Goldman Sachs has raised its wafer fab equipment (WFE) forecast to $281 billion by 2028, up from $150 billion in 2026 โ a 37% compound annual growth rate that implies the semiconductor industry is entering a structural expansion, not a cyclical uptick. The market reads this as AI conviction. I read it as a plumbing statement. Equipment spend is the earliest physical signal of compute supply โ the layer beneath every AI token, every DePIN network, every on-chain inference market. Before assessing what this means for semiconductor equities, we need to map what it means for the liquidity architecture that crypto's AI narrative depends on.
The forecast's internal math is not a smooth extrapolation. The trajectory โ $150B (2026) โ $218B (2027) โ $281B (2028) โ embeds three distinct assumptions. First, that AI capex from hyperscalers (Microsoft, Google, Amazon, Meta) remains at or above $300 billion annually through 2028. Second, that export controls on semiconductor equipment do not tighten further, allowing China to maintain $40-50 billion in annual equipment purchases. Third, that equipment suppliers โ ASML, Applied Materials, Lam Research, Tokyo Electron, KLA โ can physically deliver the tools required.
Each assumption deserves scrutiny. My audit experience with supply chain stress testing tells me that when a forecast embeds three independent optimistic assumptions, the probability of full realization is the product of their individual probabilities, not their average. If AI capex holds (70%), export controls remain rational (75%), and delivery capacity scales (80%), the joint probability drops to roughly 42%. The market is pricing this forecast as if it were a certainty. It is not.
The HBM Second Growth Engine
The most underappreciated finding in this cycle is that HBM (High Bandwidth Memory) equipment demand constitutes a second, non-overlapping growth curve. HBM3E (12-layer) and HBM4 (16-layer) production requires TSV etching, electroplating, temporary bonding and debonding, and advanced packaging tools that do not overlap with logic chip equipment. This transforms the equipment market from a single-engine (logic) to a dual-engine (logic plus memory) structure.
The math is compelling. SK Hynix, Samsung, and Micron's combined capex is projected to rise from roughly $60 billion in 2024 to $100 billion-plus by 2027, with HBM as the primary driver. DRAM supply is expected to remain tight through 2028 โ contract prices rose 10-15% in Q4 2024 and are projected to rise another 20-30% in 2025. This is not a typical memory cycle. Traditional DRAM is cyclical; HBM is structurally tied to AI training demand, which means the storage industry may be undergoing a re-rating from cyclical to growth โ a fundamental shift in how we value memory manufacturers and their equipment suppliers.
My Monte Carlo simulations on storage supply-demand dynamics suggest that the HBM demand curve is less elastic than the market assumes. The qualification cycle for HBM in AI accelerators is 12-18 months, meaning customers cannot easily switch suppliers mid-cycle. This lock-in effect extends the pricing power of memory manufacturers and, by extension, the equipment vendors that supply them. The storage equipment segment is no longer a derivative of the consumer electronics cycle; it is now a direct function of AI infrastructure spending.
The Capacity Wave
The $281 billion WFE figure implies a massive capacity addition. Using an advanced-node benchmark of $15-20 billion per 10,000 wafer starts per month, the forecast implies 1.4-1.9 million additional monthly wafer starts (12-inch equivalent) by 2028 โ equivalent to 14-19 new large-scale fabs. These are not hypotheticals; the announced projects are concrete. TSMC Arizona (three fabs, $65 billion), Samsung Taylor (two fabs, $37 billion), SK Hynix Yongin (four fabs, $90 billion), Micron New York and Idaho (two DRAM fabs, $100 billion-plus). The deployment timeline โ 2025 through 2028 โ aligns precisely with Goldman's WFE peak.
We mapped the water, not the wave. The equipment order book is the water; the capacity that comes online in 2027-2028 is the wave. And waves, by definition, crash. The depreciation overhang is the detail most analysts miss. New fabs carry 5-7 year depreciation schedules โ TSMC uses 5 years, Samsung 7. In the first 1-2 years of operation, depreciation drags gross margins down 5-10 percentage points. A new advanced fab needs 70-80% utilization just to cover depreciation costs. If the 2028 supply wave coincides with any AI capex deceleration, the utilization shortfall will be brutal.
The Export Control Assumption
Here is where the forecast gets fragile. Goldman's projection requires China to remain a meaningful equipment purchaser โ an estimated $40-50 billion annually. But the current trajectory of US export controls suggests the opposite. The December 2024 tightening on HBM exports, the Entity List designations for SMIC, YMTC, and CXMT, and the 2023 Japan-Netherlands coordination on DUV restrictions all point to further tightening, not rationalization. The probability of the pessimistic scenario โ full decoupling including mature-node equipment โ sits at 25% in my assessment. That scenario would reduce global WFE by 10-15%.
