Data centre sizes and costs are doubling about every 12 to 16 months
Read in full here:
The global AI market reached $514.5 billion in revenue in 2026, growing 19% from $390.9 billion in 2025. Forecasts project the market will compound at 30.6% annually between 2027 and 2033, hitting $3.5 trillion by the end of the period.
And according to this source.
That scale makes the unit economics harder than the headline market-size CAGR suggests. Infrastructure spend is paid in advance, while useful revenue depends on utilization and whether workloads can stay within a bounded context budget. In a couple of local agents I use [Contextpress](GitHub - Taha-azizi/contextpress: Deterministic context compression for LLM chat, RAG, and agent pipelines · GitHub), a Python library for local context compression, to trim stale tool output before inference; it cuts token spend and makes per-task cost visible instead of hiding it behind ever-growing prompts. For the $6tn claim I would want to see revenue per GPU-hour (or per completed task), utilization, depreciation, and energy/network costs—not just market revenue.