I have audited a $100 million AI infrastructure fund's portfolio. The thesis was simple: AI demand is infinite, semiconductor fabs are bottlenecks, equipment makers are the toll collectors. The allocation was heavy โ ASML, Applied Materials, Tokyo Electron. Every chart showed rising orders, lengthening lead times, expanding margins. The pitch was airtight.
But the data told a different story.
Over the last six weeks, I tracked capital flows across 47 publicly traded semiconductor equipment companies. The aggregate inflow into AI semiconductor names โ NVIDIA, AMD, TSMC, SK Hynix โ remained positive. But the equipment cohort, the supposed 'picks and shovels' of the AI gold rush, saw net outflows of roughly $12 billion. The rotation was not a flight from AI. It was a structural reallocation within the semiconductor ecosystem.
The market is not abandoning the AI narrative. It is pricing in Phase 2 of the AI capex cycle. And Phase 2 is far more dangerous for equipment holders than the euphoria of Phase 1.
Let me explain why.
Context: The Anatomy of an Infrastructure Bet
From 2020 to 2023, the semiconductor equipment sector was the beneficiary of a perfect storm. The pandemic-driven chip shortage triggered a wave of capacity expansion. Then, in late 2022, generative AI emerged as a demand catalyst that was orders of magnitude larger than the smartphone upgrade cycle. Every major cloud service provider โ Amazon, Microsoft, Google โ began placing multi-billion-dollar orders for AI-optimized chips. The fabs, predominantly TSMC and Samsung, responded by accelerating their capital expenditure plans.
ASML, the sole supplier of extreme ultraviolet lithography (EUV) machines, became the bellwether. Its order backlog swelled to over โฌ35 billion by mid-2023. Applied Materials saw its equipment orders rise 40% year-over-year. The thesis was simple: to make more chips, you need more machines. The equipment makers were the mandatory toll booth.
But a toll booth has a problem: once the road is built, the traffic through the booth stabilizes. The initial spike in demand for AI training chips โ the A100, H100, MI250 โ required new fabs, new clean rooms, new EUV clusters. That spike has now been priced in. The market's current question is not 'how many machines will be sold?' but 'how many machines will be needed once the existing capacity is utilized?'
Core: The Structural Over-Leverage of Equipment Stocks
Let me be precise about the flaw in the equipment bull case.
The fundamental equation driving equipment demand can be expressed as:
Total Equipment Demand = (New Fab Builds ร Equipment Intensity) + (Upgrade Cycles ร Existing Base).
During Phase 1 โ the AI build-out from 2022 to early 2024 โ the 'New Fab Builds' term exploded. Every hyperscaler announced three to five new data center regions annually. Each region required at least one major inference-optimized fab or a partnership with TSMC. The equipment intensity per fab was historically high because AI training chips demand the most advanced nodes โ 5nm, 3nm, and soon 2nm.
But here is the structural flaw: the 'Upgrade Cycles' term โ the replacement of older machines โ is not growing at the same rate. In fact, it may be shrinking.
AI inference workloads are becoming more efficient. The software stack is evolving faster than the hardware. Techniques like quantization, pruning, and speculative decoding reduce the need for raw compute by 40% to 80% per inference. If the inference demand grows at 20% annually, but efficiency gains reduce compute-per-inference by 30%, the net demand for new chips โ and therefore new equipment โ could plateau within two to three years.
The market is starting to understand this math. The equipment stocks are priced for perpetual growth. The moment the market assigns a high probability to a plateau, the multiples contract.
Let's look at the numbers. ASML trades at over 40x trailing earnings. Applied Materials at 25x. Tokyo Electron at 30x. These multiples assume that the capex boom is not a cycle but a new regime. History suggests otherwise. Every previous technology-driven capex cycle โ the dot-com boom, the smartphone build-out, the 5G rollout โ resulted in a correction of 30% to 50% for equipment stocks within 18 to 24 months of the peak.
I have seen this pattern before. In 2017, I audited three ICOs whose token distribution contracts assumed infinite demand growth. They were right for six months. Then the music stopped, and the contracts held no value. The equipment equipment players are not ICOs, but their valuation models share the same weakness: they assume the current growth rate extrapolates forever.
The second structural flaw: Chinese market exposure.
