Are the Hyperscalers Spending Too Much on AI, or Not Nearly Enough?
Amazon, Microsoft, Google, and Meta are collectively spending over $300 billion on AI infrastructure. Alphabet stock fell 10% in a single session on capex concerns. Here is the actual investment analysis.
The Numbers Are Staggering
Microsoft, Google, Amazon, and Meta have collectively guided for over $300 billion in capital expenditure in 2025 and 2026 combined. Microsoft guided for $80 billion in AI infrastructure capex for fiscal year 2026. Alphabet guided for $75 billion in capex for 2025, up 43% year-over-year. Amazon's AWS capex grew to $26.3 billion in Q1 2026, up 80% from the prior year. Meta guided for $60-65 billion in 2025 capital expenditures, up from $38.4 billion in 2024. This level of spending has driven NVIDIA's data center revenue to a quarterly run-rate above $35 billion, created a shortage of power infrastructure so acute that hyperscalers are negotiating directly with nuclear power plants, and generated a wave of data center construction reshaping real estate markets in Northern Virginia, Phoenix, Dallas, and other hyperscale hubs. Alphabet stock fell 10% in a single session in late June 2026 on concerns about AI talent departures. Amazon, Oracle, and Meta all declined 2-4% the same day on generalized AI capex concerns. The market is beginning to price in the possibility that the capex ramp is ahead of the monetization ramp.
The Magnificent Seven collectively increased capital expenditure by approximately 55% in 2025. Outside of that group, S&P 500 capex grew by a modest 4%. The AI investment cycle is extraordinarily concentrated in a handful of companies with the capital and competitive incentive to keep spending, which makes individual earnings reports in this group market-moving events for the broader index.
The Bull Case: Every GPU Is Generating Revenue
The most important distinction between the AI capex cycle and previous technology investment bubbles is utilization. When Cisco and Nortel built fiber optic networks in 1999-2000, they were laying capacity ahead of demand. The AI capex cycle is building infrastructure to meet demand that already exists. Every hyperscaler GPU deployed is generating revenue from day one: running inference for ChatGPT, Gemini, Copilot, or advertising optimization algorithms. Microsoft's Copilot products have over 300 million users. Google's AI Overviews appear in the majority of U.S. search queries. Meta's AI-powered recommendation system drives more than 50% of time spent on the platform. The second argument is competitive necessity: falling behind on AI infrastructure means losing market share in cloud, advertising, and enterprise software. The capex is high, but the cost of not spending may be higher.
The Bear Case: The Math Does Not Close at Current Prices
The bear case is straightforward: you cannot build a sustainable business by spending $75 billion a year on infrastructure to generate services that customers are using for free or at subsidized prices. The economics of inference are deflating rapidly as model efficiency improves, the cost to run a query through a leading AI model has fallen by roughly 90% over the past two years. To generate the same revenue, you need either dramatically more volume or meaningfully higher prices. Neither is guaranteed. And the recent concerns about AI talent departures at Alphabet, a signal that the human capital making the capex valuable may not be staying, added a new dimension to the bear case that the market priced in immediately.
The Inference vs. Training Distinction
Training a frontier AI model requires massive, concentrated compute bursts, thousands of NVIDIA H100 GPUs running for weeks or months. Inference, running that model to answer user queries, requires distributed compute that scales with usage. The economics are different, and the competitive dynamics are different. Custom silicon, Google's TPU, Amazon's Trainium, Microsoft's Maia, is increasingly competitive for inference workloads, because inference on a known model architecture can be highly optimized. NVIDIA GPUs remain dominant for training because the CUDA software ecosystem is built around NVIDIA hardware. As the AI compute mix shifts from training-heavy to inference-heavy, the hyperscalers' custom silicon investments become more valuable relative to NVIDIA. The resolution of the AI capex debate will come through earnings: watch revenue-per-compute-dollar as the metric that determines whether the investment thesis holds.