The number landed on my terminal at 7:42 AM Shanghai time. Goldman Sachs had raised its wafer fab equipment (WFE) spending forecast to $218 billion for 2027 and $281 billion for 2028. The market read it as a semiconductor bull signal. I read it as a liability schedule for every AI-token thesis currently trading on hope.
Here is the balance sheet.
Context: The Equipment Cycle as Infrastructure Truth
WFE spending is not a sentiment indicator. It is a physical commitment of capital to produce silicon. Every dollar of WFE translates into wafer starts, which translates into chips, which translates into the GPUs, ASICs, and memory modules that underpin both the AI economy and the crypto infrastructure that parasitically attaches to it.
Goldman's forecast implies a compound annual growth rate of roughly 20% from 2024 through 2028. That is not a modest assumption. It requires AI-related chip demand to remain in a hypergrowth phase for four consecutive years. It requires the hyperscalers—Meta, Google, Microsoft, Amazon—to sustain their AI capital expenditure programs without a single year of digestion. It requires HBM4 yield ramps to proceed without material delay. And it requires the equipment supply chain, currently constrained by ASML's annual EUV output of approximately 50-60 units, to somehow deliver.
The ledger does not lie, only the interpreters do.
Core: The Technical Teardown
Let me dissect this forecast the way I would dissect a smart contract audit. Layer by layer. Assumption by assumption. And then map each layer to the crypto narratives that depend on it.
Layer One: The HBM/CoWoS Bottleneck
The single most important variable in Goldman's forecast is not lithography. It is advanced packaging. High Bandwidth Memory—HBM3E in production today, HBM4 slated for the second half of 2025—requires TSV (through-silicon via) etching, hybrid bonding, and CoWoS packaging capacity. TSMC's CoWoS capacity was approximately 400,000 wafers per year (12-inch equivalent) in 2024. The company plans to double that in 2025. It will still be insufficient.
This is the bottleneck that matters for crypto. Every GPU that NVIDIA ships—H100, H200, B200—requires CoWoS packaging and HBM stacks. The AI-token ecosystem—Render, Akash, Bittensor, and the dozens of smaller DePIN projects—is priced on the assumption that GPU supply will expand to meet demand. That assumption runs directly into the CoWoS wall.
My audit experience tells me to check the dependency chain. In 2018, I reviewed the 0x Protocol v2 contracts and found three critical logic flaws in the signature verification process that previous auditors had missed. The lesson was simple: the most obvious dependency is rarely the one that breaks. For AI tokens, the obvious dependency is GPU demand. The hidden dependency is CoWoS packaging capacity. If TSMC cannot double CoWoS fast enough—and the equipment delivery cycle for advanced packaging tools runs 12-18 months—then GPU supply growth stalls, and every AI-token valuation model that assumes linear GPU expansion is built on a false premise.
The yield variable compounds the problem. SK Hynix's HBM3E yield is approximately 70-80%. TSMC's 3nm yield has improved to 80%+, but Samsung's 3nm GAA yield remains in the 60-70% range. Yield directly determines equipment purchasing rhythm. The faster yields ramp, the more urgent the expansion demand. Goldman's forecast implicitly assumes that 2nm and HBM4 yields will ramp quickly in 2026-2027. If yields disappoint, equipment spending is delayed, not cancelled—but the 2027-2028 figures of $218 billion and $281 billion carry meaningful downside risk.
Layer Two: The 2nm Transition and Equipment Delivery
Goldman's forecast implicitly assumes that 2nm GAA (gate-all-around) production ramps in 2025-2026 and that HBM4 follows the same trajectory. TSMC's N2, Samsung's 2nm GAA, and Intel's 18A are all scheduled for production in the 2024-2026 window. Each requires high-NA EUV lithography—the 0.55 numerical aperture systems that cost over $300 million per unit.