China's response has been structural. The third phase of the National IC Fund (344 billion yuan) launched in 2024, targeting equipment and materials localization. Domestic equipment localization stands at 20-25% for mature nodes, with a target of 50%+ by 2028. The equipment cycle extension creates what I call the time dividend โ longer expansion windows give Chinese equipment makers (NAURA, AMEC, Piotech) more opportunities for in-line validation and iteration. This is the hidden variable in Goldman's model. If domestic equipment achieves critical breakthroughs in mature-node etch and deposition by 2026-2028, China's equipment purchases could exceed expectations, becoming an incremental growth driver for the WFE forecast rather than a risk.
The supply chain concentration is the structural vulnerability. EUV lithography is 100% dependent on ASML, with no alternative source. High-end etch is controlled by AMAT, Lam, and TEL at 80%+. Metrology is KLA-dominated at 55%. The entire equipment supply chain is concentrated in three countries: the US, Japan, and the Netherlands. This is not a diversified system; it is a tripod. Any geopolitical shock to any of the three legs destabilizes the entire forecast.
Oligopoly Pricing Power
The competitive structure of the equipment industry remains the strongest moat in the semiconductor value chain. ASML holds roughly 85% of lithography. KLA holds approximately 55% of metrology. The top three (AMAT, Lam, TEL) control 80-90% of etch and deposition. Customer switching costs are extreme โ equipment qualification cycles run 2-3 years, creating a lock-in effect that transcends pricing power. Gross margins reflect this: ASML at 51%, KLA at 61%, with the industry average in the 45-60% range.
A ledger is a confession written in code. The equipment industry's financials confess the same truth: ROIC runs 25-45 percentage points above WACC, and operating cash flow consistently exceeds net income by 1.0-1.3x. These are not cyclical earnings; they are structural rents from technological monopoly. The industry's R&D intensity โ 10-15% of revenue โ is the moat that keeps new entrants at bay. ASML alone spent $4.5 billion on R&D in 2024; the entire Chinese domestic equipment industry spends a fraction of that. The absolute gap in R&D spending (10-30x) is the fundamental constraint on any localization narrative.
The Contrarian Angle: The Supercycle Is the Setup for the Next Downcycle
The consensus narrative is straightforward: AI demand is structural, therefore the equipment cycle is a supercycle, therefore equipment stocks deserve their 30-35x PE multiples. The contrarian reading is less comfortable. The capacity math implies that 1.4-1.9 million additional monthly wafer starts come online by 2028. That is a supply wave that will hit the market precisely when hyperscaler AI capex may be decelerating from its peak. The semiconductor industry has never escaped its cyclicality โ it has only extended the period between corrections.
The second contrarian observation concerns the equipment forecast's embedded assumption about export control rationalization. If the US tightens controls further โ particularly on mature-node equipment โ the WFE forecast faces a 10-15% downside. The industry would face a double-loss: China's $40-50 billion in purchases would contract sharply, and global equipment vendors would lose 20-30% of revenue. This is the scenario that Goldman's model quietly assumes away.
The third blind spot is the delivery bottleneck. ASML's EUV delivery cycle runs 12-18 months; high-NA EUV extends to 18-24 months. KLA metrology tools run 6-12 months. The equipment supply chain โ precision optics from Zeiss, RF power supplies, precision ceramics โ has its own capacity constraints. If equipment suppliers cannot physically deliver the tools, the WFE forecast becomes a wish, not a plan.
The Crypto Connection
For crypto markets, the WFE forecast is not a semiconductor story; it is a liquidity signal. The AI narrative in crypto โ AI tokens, DePIN compute markets, on-chain inference โ is derivative of the physical compute buildout. When equipment orders peak, that is the top of the AI capex cycle. When the supply wave hits in 2028, compute prices fall, and the economics of every AI-token project that depends on scarce compute will be repriced.
The structural insight is that crypto's AI infrastructure bet is not a bet on tokenomics; it is a bet on semiconductor supply chains. The same oligopoly that controls equipment supply controls the pace at which decentralized compute networks can scale. The decentralization narrative collides with the reality of a supply chain concentrated in three countries and five companies. This is the friction point that institutional investors will eventually price into AI-token valuations.
Takeaway
The WFE forecast is directionally correct but probably 10-15% too optimistic. The equipment supercycle is real, but it is not infinite. For crypto investors, the signal to watch is equipment order data โ specifically ASML's EUV order book and the DRAM/HBM capex trajectory. When those peak, the AI liquidity engine that powers the crypto narrative will begin its deceleration. The question is not whether the supercycle happens. It is whether you are positioned before the supply wave arrives, or after it crashes.
