ASML's strong orders ease fears AI spending boom is running out of steam
Record orders for ASML’s chipmaking machines are being interpreted by some as a sign that the current AI-driven semiconductor boom will continue for years, since fabs only recoup such investments over long horizons. Others counter that long lead times, cancellable contracts, and FOMO-driven overordering make ASML’s backlog a poor gauge of sustainable AI demand, drawing parallels to the dot-com and telecom bubbles where infrastructure spending far overshot near-term revenue. There is broad agreement that demand for compute and chips will persist, but sharp disagreement over whether today’s data center and GPU build-out reflects durable economics or an overheated cycle that will eventually correct.
ASML Orders as Indicator of AI Demand
- Many argue strong ASML orders mainly show chipmakers expect long-term demand for advanced nodes, not specifically that today’s AI economics are healthy.
- Several note ASML sells tools usable for many chip types (CPUs, GPUs, memory, consumer chips), so orders don’t uniquely validate AI.
- Some call ASML a lagging signal: by the time ASML bookings slow, the AI/datacenter bust will already be in progress.
Overordering, Contracts, and “Real Options”
- Fabs are ordering tools out to 2029–2030; some see this as overordering driven by FOMO and capacity fears.
- Debate over cancellation: some say long-dated orders are effectively “slots” that can be canceled or delayed; others stress there are contractual penalties, at least close to build time.
- Multiple comments frame these orders as “real options” or OTM call options on future demand: limited downside vs huge opportunity cost if you underbuild.
AI Demand, Profitability, and Subsidies
- Current AI compute demand is described as very high and capacity-constrained; frontier providers are at or near limits.
- Strong disagreement on whether usage is “subsidized”:
- One side says if firms are burning billions and not profitable, something (tokens or infra) is implicitly subsidized.
- Others say high R&D burn isn’t the same as subsidizing inference, and headline API prices reflect competition more than cost.
Bubble vs Sustainable Growth
- Many expect an AI bubble/correction regardless of how real the tech is, citing overinvestment, FOMO, and historical patterns.
- Others argue stock gains track earnings so far and believe AI still has large untapped economic value.
- Some think the bust will come when detailed financials (e.g., S‑1s) reveal thin margins and unsustainable capex.
Hardware vs Model Layer
- Several distinguish between:
- Model/token layer, likely to become highly competitive with squeezed margins.
- Hardware layer (GPUs, foundries) where current profits are concentrated.
- Some predict future competition from many inference-chip vendors will erode Nvidia’s pricing power, shifting value toward bottlenecks like TSMC.
- Concerns raised about highly leveraged AI infra players (e.g., CoreWeave bond yields) as early stress signals.
Historical Analogies and Limits
- Frequent comparisons to the dot-com and telecom bubbles: real underlying tech, but huge overbuild and long, painful corrections.
- Key differences noted: fiber had multi-decade life; GPUs are short-lived “consumables,” risking more capital destruction.
- Several stress Amdahl’s Law and non-automatable bottlenecks (coordination, robotics, regulation) as reasons AI may not deliver economy-wide returns fast enough to justify current bets.