Token Cost of Ownership

Narrative

The independent blog and newsletter ecosystem tells three tightly interlocking stories about AI token pricing. The first is subsidisation at scale. ScaleDown (tinyml.substack.com), the most analytically rigorous Substack on inference economics, concluded in 2024 via bottom-up hardware modelling that LLM API providers are absorbing over 90% of the true cost of each token, describing the market as a VC-funded 'land-grab phase' structurally identical to Uber's early price subsidisation. SemiAnalysis (Dylan Patel et al.) provided the forensic hardware backing — from the landmark 2023 'Inference Cost of Search Disruption' piece through the January 2025 DeepSeek Debates — consistently showing that GPU die costs, memory bandwidth constraints, power draw, cooling, networking, and staff combine to make current API prices economically unviable at face value. The GitHub Copilot parallel ($10/month subscription against $30/month compute cost) is the canonical example cited across Substack, Medium, and LessWrong. The second story is LLMflation. Andreessen Horowitz coined the term in November 2024 and quantified a 10x annual cost decline for equivalent-performance inference over three years; the November 2025 arXiv paper 'The Price of Progress' gave this academic weight, documenting a 280x cost drop for GPT-3.5-equivalent performance between late 2022 and October 2024. Epoch AI refined the picture, showing price declines of 9x to 900x per year depending on capability tier — with frontier reasoning models holding price while commodity models collapsed — and projecting training costs crossing $1 billion per run by 2027. The third story, which has dominated 2025–2026 commentary, is the Jevons Paradox: despite token prices falling a thousandfold in three years, enterprise AI inference spend reached $37 billion in 2025, up 320% year-on-year. Multiple Substack writers (AI Proem, The Substrate) documented Meta raising 2025 AI capex by 50% after DeepSeek's efficiency announcement rather than cutting it. Ben Thompson at Stratechery frames this as 'the return of marginal costs' — arguing that AI has broken the near-zero marginal cost model underpinning two decades of tech valuations — while Tiny Empires and Aspiring for Intelligence extend the argument to agentic workflows, showing that multi-step chains multiply token consumption to the point where total cost of ownership bears little relationship to headline API rates. Across all three stories, the consensus is that current prices are strategically rather than economically determined, and that energy, data centre capex ($600B+ forecast for 2026), and hardware costs are building a structural floor that will eventually force a reckoning.