Goldman Sachs Bets $281B on Silicon. Crypto Pays the Price.

The number landed without fanfare. $281 billion. That is the global wafer fab equipment (WFE) spend Goldman Sachs projects for 2028, up from a projected $149 billion in 2025. A 36% compound annual growth rate. In an industry that has historically cycled between feast and famine, this is not a projection. It is a declaration that the AI buildout will bend the physical world to its will. Code doesn't care about supply chains, but the chips it runs on absolutely do. And for crypto, which often pretends to live in a purely digital realm, this is the most important story you are not reading about. This is not a story about mining ASICs, though those will be affected. This is a story about the backbone. Every validator, every GPU-based oracle network, every Layer-2 sequencer that dreams of AI integration—they all consume compute. That compute is built on wafers. And those wafers are about to become the most strategically contested resource on the planet. Based on my experience auditing 2017 ICOs, where the gap between whitepaper claims and technical utility was a chasm, I can tell you the gap between AI revenue projections and semiconductor reality is just as wide. But unlike a token whitepaper, the physical constraints here are absolute. You cannot fork a lithography machine. The Goldman thesis is straightforward. AI training demands. HBM. Advanced process nodes. All of it adds up to a multi-year boom for ASML, Applied Materials, Lam Research, and Tokyo Electron. The logic is sound. The assumption is the risk. The assumption is that AI capex sustains a 40%+ growth rate through 2028. That is a massive bet. From my 2020 analysis of yield farms, where I built spreadsheets to track token emissions against real revenue, I learned that unsustainable growth rates eventually meet the mathematics of reality. AI capex is not a token emission schedule, but the principle holds. The music can stop. But for now, the cycle is undeniable. Let's break down the mechanics. The DRAM and HBM story is the engine. Goldman names memory as the first growth driver. This is a critical detail. Logic foundries are also expanding, but memory is where the equipment spend intensity is exploding. An HBM3E stack consumes 3-4x the DRAM die area of a standard DDR5 module. HBM4 moves to hybrid bonding, requiring even more precision. This is not an incremental expansion. It is a full-scale rebuilding of the memory ecosystem. SK Hynix, Samsung, and Micron are on track to pour over $50 billion into HBM related capex from 2025 to 2027. Each dollar of that capex flows directly into the revenue lines of the equipment oligopoly. Then you have the process node race. The 2026-2028 window is the critical ramp for TSMC's N2 (2nm GAA), Intel's 18A/14A, and Samsung's 2nm GAA. These are not simple nodes. They are the first large-scale production of Gate-All-Around architecture. The yield curve is painful. Reaching profitable yields requires massive over-investment in process control, metrology, and inspection. KLA, the metrology monopoly, will see a direct benefit. ASML, with High-NA EUV, is the toll booth for the entire industry. Each of these machines carries a price tag of 300-400 million euros. They are not just equipment. They are a national investment. The transition to High-NA is not optional for the leading edge. It is mandatory. And ASML's annual capacity remains at only 50-60 EUV units. That is the bottleneck. Goldman's forecast implicitly assumes High-NA EUV will be deployed in volume by 2026-2027. If ASML's delivery slips, the entire 2028 forecast becomes the target. This is not a minor risk. This is the supply chain underwriting the bull case for every AI token and every infrastructure protocol. The machine cannot be printed. It cannot be deployed. It must be delivered, installed, and calibrated. That process takes 12-18 months for the order, and a year for installation and yield. The timeline is brutal. The second structural issue is the shift in bargaining power. The WFE market is a seller's market. ASML has a gross margin of 50%+, KLA has 60%+. The customer base is concentrated—TSMC, Samsung, Intel, SK Hynix, Micron—and the top 10 customers account for 60-80% of revenue. Yet, these customers are at the mercy of the suppliers. This is the best position in the entire semiconductor value chain. The "picks and shovels" are the only game in town. As demand outstrips supply, expect equipment makers to raise prices 5-10% per year. This will not be a drag on their margins. It will be a tailwind. Now, let's look at the blind spots. The Goldman forecast is, in my view, underweight on China. Chinese WFE spend is about 20-25% of the global total. The forecast implicitly assumes that this will not change significantly. But the National Big Fund Phase III is a 344 billion RMB (about $48 billion) commitment to domestic equipment and