Goldman's 2028 WFE Forecast: A Liquidity Cascade in Silicon

While the market fixates on Nvidia's earnings and the daily chatter of AI bubbles, the liquidity structure reveals something else entirely. On August 25, Goldman Sachs raised its semiconductor equipment cycle forecast, extending the WFE (Wafer Fabrication Equipment) spending boom to 2028. The headline numbers are stark: 2026 spending projected at $150 billion, ballooning to $281 billion by 2028. That is not a cyclical uptick. That is a structural repricing of capital allocation across the entire global technology stack. Most analysts will read this as a bullish signal for ASML, AMAT, and Lam Research. They are looking at the order books. I am looking at what this means for the liquidity cascade. This is not a story about chipmakers. This is a story about where the next five years of global liquidity will be parked. And for those of us watching the macro flows, this forecast is the clearest signal yet that the AI-driven buildout is not a speculative mania — it is the new physical layer of the machine economy. The numbers demand a forensic breakdown. Goldman's trajectory implies a 36% growth in 2026, a 45% surge in 2027, and a 29% expansion in 2028. The shape matters more than the absolute figures. The peak in 2027, followed by a deceleration, tells me something critical: Goldman is modeling a first wave of AI infrastructure saturation. They are pricing in the buildout of data centers, HBM capacity, and advanced logic. But they are also telegraphing that the second wave — the application layer, the edge devices, the autonomous agents — has not yet fully entered their model. This is where my 2022 forensic lens kicks in. During the Terra/Luna collapse, I calculated that $60 billion evaporated in 48 hours because algorithmic feedback loops, not ideology, broke the peg. The same mechanical reasoning applies here. The semiconductor equipment cycle is not a narrative. It is a hard constraint of physics, capital, and supply chains. The liquidity cascade is already flowing. The question is not whether the cycle is real — it is which balance sheets are positioned to absorb the shock when the inevitable 2028 deceleration arrives. Let me break down the technical architecture first, because the signals are hiding in plain sight. The forecast is not merely about expansion; it is a prediction of where the technology is heading. The three expansion areas — DRAM, HBM, and advanced process foundry — are not separate silos. They are an interconnected lattice of manufacturing complexity. DRAM process migration is moving from the 1-alpha/1-beta node to 1-gamma/1-delta. This is a 10-to-15-nanometer level shrink. Every node transition requires a step function in equipment spending, not a linear one. The precision required for the overlay, the etching, the deposition — it is all exponentially more demanding. The HBM technology curve is even more telling. The jump from HBM3E to HBM4, with the stack layers going from 8/12 to 16, is not a minor spec bump. This is a redesign of the entire thermal and signal integrity profile. And the advanced logic transition from 3nm to 2nm GAA (Gate-All-Around) — with the move from EUV single-patterning to high-NA EUV — represents a massive tooling upgrade. Goldman's forecast implicitly assumes these technology transitions will go smoothly. If the yield curve fails on high-NA EUV, the forecast is not just wrong — it is dead. A $300 million high-NA EUV machine is a capital allocation gamble, and if the tooling does not deliver on efficiency, the spending becomes repetition, not expansion. The confidence level here is 7/10. The physics are known, but the yield schedules are the hidden variable. Now, let me take a step back and look at the architecture of the supply chain. This is where the macro watcher's blood starts pumping. The equipment segment is the "pick-and-shovel" of the semiconductor world. It captures 10-15% of the entire industry's profit pool. But that is the surface-level stat. The deeper truth is in the bargaining power. The top equipment players — ASML, AMAT, Lam, TEL — are not just oligopolists. They are a chokehold. ASML has 100% monopoly on EUV. There is no alternative. This is not a supply chain; it is a tax on the entire global semiconductor industry. Downstream, the customer concentration is just as severe. The top ten fabs — TSMC, Samsung, Intel, SK Hynix, Micron, SMIC — account for 80% of equipment purchases. That gives the equipment manufacturers an enormous advantage. They have a stable customer base that is expanding at a rate of 36% annually. The upstream is concentrated, but the downstream is oligopolistic. The suppliers have the pricing power. The margins will remain at 45-50%+. This is not a healthy market. It is a machine. The supply chain vulnerability rating is medium-high. EUV is 100% dependent on ASML. The etching market is controlled by AMAT/Lam/TEL. The high-end photoresist comes from Japan, 90%. The silicon wafers are 60%+ from Shin-Etsu and SUMCO. These are not just supply chains. They are geopolitical leverage points. And the forecast is assuming that this fragile lattice will not break. Here is a hidden signal in Goldman's forecast that few will pick up: the confidence in the supply chain's resilience. To predict 2026-2028 high growth, Goldman implicitly assumes ASML's EUV capacity — roughly 50-60 units per year — will keep pace with demand. If the demand exceeds the projection, the delivery lead times will stretch, and the WFE spending will be a bottleneck. It is not a demand problem; it is a supply constraint. This is the same dynamic that is happening in HBM. I'm going to break down the capital expenditure in a way that matters to the balance sheet. The storage trio — SK Hynix, Samsung, and Micron — will exceed $80 billion in combined capex in 2025, representing 30-40% of revenue. TSMC's capex is $38-42 billion, 35-40% of revenue. This is a massive, coordinated investment phase. Now let's talk about the depreciation impact because this is where the liquidity trap is. Storage manufacturers typically use accelerated 5-year depreciation. Foundries use straight-line over 7 