The WFE spending trajectory is a stack trace. And this stack trace tells a specific story: $1,505 billion in 2026, $2,160 billion in 2027, $2,810 billion in 2028. Goldman Sachs raised its semiconductor equipment cycle forecast on August 25, 2025, extending the growth window to 2028. The pattern: 36% growth in 2026, 45% in 2027, 29% in 2028. The peak arrives in 2027. Then it fades.
This is not an opinion. It's a structural prediction embedded in a series of assumptions about DRAM, HBM, and advanced foundry expansion. My job is to read those assumptions, trace their failure modes, and identify what the forecast silently takes for granted. As someone who has spent nearly two decades auditing systems for hidden vulnerabilities — first smart contracts, now the infrastructure layer that powers AI and crypto alike — I know that the most dangerous predictions are the ones that look like facts but are actually compound bets on unverified premises.
Let me walk through the architecture of this forecast like I would walk through a protocol's smart contract code. Line by line. Assumption by assumption. Because the stack trace doesn't lie — but only if you know how to read the full trace.
The Three Expansion Vectors
Goldman's forecast rests on three pillars: DRAM process scaling, HBM technology iteration, and advanced foundry expansion. Each has its own failure modes, its own bottlenecks, and its own timeline. The forecast treats them as if they move in concert. In practice, they move independently. That discrepancy is where the risk lives.
DRAM process scaling is the first pillar. The industry is transitioning from 1-alpha/1-beta nanometer nodes to 1-gamma/1-delta nodes. That's the 10-15nm range. Each node transition requires new equipment: more deposition tools, more etching systems, more metrology. But DRAM scaling is no longer driven by density alone. It's driven by AI's insatiable appetite for memory bandwidth. HBM3E is already in production. HBM4 is expected in 2025-2026, with stack heights moving from 8 layers to 12 to 16. Each layer addition compounds the equipment requirements.
The hidden assumption here is that HBM4 and subsequent generations will hit their production targets on schedule. If HBM iteration slips by even one quarter, the equipment spending curve shifts down. Goldman's forecast has no slack for schedule slippage. That's a structural fragility, not a minor detail.
The second pillar is advanced foundry. The transition from 3nm to 2nm GAA (Gate-All-Around) nodes is the primary driver. This transition requires high-NA EUV equipment. ASML is the sole supplier of EUV lithography. The unit price of a high-NA EUV scanner exceeds $300 million. Delivery times run 24 months or longer. ASML's annual EUV capacity is roughly 50-60 units. If demand outstrips that capacity, the delivery queue extends. And extension in the equipment delivery queue translates directly into delayed wafer starts. Delayed wafer starts translate into delayed revenue. The whole chain is latency-bound.
The third pillar is HBM itself. And here is where the cost structure gets interesting. HBM is not just a memory product. It is a packaging miracle. TSV — through-silicon via — requires advanced stacking. HBM3E's 8-layer stack consumes three to four times the equivalent wafer capacity of a standard DDR5 product. This means every HBM wafer allocated is a DDR5 wafer not produced. The DRAM supply tightness that Goldman's forecast relies upon is directly caused by this wafer consumption ratio. The forecast is not predicting a shortage. It is predicting a structural shift in how wafers are allocated.
The Equipment Bottleneck Chain
Now let me trace the supply chain, because this is where the forecast's fragility is most visible.
ASML holds 100% of the EUV lithography market. This is not a competitive advantage. It is a monopoly. And monopolies create single points of failure. If ASML's production line hits any disruption — a supply chain shock in precision optics, a skilled labor shortage, a logistics break — the entire global advanced-node roadmap slips. There is no substitute supplier. There is no fallback. There is only the queue.
For etching and deposition, the market is an oligopoly. Lam Research holds roughly 35% of the etching market. Tokyo Electron holds about 30%. Applied Materials holds about 20%. These three companies control the equipment that creates the structural integrity of every advanced chip. Their pricing power is extraordinary. Their margins reflect it. But their concentration also means that a single supply chain disruption at any one of these firms — a fire in a plant, a materials shortage, a labor dispute — hits every wafer fab simultaneously. The system has no redundancy.
The downstream customer base is equally concentrated. The top ten wafer fabs — TSMC, Samsung, Intel, SK hynix, Micron, SMIC and others — account for over 80% of all equipment purchases. This is not a diversified market. It is a small group of enormous buyers dependent on a small group of enormous suppliers. That kind of structure is efficient until it breaks. When it breaks, it breaks everything at once.
The Geopolitical Wildcard
Goldman's forecast silently assumes that geopolitical risk remains contained. That's the most fragile assumption in the entire prediction.
Consider the export control regime. The US restricts advanced semiconductor equipment sales to China. This covers logic below 14nm, DRAM below 18nm, and NAND above 128 layers. The license approval rate for advanced equipment exports to China is less than 10%. ASML has been blocked from selling EUV and advanced immersion DUV tools to Chinese buyers since 2024. The only equipment that flows through is mature-node gear — the KrF and i-line systems that power the 28nm and above markets.
China's response is predictable. The National Fund III, capitalized at 344 billion RMB, is focused on equipment, materials, and EDA tools. Domestic suppliers like Naura and AMEC are pushing into the mature process market with increasing success. The equipment localization rate is roughly 20-25% by value. The policy target is 50%. The timeline is around 2030.
