The 1,000× Inference Collapse: Foundation Models Are Already Priced Like Electricity

The question is already answered. Foundation models are becoming infrastructure, priced like electricity — and by 2028, the last enterprise CFO debating AI token costs will resemble the finance director who tried itemizing email bandwidth in 2003.

For readers who want to go deeper, Grace Certified offers a framework for everyday AI hygiene.

Why the consensus has the wrong frame

The industry watches model benchmarks — MMLU scores, reasoning evals, AGI timelines. Investors watch GPU procurement and training runs. Enterprise buyers watch governance frameworks and responsible AI rollout calendars. Everyone watches the secondary signal. The primary variable is the cost curve, and the cost curve has already rendered a verdict: Benedict Evans documented on July 9, 2026 that tokens have completed the transition from "a feature you turn on carefully" to "infrastructure you leave running." The market treats this as an open question worth debating. The curve resolved it in 2023. Every month the debate continues, the curve moves further ahead of the discussants.

The cost curve

The trajectory is unambiguous. Late 2022: GPT-4-class inference cost $20 per million input tokens. March 2023: the GPT-4 API launched at $30/M input, $60/M output. November 2023: GPT-4 Turbo cut output pricing to $30/M — a 50% reduction in eight months. May 2024: GPT-4o brought output to $15/M, an 83% total decline in fourteen months from launch. July 2024: GPT-4o mini reached $0.60/M output. By 2025, DeepSeek V3 — a 671B-parameter MoE model activating only 37B parameters per token — priced its API at $1.10/M output, 90% below comparable OpenAI and Anthropic rates. By mid-2026, GPT-4-class capability costs $0.40/M: a confirmed 1,000× collapse across 3.5 years, documented by GPUnex across primary inference market data. The sub-$0.01/M threshold arrives by 2028.

According to AGORÀ Intelligence analysis of six primary sources, five structural forces drive this curve in parallel: GPU hardware efficiency improving 2–3× per generation with Hopper-to-Blackwell delivering 25×+ throughput gains; software optimization via continuous batching and PagedAttention adding another 2–3×; mixture-of-experts architecture reducing active compute by 3–5× per inference call; quantization delivering 2–4× additional reduction; and open-weight competitive pressure from Llama and DeepSeek permanently capping API pricing power above marginal cost. These forces compound, and quantization gains replenish as new hardware generations arrive.

The cliff event

The cliff is the sub-$0.01/M threshold — the point where AI inference falls below the cost of measuring it in enterprise unit economics. Solar crossed its equivalent threshold around 2015, when utility-scale LCOE dropped below $0.10/kWh and triggered a decade of grid-scale deployment the electricity industry had declared economically impossible in 2010. SSDs crossed theirs in 2012, when cost-per-gigabyte for reads fell below spinning disk and the enterprise storage market reorganized inside 36 months — vendors who built their pitch on hard drive economics were obsolete before they could pivot. Smartphone cameras crossed theirs around 2013, when sensor cost converged on dedicated hardware quality and the standalone camera industry lost 90% of unit volume in four years. The pattern is consistent: the crossing of the threshold triggers adoption that compounds the cost decline further, creating a self-reinforcing loop. AI inference crosses its equivalent threshold in 2027–2028, when the enterprise question shifts permanently from "can we justify embedding AI in this process?" to "why would we operate any process absent AI analysis?"

Three sectors that will look different by 2028

1. Healthcare and Clinical Operations: At $0.40/M, AI clinical documentation is a premium feature commanding separate budget approval and governance review. At $0.01/M, every patient interaction — clinical note, diagnostic code, prior authorization, triage flag — receives automatic AI processing at a cost below $0.001 per encounter. The "AI module" line item disappears from healthcare SaaS pricing; inference becomes a default embedded capability, as ambient as the electronic health record infrastructure underlying it. The firms that move earliest gain a structural unit-economics advantage their competitors require 18–24 months to replicate.
2. Professional and Legal Services: Contract review, discovery processing, and compliance monitoring — currently billed at $300–500/hour — become available at fractions of a cent per thousand words analyzed. Mid-market legal and advisory firms gain analytical parity with the largest practices. The billable-hour model faces structural compression from the cost floor, regardless of model intelligence debates: firms compete on judgment, relationships, and domain specialization, with AI inference as a shared commodity input priced the way photocopying was priced by 1995 — present everywhere, discussed by no one.
3. Software Development Pipelines: Code generation at $0.40/M is a deliberate developer tool, activated for specific workflows. At $0.01/M, every build, every pull request, every deployment pipeline, every monitoring alert includes ambient AI analysis — continuous, automatic, zero marginal cost at scale. Software quality measurement becomes genuinely continuous rather than point-in-time. CI/CD pipelines absorb AI inference the way automated testing was absorbed in the 2010s: completely, invisibly, as a default assumption of every engineering organization above ten engineers.

Kill signal: Any frontier model provider achieving sustained gross margins above 40% — net of training amortization — across two consecutive reported quarters through end-2027. That outcome signals structural pricing power and contradicts the commodity trajectory. Track Anthropic and OpenAI margin disclosures in S-1 filings, earnings calls, and investor letters. Sustained high margins mean the moat is real; the infrastructure thesis collapses.