Amazon, Google, Meta, Microsoft Plan Over $500bn AI Capex as Inference Costs Shift to Enterprise
This digest was compiled by AI from multiple sources — links to the originals are below.
Amazon, Google, Meta and Microsoft will allocate more than $500 billion in capital expenditure to AI infrastructure in 2026, according to an analysis by AVEVA's CEO in TechRadar. The author argues enterprise boards are underestimating the shift from one-time model training to continuous inference, where token-based consumption becomes the key cost driver. The piece arrives amid research suggesting the AI inference market will more than double within five years.
Hyperscaler AI Spending
Amazon, Google, Meta and Microsoft plan more than $500 billion in combined AI infrastructure capex for 2026, concentrated in data centers. Enterprise demand expectations have driven investor interest in hyperscalers, frontier labs and GPU makers. According to the AVEVA analysis, industrial sectors such as manufacturing, energy and transportation will never train frontier models themselves.
Inference Cost Shift
Training OpenAI's GPT-4 alone cost more than $100 million, but the AVEVA analysis argues that figure is largely irrelevant to industrial enterprises. Inference, by contrast, runs continuously in live operating environments to generate recommendations, support decisions and take actions. Agentic workflows require multiple model calls per task, large context windows and repeated reasoning loops, making them many times more compute-intensive than earlier generative AI tools.
Tokenomics and Cloud Parallel
Tokens, the units through which generative AI systems process and generate information, are described as the kilowatt-hour equivalent for AI consumption. Just as utilities charge per kilowatt-hour, AI providers can bill by token, aligning cost to usage. Research cited in the piece suggests the AI inference market will more than double over five years, echoing the earlier cloud migration.
What's Next
Enterprise finance and technology leaders will need to introduce token-based metrics as agentic deployments scale through 2027. It remains unclear whether board-level budgeting models will adapt quickly enough to prevent inference costs from outpacing planned IT allocations.
1 source
Amazon, Google, Meta, Microsoft Plan Over $500bn AI Capex as Inference Costs Shift to Enterprise