If China's domestic equipment suppliers succeed faster than expected, the global WFE market faces a structural downward revision. The addressable market for ASML, Applied Materials, and Lam Research shrinks. Goldman's forecast assumes a stable global demand curve. But the curve is not stable. It is actively being reshaped by export controls and retaliation. The geopolitical variable is not a tail risk. It is a present, active force.
The 2027 Peak Pattern
The forecast's growth trajectory deserves closer scrutiny. The shape of the curve — 36% growth in 2026, 45% in 2027, 29% in 2028 — is the diagnostic fingerprint of a cycle that peaks in 2027. The slowdown in 2028 is not an accident. It is the architecture revealing itself.
This pattern implies that Goldman's analysts expect the first major wave of AI infrastructure investment to reach saturation in the 2027-2028 window. The spending increase has a front-loaded profile. The second derivative turns negative. The direction of the curve matters more than the absolute level. The WFE spending is not heading to a plateau in 2028. It is heading to a ceiling.
The DRAM Supercycle Question
The DRAM supply tightness is real. I do not dispute that. HBM's wafer consumption ratio is a mechanical fact. The base of HBM expansion is consuming capacity that would otherwise produce standard DRAM. This is a structural shortage, not a cyclical one. The DRAM contract prices rose 15-25% quarter-over-quarter in the second and third quarters of 2025. HBM pricing remains three to five times above standard DRAM. The semiconductor memory market is in a genuine supply-demand imbalance.
The question is whether this imbalance persists through 2028, as Goldman's forecast implies. The comparison to the 2017-2018 memory supercycle is instructive. That cycle lasted roughly two years. It was driven by smartphone growth and cloud expansion. It collapsed when supply caught up with demand. The current cycle is driven by AI infrastructure spending, which has a longer build-out timeline. But it is not immune to the same oversupply dynamics. If HBM yields improve faster than expected — if SK hynix, Samsung, and Micron all hit their HBM4 targets on schedule — the supply will catch up. The shortage will normalize. The equipment spending will follow.
The point is not that Goldman is wrong. The point is that the forecast has no margin of safety built in. It assumes the cycle runs its course at a predictable, monotonic path. No scenario. No downside case. No stress test.
What the Bulls Got Right
I've been the critic in this piece. Let me now do what a good auditor does: acknowledge what the forecast gets right.
The first is the persistence of AI demand. This is not the same as the crypto mining boom of 2017 or the DeFi summer of 2020. AI infrastructure is not a speculative asset. It is productive capital. Large language models need compute. Inference needs compute. The compound growth of training compute is roughly doubling every three to four months. That is not a bubble. It is an exponential curve that has not yet flattened.
Second, the HBM bottleneck is genuine. The TSV stacking technology, the CoWoS packaging, the advanced interconnects — these are not marketing narratives. They are physical constraints. The yield rates on HBM4 will not improve linearly. They will improve through iterative engineering. That process takes time. The forecast correctly identifies that the equipment spend will continue during this iteration phase.
Third, the memory oligopoly is real. SK hynix, Samsung, and Micron control the HBM market. SK hynix has more than 50% of the HBM market. These players have deep pockets and high barriers. They can fund their own expansion. They can sustain long capital expenditure cycles without collapsing. The oligopoly structure gives the forecast some underlying stability.
The Auditor's Verdict
So what is my final verdict on the Goldman forecast?
The 2026-2028 WFE cycle is a defensible prediction. The structural drivers are real. The HBM bottleneck is real. The advanced foundry transition is real. The AI demand curve is real. But the forecast's architecture has three critical failure points.
First, it assumes no HBM yield slippage. If HBM4 yield ramps slower than expected — if SK hynix hits a 70% yield threshold that takes two quarters longer than planned — the entire equipment spending trajectory shifts. The capex and equipment demand is still there. But the timing changes. And in a cycle that peaks in 2027, timing is everything.
Second, it assumes geopolitical containment. The Taiwan scenario is the tail risk that ends all forecasts. If the Strait escalates, the semiconductor supply chain fractures. The WFE spending would not decline 30%. It would collapse. No forecast survives that scenario. But a credible forecast should at least acknowledge the probability.
Third, the forecast assumes the demand curve remains linear through the cycle. It does not model the possibility that AI infrastructure investment becomes saturated in 2027, that the cloud providers pause their capex to absorb the massive new capacity they've already deployed. The 2028 slowdown is baked into the numbers. The question is whether it arrives on schedule.
The Takeaway
Here is what I want you to take from this analysis. The forecast is a framework, not a guarantee. It is useful for structuring how you think about the semiconductor cycle. It is not useful for timing specific equipment purchases. The 2027 peak is the signal to watch. When WFE spending growth decelerates, the market will not see it coming. The market will be focused on the absolute level, not the direction. The stack trace will show the first signs of a cycle turning, but you have to know where to look.
I've spent years auditing systems that promise reliability. The best ones are built with fail-safes. They are built with the assumption that something will fail. The semiconductor equipment cycle forecast, as presented, has no fail-safe. It assumes the cycle runs its course without interruption. The stack trace doesn't lie, but it only tells you what you have already observed. The question is what you don't know — the yield data that hasn't been published yet, the geopolitical event that hasn't occurred yet, the demand curve that hasn't bent yet.
Those are the variables that will determine whether Goldman's forecast holds or breaks. Those are the variables you should be tracking. Not the headline numbers. The stack trace beneath them.