ASML's annual EUV production capacity is approximately 50-60 units. High-NA EUV is even more constrained. The delivery pipeline is the constraint. A fab that orders high-NA EUV today waits 12-18 months for delivery, then spends another 12-24 months on installation and yield ramp. The full cycle from equipment order to volume production is 24-36 months.
The capital intensity math is unforgiving. The equipment investment density per 10,000 wafers of monthly capacity at 2nm and below is more than 50% higher than at 5nm. This is not incremental. It is a step change in the cost structure of leading-edge semiconductor manufacturing. Every advanced fab built in the 2024-2026 window carries a capital burden that previous generations did not.
This matters for crypto because the mining industry and the AI-infrastructure sector are both downstream of the same equipment pipeline. Bitcoin ASIC miners compete with AI chips for the same wafer capacity at TSMC and Samsung. When AI demand absorbs leading-edge capacity, mining ASICs get pushed to mature nodes or face longer lead times. The 2021 mining bull run was partly enabled by excess leading-edge capacity. That excess no longer exists. The equipment cycle is now structurally tight, and mining hardware supply will remain constrained through at least 2026.
Layer Three: The Storage Supercycle
Goldman's forecast identifies DRAM and HBM as primary drivers of WFE growth through 2028. The implication is that DRAM supply remains tight for the next four years. This is a structural claim, not a cyclical one. AI's demand for HBM is a new consumption category that did not exist in previous memory cycles. Every HBM stack consumes roughly three times the wafer area of a conventional DRAM die. The math is unforgiving.
The capacity expansion plans confirm the direction. SK Hynix's Yongin cluster represents a $90 billion+ investment in DRAM and HBM expansion, with production starting in 2027. Micron's New York and Idaho fabs represent $100 billion+ in planned investment for DRAM and HBM, with production in the 2026-2028 window. These are not speculative projects. They are committed capital expenditures backed by confirmed customer demand from the hyperscalers and GPU vendors.
For crypto, the storage supercycle has a specific implication: the cost of running blockchain infrastructure is about to increase. Validator nodes, archive nodes, and data-availability layers all require storage. If DRAM prices enter a sustained uptrend—and Goldman's forecast implies exactly that—the operating cost of node infrastructure rises. Projects that budgeted for storage costs based on 2023-2024 prices will face margin compression.
Trust is a bug, not a feature. But storage is a cost, and costs are real.
The equipment demand structure is also shifting. Memory manufacturing requires a different equipment mix than logic. Etch tools from Lam Research, deposition tools from TEL and AMAT, and advanced packaging tools from Besi and ASM International will see disproportionate demand growth. The equipment competitive landscape is not static. The HBM buildout is creating new winners in the back-end packaging segment—a shift that the market has not fully priced.
Layer Four: The China Variable
The report notes that China's equipment localization rate is approximately 20-30% for mature nodes and below 5% for advanced nodes. The gap is structural. EUV lithography is 100% dependent on ASML, with no alternative source. Advanced DUV is 90%+ dependent on ASML, Nikon, and Canon. The export controls imposed by the United States, followed by the Netherlands and Japan, have created a permanent ceiling on China's advanced-node ambitions.
The supply chain vulnerability assessment is sobering. For EUV lithography, import dependence is 100% with no substitute. For immersion DUV, dependence exceeds 90%. For etch equipment, dependence is approximately 70% on Lam Research, AMAT, and TEL. For thin-film deposition, 75% dependence on AMAT, TEL, and Lam. For high-end photoresist, over 90% dependence on Japanese suppliers—JSR, Shin-Etsu, Tokyo Ohka. For 12-inch silicon wafers, over 80% dependence on Shin-Etsu, SUMCO, and SK Siltron. The only categories with meaningful domestic alternatives are specialty gases and mature-node etch/deposition tools.
