Hook: The Number That Changes Everything
Goldman Sachs just dropped a number that should make every semiconductor executive sweat: $281 billion in wafer fab equipment spending by 2028. That's not a typo. The investment bank has revised its WFE expenditure forecast upward, projecting a compound annual growth rate north of 20% from 2024 through the end of the decade.
Here's what's wild about that figure—it's not just about more of the same. The structure of that spending is fundamentally shifting. DRAM and HBM are becoming the primary drivers, not logic chips. Advanced packaging is eating a bigger slice of the pie. And high-NA EUV lithography—at $300 million per machine—is about to blow up the capital intensity math for every leading-edge fab on the planet.
But here's the part nobody's talking about: Goldman's prediction contains an implicit assumption about AI demand that could be catastrophically wrong.
Let me show you what I mean.
Context: Why This Cycle Is Different
The semiconductor equipment market has always been cyclical. Boom, bust, repeat. But the 2024-2028 cycle has a different DNA. Three structural forces are converging that didn't exist in previous upcycles.
First, AI is not a cyclical demand driver—it's a paradigm shift. The HPC/AI training segment alone accounts for 25-30% of semiconductor revenue in 2024, growing at 40%+. AI inference is growing even faster at 50%+. NVIDIA's H100/H200/B200 GPUs are sold out months in advance. This isn't a typical replacement cycle; it's the creation of an entirely new compute category.
Second, memory is back in a big way. DRAM contract prices rose 10-15% in Q3 2024. NAND jumped 15-20%. SK Hynix, Samsung, and Micron are all racing to expand HBM capacity, and Goldman expects DRAM supply tightness to persist through 2028. That's not a blip—that's a super-cycle.
Third, geopolitical fragmentation is forcing duplicate fab construction. The US CHIPS Act, Europe's Chip Act, Japan's semiconductor revival plan, and China's Big Fund III (¥344 billion) are all funding parallel semiconductor ecosystems. Every region wants its own fabs, which means every region is buying equipment. Goldman's forecast partially captures this, but I'd argue it doesn't fully price in the inefficiency of this fragmentation.
The current capacity utilization numbers tell the story: leading-edge (5nm and below) is running at 90%+, while mature nodes (28nm+) sit at 75-80%. The market is bifurcated, and that bifurcation is about to get more extreme.
Core: The Equipment Spending Bonanza—And Its Hidden Fault Lines
Let me walk through what Goldman's $281 billion forecast actually implies, because the devils in the details.
The HBM and CoWoS Bottleneck
Here's the first thing most analysts miss: CoWoS advanced packaging capacity is the real constraint on AI chip supply, not wafer fabrication. TSMC's CoWoS capacity was roughly 400,000 wafers per year (12-inch equivalent) in 2024. The company plans to double that to 800,000 in 2025, and it's still not enough.
The equipment implications are massive. HBM requires TSV (through-silicon via) etching, hybrid bonding, and thin-wafer handling equipment that traditional logic fabs don't need. Companies like Besi and ASM International—historically considered second-tier equipment makers—are suddenly critical players. The advanced packaging equipment segment is growing faster than the overall WFE market, and that trend accelerates through 2028.
Based on my audit experience tracking equipment orders, I can tell you the bottleneck isn't demand—it's delivery. ASML's EUV annual capacity is only 50-60 units. The delivery lead time for EUV systems is 12-18 months. High-NA EUV is even tighter. You can't just flip a switch and produce more $300 million machines.
The 2nm and HBM4 Assumption
Goldman's forecast implicitly assumes that 2nm GAA production ramps smoothly in 2025-2026 and that HBM4 hits volume production in 2026. Both are reasonable assumptions, but they carry significant execution risk.
TSMC's N2 is on track, but yields at the 2nm node with GAA architecture are unproven at scale. Samsung's 3nm GAA yields are reportedly 60-70%—meaningfully below TSMC's 80%+ at comparable nodes. If 2nm yields disappoint, equipment spending gets delayed, not canceled, but the 2027-2028 numbers could slip.
For HBM4, the move to hybrid bonding is a fundamental process change, not an incremental improvement. SK Hynix's HBM3E yields are 70-80%, but HBM4's more complex stacking requirements could reset that learning curve. Every yield percentage point matters because it directly determines how much equipment gets ordered.
The Memory-Led Structural Shift
Here's something the market hasn't fully digested: the equipment spending mix is shifting from logic to memory. Historically, logic fabs accounted for roughly 60% of WFE spending. Goldman's forecast implies memory will be the primary growth driver through 2028.
This matters because memory fabs require a different equipment mix. More etch tools from Lam Research. More deposition systems from TEL and AMAT. More advanced packaging equipment. The winners in this cycle won't be the companies with the best EUV story—it'll be the ones with the strongest memory and packaging equipment portfolios.
Contrarian: What Goldman's Forecast Gets Wrong
Now let me flip the narrative. Because there's a version of this story that ends very differently.
The AI capex assumption is the single point of failure. Goldman's WFE forecast requires hyperscalers—Meta, Google, Microsoft, Amazon—to sustain AI capital spending growth through 2028. That's a massive bet. If AI monetization disappoints, if CSPs decide to throttle their AI infrastructure buildout in 2026-2027, the entire WFE projection collapses.
The historical precedent is sobering. In 2018-2019, WFE spending dropped over 10% when memory prices crashed. The semiconductor industry has never sustained 20%+ CAGR over five years without a significant correction. The "herd effect" in equipment spending has historically led to overcapacity—just look at the 2010s memory expansion cycle that ended in a price war.
China's mature-node overcapacity risk is being ignored. Chinese fabs are adding 28nm+ capacity at a furious pace, with SMIC, Hua Hong, and others expanding across Beijing, Shanghai, and Shenzhen. By 2025-2027, this capacity comes online simultaneously. The result: mature-node price wars that could pressure everyone's margins. Goldman's forecast focuses on leading-edge and memory, but the oversupply in mature nodes could create a deflationary shock that ripples through the entire equipment ecosystem.
The geopolitical wildcard. China accounts for 20-25% of global WFE spending and is ASML's third-largest customer at ~15% of revenue. If the US tightens export controls further—including mature-node equipment—the impact on equipment makers' revenue expectations would be significant. The current regulatory framework is a moving target, and Goldman's global view may not adequately weight this risk.
Takeaway: The Verification Protocol
Here's what I'm watching to validate or invalidate Goldman's thesis:
Equipment order books. ASML, AMAT, and Lam Research's book-to-bill ratios are the canary in the coal mine. Current ratios of 1.5-2.0 support the bullish case. Any sustained decline below 1.0 would signal the cycle is turning.
Memory pricing. DRAM contract prices need to keep rising. If they flatten or reverse before 2026, the memory super-cycle thesis is wrong.
AI capex guidance. Every hyperscaler earnings call needs to be scrutinized for AI infrastructure spending commitments. Any hesitation is a bearish signal for the entire equipment complex.
Yield reports from 2nm and HBM4. TSMC's N2 yield updates and SK Hynix's HBM4 progress will determine whether the 2027-2028 equipment peak happens on schedule.
The equipment cycle is real, but it's not guaranteed. Goldman's $281 billion forecast is a roadmap, not a destiny. The smart money isn't betting on the number—it's watching the data that will confirm or break it. Follow the equipment orders, not the analyst reports. The chart didn't lie in 2018, and it won't lie now.