evolving. The trick is to see the evolution before the crowd does.
That's the hunt. That's the game. And it's not over yet.
History doesn't repeat. But it rhymes. And the rhyme I hear right now is the sound of a cycle peaking. Not collapsing. Just peaking. The equipment makers will have their moment. But it's not the moment that matters. It's what comes after.
I'll be watching. I'll be auditing. And I'll be positioned for the turn. Not because I'm bearish. But because I've seen this movie before. The ending never changes.
Just the details.