years. The massive expansion of 2025-2027 will release a concentrated depreciation hit in 2027-2029. This will squeeze storage gross margins by 5-10 points, and foundry margins by 3-5 points. The break-even calculation is brutal: storage manufacturers must maintain 85% utilization to cover the new depreciation; foundries need 80%. If the AI demand softens, the depreciation will act as a downward spiral on earnings. Here is my key insight. Goldman's forecast is not just about HBM, it is about a storage supercycle. The DRAM supply tension extending to 2028 is a structural change, not a temporary peak. This is the opposite of the 2017-2018 cycle. In 2017, the DRAM shortage lasted about two years. The current cycle is being driven by the AI data tsunami. The key logic is the HBM capacity squeeze. Let me break down the HBM math. HBM3E with 8-layer stacking consumes the equivalent of 3-4 times the wafer capacity of a standard DDR5. This is the core of the DRAM supply crunch. The AI server demand is not a blip; it is a massive shift in the memory landscape. As HBM production ramps, it eats up the wafer capacity that would otherwise go to standard DRAM. This is a permanent shift in supply elasticity. The market is not just buying more DRAM; it is re-routing its existing production capacity to a different, higher-value product. The second hidden insight is the "front-high, back-low" shape of the WFE spending. 2026's 36%, 2027's 45%, 2028's 29% — the peak in 2027 tells me Goldman believes the first wave of AI infrastructure investment will mature around 2028. The data centers, the training clusters, and the HBM fabs will be built. The next wave — the edge, the robots, the AI agents — is not yet in the model. This is the potential for a market gap. This brings us to the geopolitical layer. The forecast is essentially a bet on manageable geopolitical risk. The prediction of WFE growth until 2028 assumes that the US-China tech war will not break the global supply chain. The Chinese fabs — SMIC, Hua Hong, CXMT, YMTC — are part of the global WFE demand. But the export controls have already limited their access to advanced equipment. The Chinese market will shift to domestic alternatives. That shift will not help ASML, AMAT, or Lam. It will help Chinese domestic players like NAURA and AMEC. This is the contrarian angle. While the market sees a unified global growth, the structure shows a fragmented expansion. The US, Europe, Japan, and Korea will build their own capacity. China will build a separate, less advanced one. The cost of this decoupling is a 10-20% increase in long-term costs. The market is broken into two supply chains. The forecast is, in a sense, a prediction of that fragmentation. It is a bet that the advanced supply chain will keep expanding, while the Chinese one is built in parallel. The competitive landscape is clear. ASML has 100% EUV monopoly, Lam has 35% etching, AMAT has 35% deposition. The HBM market is SK Hynix with 50%+, Samsung with 30%, and Micron with 15%. The advanced foundry is TSMC with 90% of the 5nm and below. These players are the architecture of the machine economy. They are not just chipmakers; they are the enforcers of the new digital infrastructure. Now, the valuations. This is where the discipline of the balance sheet comes in. ASML trades at 35-40x P/E, AMAT at 25-30x, Lam at 25-30x. SK Hynix is at 15-20x. TSMC is at 25-30x. The equipment makers are priced for perfection. They are trading at historical high levels, driven by AI expectations. The storage players, on the other hand, are not. Their earnings improvement has not been fully priced in. This is the asymmetry in the market. From a value creation perspective, the equipment makers' ROIC is 25-35%, well above their WACC of 8-10%. They are value creators. The storage players' ROIC is 10-15%, which is above their WACC, but their current expansion cycle is a cash burn. The free cash flow is negative or barely positive. The leverage is rising. If the cycle turns, the storage companies are the most vulnerable. The hidden message is that the storage players have the most earnings elasticity. DRAM price increases and HBM mix improvements mean SK Hynix could hit a record net income in 2026. The equipment makers have high visibility, but their orders are already known. The real upside is in the memory side. Now, let's address the risk. The AI demand bubble is the biggest threat. If the commercial application of large language models fails to materialize, or if the cloud providers cut their capex guidance, the WFE spending could drop 30-50% below forecast. The trigger would be a missed NVIDIA earnings report. The probability of that is 20-30% in the next 12-24 months. The geopolitical risk is the second one. If there is a Taiwan conflict, the entire forecast is invalid. That is a 10-15% chance in three years. It is not the base case, but it is a tail risk. The third risk is the storage cycle peaking early. If the memory expansion comes faster than expected, the shortage will ease before 2026. That's a 25-35% probability. This is a viable risk because HBM yields are improving fast. The cycle is real. The capital is flowing. The liquidity is cascading. But the key is that the market is now pricing in a 2027 peak. As a macro watcher, I see this as a clear path. I see the technology signal and the market structure. My view is that the cycle will hold, but the base case is not in the consensus. The consensus is a linear extension of the past. My base case is a measured, volatile expansion, with a potential for a pause in 2028. The key is to be positioned not for the buildout, but for the transition. The buildout is the first wave, and the transition is the second. This is not a 2025 story. This is a 2026-2028 story. The real question is how the market will be positioned when the first wave of AI capex is fully built and the second wave — the machine economy, the autonomous agents, the edge intelligence — begins to emerge. The liquidity has a direction. The question is whether you are on the right side of the flow. The market is a machine, and the machine is being built. The question is who is going to own the machine.