The hidden menace behind Big Tech’s AI arms race: Meta, Amazon and others are spending billions on hardware that’s worthless in 3 years

Research Affiliates argues AI capex has surged from $250 billion in 2024 to $650 billion this year, but much of that spending may generate poor returns because GPUs and related hardware become economically obsolete in about 3 years. The article says hyperscalers such as Microsoft, Amazon, Alphabet, and Meta are using AI primarily to defend core franchises, even as they take losses on AI services and must keep reinvesting to preserve competitive position. The key implication is that AI spending may boost hardware demand and product adoption without translating into strong shareholder value for the Big Four.

Analysis

This is less an AI supercycle thesis than a capital-intensity trap for the incumbent platforms. The market is still pricing the hyperscalers as if AI spend compounds into durable operating leverage, but the more relevant frame is a zero-sum distribution war where each dollar of AI capex mostly preserves share rather than expands the profit pool. That implies the biggest second-order winner is not necessarily the platform owner, but the layer above and below it: model-agnostic software, enterprise workflow vendors, and energy/infra suppliers that monetize the buildout without taking product margin risk.

The key near-term risk is not that AI demand disappoints, but that depreciation and replacement cycles force earnings revisions before revenue catches up. Because the economic life of the hardware is shortening faster than the accounting life, reported margins can look resilient for 2-4 quarters and then slip abruptly once refresh intensity and price/perf gains reset the cost base. That creates a classic late-cycle setup: consensus extrapolates operating leverage, while the real P&L is getting hit by stealth maintenance capex and lower monetization per watt.

For NVDA and AMD, the bullish argument is still intact tactically, but the medium-term valuation case is more fragile than the revenue story suggests. If customers increasingly treat accelerators as consumables, order visibility improves, yet pricing power may normalize faster as buyers demand shorter depreciation assumptions and faster ROI hurdles. The better way to express the view is not a naked short on semis, but a relative underweight versus businesses with recurring software economics and lower reinvestment drag.

The contrarian miss is that the AI spend surge may be value-destructive for the mega-caps while still being highly accretive to their ecosystem. Investors are focused on who owns the model layer, but the real monetization could accrue to cloud infrastructure, networking, power, cooling, and vertical software that captures productivity gains from end users. In other words, the capex is a tax on hyperscaler ROIC, but a subsidy for everything that helps customers extract value from the output.

Key Decisions for Investors

• Short basket MSFT/AMZN/GOOGL/META vs long QQQ-neutral software enablers for 3-6 months; thesis is margin compression from maintenance capex outpacing AI monetization, with better risk control than an outright market short.
• Long NVDA Jan-2026 calls only on 10-15% pullbacks, but hedge with short GOOGL or META against the view that accelerator demand stays strong while platform ROIC disappoints; best if the market re-rates capex as consumptive rather than growth.
• Pair long ANET and/or EQIX vs short AMZN on a 3-9 month horizon to express that the monetization sits in infrastructure and colocation rather than in the hyperscaler P&L.
• Avoid chasing AMD here; if anything, use strength to trim exposure and rotate into companies with higher switching costs and lower refresh dependence. Risk/reward is inferior until the market prices in faster obsolescence and slower AI ARPU expansion.
• Set a trigger to add to shorts if one of the Big Four guides capex up again without raising AI revenue disclosure; that would likely be the first sign that the market is entering the 'maintenance spend' phase and estimates are still too high.