materials. This is a forced upgrade cycle. Chinese fabs will continue to build out mature process nodes. And domestic equipment makers like Naura (北方华创), AMEC (中微公司), and ACM Research (拓荆科技) are moving up the curve. In 2026-2028, I expect them to achieve 30-50% annual growth in revenue, driven by the dual engine of policy support and the need for supply chain resilience. This is not just a China story. The geopolitical overlay is the biggest risk to the Goldman forecast. The US export controls on advanced process equipment are already strict. But the recent trend is moving to "mature" node controls. If the US limits equipment to China even at 28nm, the global WFE forecast will be significantly lower. But it would also accelerate the domestic substitution effort. This is a lose-lose scenario for the global efficiency. The technology decoupling will create 10-15% inefficiency due to duplicate fabs and fragmented supply chains. But it will create a new world. The "Global Semiconductor" is dead. My prediction is that the real state is not whether the $281 billion will be spent, but how much of it will be spent on redundant capacity. The US CHIPS Act, EU Chip Act, and Japan's subsidy are all aimed at creating domestic capacity. This is a positive for WFE in the short term. But it is a huge negative for the long-term ROIC of the industry. I have seen this in 2022 with the Terra/Luna collapse—the fragility of the ecosystem is in the dependencies. The current crypto ecosystem is heavily dependent on AI narratives. The AI narrative is heavily dependent on the physical ability to produce chips. Let me talk about the memory cycle. The Goldman forecast has DRAM supply tightness through 2028. That is a long time. In my experience, memory cycles are 3-4 years. This bull cycle started in 2024. It should peak in 2027-2028. The forecast of a persistent shortage suggests that HBM demand will be much higher than current estimates. This is the same logic as the DeFi yield analysis. You must check the underlying emissions. In this case, the "emission" is the HBM demand from AI accelerators. If NVIDIA's B200 (and future Rubin) and AMD's MI400 series continue to sell out, the demand for HBM will be insatiable. But if the AI bubble bursts, the DRAM market will have a massive supply glut. The same equipment that is now a gold mine will become a liability. The depreciation cost is massive. The depreciation is a key factor. Fabs are taking on billions in capex. The depreciation schedule is typically 5-7 years. This will pressure margins. I expect 2-4 percentage points of gross margin drag for the fabs in 2026-2028. But AI chip pricing power can offset that. NVIDIA can charge $30-40k per GPU. That is a high pricing power. However, if the AI market consolidates and competition intensifies, pricing will drop. And the cycle will break. The bottom line is: the goldman forecast is a map of the future. But the road is not paved. It is a mud. The equipment supply chain is the first and most critical risk. The delivery times for Lam and AMAT are already 12-18 months. This will not get better. It will get worse. The only way to mitigate this is to increase the prices. And they will. What does this mean for the crypto? The next few years will see a "pay-to-play" dynamic for AI-centric crypto projects. If you are building a decentralized training network or an AI oracle that requires high compute, you are not just competing with other blockchain projects. You are competing with OpenAI and Microsoft for the same wafers. The cost of compute is not going down. It is going up. The bull market for AI is a bull market for the physical infrastructure. And the physical infrastructure is owned by a very small cartel. The "decentralized" aspect of crypto is an illusion if you are dependent on centralized hardware. This is my biggest concern for the next cycle. The crypto industry is moving from a software-defined asset to a hardware-constrained asset. The future is not just code. It is silicon. This is the moment to watch the ASML order book and the SK hynix capex cycle. The signals are in the backlog. The market is currently paying for the AI narrative. But the price of the narrative will be written in the order books of the equipment makers. I will be watching the capex / revenue ratio of the memory makers. If it goes above 40%, it is a sign of over-building. The 2028 forecast is a peak. The following year will be a slowdown. The equipment makers will be the first to see it. In my 2021 NFT analysis, I found that the smart contract code was often a trap for the uninitiated. The same is true for the WFE cycle. The trap is the assumption that the growth is linear. It is not. It is cyclical. The question is not whether the WFE will go up. It is when it will go down. The 2028 peak is the warning. The time to prepare is now. Code doesn't. Physics does. And physics is the most reliable oracle in the world.