But here is the hidden variable: China's mature-node expansion is proceeding regardless. SMIC, Hua Hong, and the memory makers are adding 28nm and above capacity at scale. The third phase of the National Integrated Circuit Industry Investment Fund—344 billion yuan, established in 2024—is specifically directed at equipment, materials, and EDA localization. This creates a two-tier market. Advanced nodes remain tight, driven by AI demand. Mature nodes face oversupply, driven by Chinese capacity additions. The price pressure in mature nodes will compress margins for any crypto project that relies on commodity silicon—which includes most mining operations outside the leading-edge ASIC segment.
The geopolitical risk is the largest unquantified variable in Goldman's forecast. If export controls tighten further—if the United States restricts mature-node equipment, for example—the global equipment supply chain is disrupted. ASML derives approximately 15% of its revenue from China. A full decoupling scenario would reduce global WFE efficiency and raise costs by an estimated 20-30%. Crypto infrastructure would feel this through higher hardware costs and longer delivery times.
The onshoring trend compounds the demand picture. The United States CHIPS Act ($52 billion) aims for 20% of global advanced-node capacity by 2030. The European Chips Act (€43 billion) targets 20% global share. Japan's semiconductor revival plan commits over ¥2 trillion. Each region building its own fabs means duplicate equipment purchases. This is a structural demand increment that did not exist in previous cycles. Goldman's forecast partially captures this, but the full effect of geopolitical fragmentation on equipment demand is likely understated.
Layer Five: The Depreciation Cliff
Here is the number that most market participants are not modeling. Semiconductor equipment is depreciated on a straight-line basis over 5-7 years. The fabs that begin construction in 2024-2025 will reach volume production in 2026-2028. That is precisely when Goldman's WFE forecast peaks. The depreciation expense from these new fabs will suppress gross margins at TSMC by an estimated 3-5 percentage points—from approximately 55% to 50-52%. Memory manufacturers will feel it more acutely.
The breakeven utilization rate for an advanced-node fab is 70-80%. If demand softens in 2029-2030—which is the historical pattern after every equipment spending surge—the industry faces a capacity glut. The 2010s memory expansion cycle followed exactly this trajectory. The semiconductor industry has a documented history of herding behavior. Goldman's forecast, if fully realized, sets up the conditions for the next oversupply crisis.
The financial metrics of the equipment oligopoly tell a consistent story. ASML's gross margin runs 50-55%, with R&D intensity of 15-18% and operating cash flow of $80-90 billion annually. AMAT runs 45-48% gross margin with $70-80 billion in operating cash flow. KLA leads the group at 60%+ gross margin. Return on invested capital for ASML is 30-35%, versus a weighted average cost of capital of approximately 8-10%. These are exceptional value-creation metrics. They also explain why the equipment stocks trade at 25-40x trailing earnings—the market has already priced in a substantial portion of Goldman's forecast.
For crypto, the depreciation cliff has a delayed but inevitable effect. The hardware that mining companies and DePIN networks purchase today will be worth a fraction of its purchase price in 2028-2029. Projects that financed hardware purchases on the assumption of stable resale values are carrying unmarked liabilities. The ledger does not lie, only the interpreters do.
Layer Six: What This Means for Crypto's AI Narrative
The AI-token sector is currently trading on a narrative that GPU supply will expand to meet demand. That narrative is now quantifiable. Goldman's WFE forecast implies a specific trajectory of GPU availability. If the forecast is accurate, GPU supply grows at roughly 20% annually through 2028. If it is inaccurate—if the equipment supply chain fails to deliver, or if AI demand peaks earlier than expected—GPU supply growth slows, and every AI-token valuation model must be revised downward.
The more interesting implication is structural. The AI-token thesis is not merely about GPU supply. It is about the marginal cost of compute. As the equipment cycle matures and depreciation begins to bite, the marginal cost of compute rises. This favors projects that have secured long-term compute contracts at fixed prices. It punishes projects that rely on spot-market GPU rental. The difference between these two business models is the difference between a fixed-rate mortgage and a variable-rate loan. In a rising rate environment, the variable-rate borrower gets liquidated.
