2026-08-08 · Gold Rush · markdown version
Looking back from December 2025, the thing everyone missed in August was that Amazon's quiet replacement of twenty-month-old Trainium 2 chips wasn't just a product refresh—it was a depreciation bomb. The hyperscaler capex boom had been framed as a rising tide lifting all AI infrastructure boats: chip designers, memory makers, data center REITs, GPU suppliers all riding the $220 billion secular spend wave. But the custom silicon cycle turned out to have a half-life closer to industrial fashion than industrial equipment. In this branch of the distribution, the market spent the autumn repricing what a two-year useful life does to unit economics when the chips cost what chips cost. The consensus had assumed Moore's Law depreciation schedules; the realised path was closer to smartphone obsolescence.
The August 2024 consensus read was straightforward: hyperscaler capital expenditure on AI infrastructure represented a multi-year secular tailwind, and exposure to any part of the stack—NVIDIA's GPUs, Micron's high-bandwidth memory, Equinix's data center REITs, TSMC's foundry capacity—captured a slice of that tailwind. The $220 billion figure floated by analysts assumed that chip useful lives would follow historical semiconductor norms, with depreciation spread over four to seven years and replacement cycles governed by performance improvements, not obsolescence. The narrative was that AI workloads were sticky, architectures were converging, and the capex wave would compound as models scaled. Every earnings call in the sector referenced the same McKinsey slide deck. The distribution of outcomes, in this view, had a fat right tail and a narrow left tail: the boom might be bigger or smaller, but it was a boom.
The contrarian read—one branch among many—was that Amazon's Trainium 2 replacement after twenty months was not an anomaly but a signal. Custom silicon optimised for specific AI workloads could have a much shorter useful life than the market had priced, because the workloads themselves were moving targets and architectural bets were cheap to abandon when you control the full stack. If custom ASICs depreciate in two years instead of five, the $220 billion hyperscaler spend starts to look worse on a unit-economics basis: more capex per unit of durable compute, compressed margins, and a treadmill that accelerates rather than slows. The second-order read was that this depreciation trap—if it materialised—would force hyperscalers to run fabs harder, not slower, because the only way to justify the spend was to maintain utilisation and hope for workload growth to outrun obsolescence. Faster write-downs could mean more wafer starts, more fab throughput, more cleanroom hours. And every additional wafer start, regardless of whether the chip architecture wins or loses, consumes ultrapure nitrogen, argon, helium, and specialty gases supplied by an oligopoly: Linde, Air Products, Air Liquide, Nippon Sanso, Messer. The crazy read was that the real toll-road wasn't the chip designers—it was the industrial gas suppliers who collect revenue on every fab generation, every process node, every architecture pivot. Linde and Air Products were the two largest players in the Western fab supply chain, and the read was that their pricing power and contract structures insulated them from chip obsolescence risk in a way that semiconductor equities were not. The Play fed into the Hoard Engine modeled a ten-percent reallocation from reference equities into commodities exposure, a tilt toward toll-road industrials over chip hype.
Across 10,000 bootstrap paths (seed 921538), this play moves the reference hoard's 15-year median from $551,192 to $543,392 ($-7,800), and its goal probability from 6.4% to 5.2%. The simulation shows the median terminal outcome falling, the fifth percentile rising from $283,721 to $294,426, and the ninety-fifth percentile falling from $1,048,249 to $1,008,777. The maximum drawdown at the ninety-fifth percentile compresses from 39 percent to 34.7 percent. The distribution narrows: the play trades upside for downside protection, a classic barbell collapse. The goal probability drops by more than a percentage point, meaning fewer paths in the simulation reach the reference hoard's target. The machine is saying that reallocating into commodities exposure—even if the toll-road thesis proves correct—costs optionality in this particular hoard's distribution. The Play does not improve the odds; it changes the shape. Whether that shape is preferable depends on preferences the engine does not model.
The duopoly framing is already wrong: the industrial gas market is an oligopoly with at least five global players, and Air Liquide alone commands comparable or larger market share than Linde in European and Asian fab supply chains. Pricing power in oligopolies is weaker than in duopolies, and long-term contracts with hyperscaler-backed fabs often include volume ratchets and renegotiation clauses that limit toll-road economics. The chain of causation—accelerated chip depreciation leads to sustained or increased fab throughput—is an assumption, not a mechanism: hyperscalers could respond to compressed useful life by slowing deployment, shifting to general-purpose architectures with longer cycles, or renegotiating foundry contracts to spread risk. Faster write-downs do not mechanically equal more wafer starts; they could equally signal a pullback in custom silicon investment. The commodities allocation in the Play is a broad basket, not a pure-play bet on industrial gas suppliers, so the simulation is testing a diluted version of the thesis. Linde and Air Products derive revenue from healthcare, food processing, and petrochemical markets; their exposure to semiconductor fabs is material but not dominant, and a fab boom that fails to materialise would not crater their businesses the way it might crater a pure-play chip designer. The August 2024 to December 2025 window in the thought experiment assumes that depreciation signals and margin compression would become consensus within sixteen months, but markets have repriced semiconductor cycles on much longer lags historically. The Play could be right about the thesis and wrong about the timing, or right about the timing and wrong about the commodities basket as the expression vehicle. New entrants in industrial gas supply—particularly in Asia, where fab buildouts are concentrated—could erode oligopoly pricing before the depreciation wave crests. The entire read depends on hyperscalers continuing to prioritise custom silicon despite deteriorating unit economics, and the alternative branch is that they pivot back to off-the-shelf GPUs and extend useful lives, collapsing the thesis entirely.
Disclosure of actual useful-life assumptions in hyperscaler financial statements would falsify or confirm the depreciation thesis: if Amazon, Google, and Microsoft are booking custom AI chips on seven-year schedules, the Trainium replacement is an outlier; if they are booking on three-year schedules, the thesis has already been priced. Public fab utilisation data from TSMC, Samsung, and Intel Foundry Services would show whether wafer starts are accelerating, decelerating, or holding steady in the face of chip architecture churn—if utilisation is falling while chip generations are accelerating, the second-order mechanism breaks. Pricing trends in industrial gas long-term contracts, particularly any disclosed renegotiations or volume rebates tied to fab throughput, would reveal whether the toll-road economics are durable or contested. A hyperscaler announcing a return to general-purpose GPU architectures for a major workload category, or extending the useful life of an existing custom chip beyond thirty months, would suggest the depreciation trap is avoidable. Air Liquide, Nippon Sanso, or Messer winning a marquee fab supply contract away from Linde or Air Products would weaken the oligopoly pricing assumption. A sustained period—say, four consecutive quarters—in which Linde's and Air Products' semiconductor-segment revenue growth decouples from hyperscaler capex growth would indicate that fab throughput is not rising in lockstep with AI spend. Commodity basket performance diverging sharply from industrial gas equities would show that the Play's vehicle does not track the thesis. Any of these observations, if they appeared in the data by December 2025 in this hypothetical branch, would require abandoning or revising the read.