Token Cost of Ownership

Narrative

Bloomberg has emerged as the dominant financial outlet for AI cost and infrastructure coverage. Its February 2026 news pieces quantified combined Big Tech AI capex at $650 billion for 2026 alone and characterised the associated debt financing as a '$3 trillion' market event — establishing the sheer scale of capital that undergirds current below-cost token pricing. Bloomberg's April 2025 newsletter marked the structural pivot from training to inference spend, while its January 2025 live video coverage of DeepSeek's market shock captured the moment financial markets first confronted the possibility that frontier-equivalent performance was achievable at 5–10% of incumbent labs' cost — a signal that the subsidy model could not persist indefinitely. Bloomberg's December 2024 interview with OpenAI's CFO, targeting $2,000/month enterprise subscriptions, revealed the commercial logic: pivot high-value users to flat-rate agreements that decouple revenue from per-token cost exposure. The April 2026 reporting on OpenAI pausing its UK Stargate data centre explicitly named energy cost as the binding constraint — confirming what Bloomberg Intelligence had warned: AI queries consume up to 10x the energy of traditional searches, making power a rising structural cost floor that token price declines cannot escape. Goldman Sachs' 2024 'Gen AI: Too Much Spend, Too Little Benefit?' report, featuring head of global equity research Jim Covello's scepticism that the projected $1 trillion in AI infrastructure would generate adequate returns, became the reference document for financial market bears — its core question about cost versus benefit remains unanswered in 2026.

The picture that emerges from combining financial press coverage with high-quality analytical sources (Epoch AI, SemiAnalysis, McKinsey) is one of a market in deliberate subsidy mode. Epoch AI's empirical tracking shows token prices for comparable capability falling 40x per year for some performance tiers since January 2024 — but frontier reasoning models have not followed commodity curves. Fortune's November 2025 reporting on OpenAI's leaked financial documents revealed the starkest number: $143 billion in projected cumulative cash outflow from 2024 to 2029, with inference costs at $8.4 billion in 2025 rising to $14.1 billion in 2026 — against a gross margin held below 35%. McKinsey's $7 trillion compute analysis and SemiAnalysis's bottom-up hardware economics confirm that the hardware and energy cost stack is rising, not falling, at the infrastructure level, even as algorithmic efficiency improvements drive per-token software costs down. The financial press story, as of April 2026, is that investor capital — not unit economics — is the primary determinant of what users pay per token, and the timeline to cost normalisation remains the central unresolved question for enterprise AI budgeting.