The $150B Question: Why Amazon and Alibaba's AI Capex Is a Depreciation Time Bomb

Chasing shadows in the liquidity fog of 2017 taught me one thing: when a market narrative shifts from "potential" to "infrastructure," the bills stop being theoretical. Fast forward to 2025, and the narrative has shifted decisively. Amazon and Alibaba have committed a combined $150 billion to AI infrastructure. The markets cheered. I dug into the fine print. The systemic rot is hidden in the depreciation schedules, not the press releases. A quick framing exercise. The headlines lump Amazon's ~$100 billion annual capex guidance and Alibaba's ¥380 billion three-year plan into a single "AI arms race" bucket. That's a category error. Amazon's annual run rate is roughly 5.6x Alibaba's. The latter's ¥380 billion translates to about $53 billion total, or ~$17.7 billion annually. Even accounting for China's lower construction costs, the raw compute acquisition gap is 3-4x. These are parallel universes, not competitors. AWS is fighting Azure and GCP in the West. Alibaba is fighting Huawei and Tencent in the East. But here's the core analytical problem: neither company's narrative addresses the unit economics of what they're building. The market has been conditioned to price cloud giants on P/S ratios. That anchor has shifted. The new valuation lens is EV/EBITDA with a heavy weight on capital efficiency. And this is where the story gets uncomfortable. Let's do some forensic accounting. Alibaba's annualized capex of ~¥126.7 billion, straight-lined over six years, generates ~¥21.1 billion in annual depreciation. That's a significant drag on a company whose core commerce margins are already under pressure. Amazon's math is more alarming. At ~$100 billion annual capex and a five-year useful life, you're looking at ~$20 billion in annual depreciation. AWS's operating profit hovers around $40 billion. In a steady state, depreciation alone could consume half of AWS's incremental profit. Volatility is the tax on certainty, but depreciation is the tax on ambition. The real question is whether those depreciation assumptions hold. AI accelerators have a functional lifespan of 2-3 years before they're obsolete. If hyperscalers are forced to accelerate depreciation — and they will be, because the physical reality of silicon degradation and architectural obsolescence doesn't care about accounting policies — then current profits are overstated. The gray zone here is the depreciation policy itself. Microsoft and Google have historically extended server useful lives to flatter earnings. If Amazon or Alibaba follow suit, they're simply deferring a profit cliff. Correlation is the siren song of fools, and so is treating "capex announced" as "compute deployed." The lag between announcement and actual power-on is 18-36 months. That's the hidden detail in the fine print. Amazon's binding constraint isn't chip supply; it's power grid interconnection queues that stretch for years. Alibaba's constraint is geopolitical: export controls on advanced silicon force reliance on Huawei Ascend and Cambricon alternatives, which carry a real efficiency penalty. You're not comparing like-for-like dollars. Same capex, different compute — and 2025's earnings don't yet reflect this structural asymmetry. Here's the contrarian angle the headlines miss. Amazon's real strategic anxiety isn't capex payback — it's that AWS lacks a frontier model of its own. Its Nova series is an also-ran. Amazon is the only major cloud provider in a position where its AI revenue depends on Anthropic's success. That's the "pipeline" risk: becoming the infrastructure layer for someone else's intelligence. Alibaba's Qwen open-source strategy has the opposite problem: it gives away the crown jewels for ecosystem share, undermining direct API monetization. Their AI revenue will come from compute and platform, not model calls. The credit market adds another layer. Meta, Oracle, and CoreWeave have issued massive debt. This isn't just operating expenditure anymore — it's leveraged investment. The phrase "bills coming due" in the original analysis carries a double meaning: operational depreciation and literal debt servicing. If AI demand softens or interest rates stay higher for longer, that leverage compounds the downside. Innovation often precedes regulation by a decade, but leverage precedes crisis by about 18 months. The bottom line: the capex cycle is real, the demand signals are real, but the accounting fiction is the most dangerous variable. The market has already repriced Alibaba's story around AI narrative; a capex efficiency miss will hit that valuation hard. For Amazon, the risk is slower — but it's a slow bleed through depreciation and a strategic void at the model layer. So where does that leave us? The AI infrastructure buildout is not a fraud. It's a massive, real allocation of capital. But the "bills coming due" warning isn't about bankruptcy — it's about margin compression, asset impairments, and a brutal accounting reckoning. The market is betting on a future where AI demand fills every new data center. If that demand curve bends even slightly, the depreciation tax becomes an extinction event for earnings quality. History doesn't repeat, but it rhymes in code — and this particular rhyme was written in 2000, with fiber optic cable instead of GPUs.