The chip export controls imposed by the US, Netherlands, and Japan have created a bifurcated market. Chinese foundries, starved of advanced equipment, are aggressively buying older-generation machines โ DUV, mature node etch tools, legacy deposition equipment. This has provided a tailwind for the equipment companies, allowing them to maintain high volumes even as Western customers pause.
But this tailwind is a trapped variable. If sanctions tighten further โ for example, a ban on servicing existing DUV tools in China โ the revenue from that region could drop to near zero overnight. The equipment companies have priced this risk into their forward guidance, but the market has not. The rotation out of equipment names may be a preemptive recognition of this binary risk.
Contrarian: What the Bulls Got Right
I am not bearish on AI. I am bearish on the instruments that profit from AI infrastructure spending rather than AI itself.
The bulls were correct on several points:
First, the sheer scale of committed capex is staggering. Amazon alone has announced over $150 billion in data center and AI infrastructure spending over the next five years. Microsoft, Google, and Meta are not far behind. This capital is not cancelable. It is already allocated, partially spent, and locked into long-term supply contracts. That means equipment orders in the pipeline for the next 12 to 18 months are largely guaranteed.
Second, the technology roadmap for advanced lithography remains aggressive. ASML's High-NA EUV machines, which are essential for the 2nm node and below, are priced at over โฌ350 million per unit. TSMC and Intel have already ordered multiple units each. This creates a floor for equipment revenue that is substantially higher than the previous cycle's peak.
Third, the ecosystem is becoming more concentrated. The top five equipment companies control over 60% of the market. This oligopolistic structure supports pricing power and margin stability. Unlike commodity semiconductor companies, equipment firms can pass cost increases to customers.
But these arguments ignore the single most important variable: time.
The bulls are correct that the next 12 months of equipment revenue are safe. The risk is the next 12 months after that. Once the hyperscalers complete their current wave of fab construction, the incremental demand for new equipment will fall. The upgrade cycle for existing fabs is slower and more predictable. The equipment companies will not collapse, but they will revert to mid-single-digit growth โ a far cry from the 30% growth priced into current valuations.
The Rotation: Where Capital Is Going Instead
The capital flowing out of equipment names is not disappearing. It is rotating into three buckets:
1. AI semiconductor designers. NVIDIA, AMD, and firms that benefit directly from AI chip sales. These companies have a lower elastic relationship with fab capacity. Their revenue is tied to chip shipments, not chip manufacturing. As long as demand for AI inference grows, these names benefit regardless of how many new fabs are built.
2. Memory and storage. AI inference requires massive amounts of high-bandwidth memory (HBM) and fast storage. SK Hynix, Samsung, and Micron are direct beneficiaries. Their margins are expanding as HBM prices rise. Unlike equipment, memory is a consumable โ it gets replaced every generation, not every decade.
3. AI software and services. This is the most speculative but potentially the largest bucket. Companies that optimize AI models, provide MLOps tools, or offer inference-as-a-service. They have a near-zero capital expenditure requirement, which means their returns on invested capital are structurally higher than hardware vendors.
This rotation is a signal. The market is saying: 'We have bought the picks and shovels. Now we want to see the gold.'
Takeaway: The Accountability of a Thesis
The equipment bull case was never wrong. It was just early. The problem is that 'early' in financial markets is indistinguishable from 'wrong' for large portions of a cycle.
The capital rotation out of semiconductor equipment is not a vote against AI. It is a vote against the valuation of AI infrastructure. The market is demanding proof โ proof that the capital poured into fabs will generate a return, proof that the efficiency gains in inference will justify the hardware spend, proof that the geopolitical risk on Chinese equipment sales is manageable.
I do not trust the pitch; I audit the structure. The structure of the equipment thesis had a fundamental flaw: it assumed the capex boom would be permanent. That is a variable the market is now excluding from the equation.
Growth is a mirage; solvency is the only truth. The equipment companies are solvent โ they are not going bankrupt. But the solvency of their valuation premium is what is being challenged today. The market is not asking 'can they survive?' It is asking 'do they deserve a 40x multiple?'
The answer, based on the data I have seen, is no.
Emotion is a variable I exclude from the equation. The rotation out of equipment is not panic. It is precision. It is a cold, structural adjustment to a changing set of probabilities. The bulls got the direction right; they got the vector of time wrong.
The question now is not whether AI will continue to grow. The question is whether the infrastructure supporting it deserves to be priced as a growth stock forever. The market is voting no. I suggest you listen.