The competitive landscape of the semiconductor industry reinforces the concentration risk. ASML holds over 80% of the lithography market and 100% of EUV. The etch market is a three-way split—Lam Research at approximately 30%, AMAT at 25%, TEL at 25%. Deposition is similarly concentrated—AMAT at 35%, TEL at 25%, Lam at 20%. Inspection is a duopoly—KLA at 50%, AMAT at 25%. This is not a competitive market. It is an oligopoly with pricing power that has only strengthened through successive cycles. The customer base is equally concentrated—the top five customers (TSMC, Samsung, SK Hynix, Micron, Intel) account for 50-60% of equipment vendor revenue.
The new-entrant threat is real but distant. Chinese equipment vendors—AMEC, NAURA, Piotech—are making progress in mature-node etch, deposition, and cleaning. But the advanced-node gap is measured in decades, not years. The equipment industry's barriers—technology, customer qualification cycles, patents—require 10-15 years for a new entrant to become competitive. The oligopoly is secure through the forecast window.
Contrarian: What the Bulls Get Right
I have spent this analysis dissecting the assumptions in Goldman's forecast. Intellectual honesty requires me to acknowledge what the bulls get right.
The storage supercycle is real. AI's demand for HBM is a structural addition to the memory market, not a cyclical blip. The wafer area consumed by HBM is approximately three times that of conventional DRAM. This is not a demand shock that will reverse. It is a permanent shift in the consumption function. Memory manufacturers—Samsung, SK Hynix, Micron—are entering a multi-year upcycle that their current valuations do not fully reflect. The current valuations of memory stocks, at 10-15x earnings, do not price the earnings elasticity of a supply-constrained upcycle.
The AI demand is also more durable than skeptics assume. The hyperscalers are not making speculative bets. They are building infrastructure for revenue-generating services. The inference demand—not just training—is growing at 50%+ annually. This is not a bubble in the 2021 sense. It is a capital expenditure cycle backed by actual revenue. The application mix is diversifying—HPC and AI training at 25-30% of semiconductor revenue with 40%+ growth, AI inference at 10-15% with 50%+ growth, automotive electronics at 10-12% with 10-15% growth. The demand base is broader than any single application.
And the equipment cycle itself is underappreciated. The transition to high-NA EUV, the expansion of advanced packaging, and the storage buildout represent a multi-year investment supercycle. The equipment oligopoly has pricing power that was absent in previous cycles. Their backlogs are at record levels. The order-to-revenue ratio of approximately 1.5-2.0 provides visibility that the market is not fully pricing. The equipment vendors' cash flow generation—OCF/net income ratios above 1.0 across the group—supports continued capital returns to shareholders.
History repeats, but the gas fees change. The pattern is the same—capital expenditure supercycles followed by capacity gluts—but the specific technology drivers are different. AI and HBM are not the same as the smartphone cycle of the 2010s. The demand is more concentrated, the capital intensity is higher, and the geopolitical overlay is more complex. The cycle may be longer than historical precedents because the demand drivers are structural rather than cyclical.
Takeaway: The Accountability Call
Goldman's forecast is not a prediction. It is a scenario. The scenario requires AI demand to remain in hypergrowth, equipment supply chains to deliver on schedule, and no major geopolitical disruption. Each of these assumptions has a non-trivial failure probability. The market is currently pricing the scenario as if it were certain.
The crypto projects that will survive this cycle are those that have modeled the downside. Projects with fixed-price compute contracts. Projects with hardware that is already depreciated. Projects that do not need to raise capital in 2027-2028 to fund equipment purchases at peak prices.
The projects that will fail are those that have built their business models on the assumption that GPU supply expands linearly, that storage costs remain flat, and that the semiconductor industry has somehow escaped its historical cycle of boom and bust.
Code is law; intent is irrelevant. The equipment cycle is the law. The intent of every AI-token whitepaper is irrelevant to the physical reality of wafer starts, packaging capacity, and depreciation schedules.
Verify the hash. Ignore the hype. The silicon ledger is the only ledger that matters.