Skip to content

Kelford Press

Signal from the noise

§ The Long FormEconomics · Procurement

When AI costs beat human labor.

Token and compute prices are climbing fast. For a growing share of workloads, the AI is now more expensive than the human it replaced. The question is no longer whether AI works — it is whether the unit economics survive contact with production.

10 May 2026 · 8 min read · Kelford Press editorial

The 2024 thesis was simple: AI is cheap, humans are expensive, the substitution math is one-way. Two years in, the math is uglier. Token prices on the frontier tier have risen 30–40% on a per-call basis, agentic workloads multiply call volume by 10–100x, and the all-in cost of a real production agent on a real workload now routinely exceeds the cost of the analyst it was supposed to free up. The strategic question this raises is the one nobody wanted to ask in 2024: is this a generational productivity unlock, or is it a margin-compression race?

§ 01

The unit-economics squeeze

Three things are happening simultaneously. First, frontier-tier inference prices are climbing — partly because the compute supply chain is tight, partly because labs are pricing for reasoning workloads, not chat. Second, agentic patterns multiply call volume: a single user task can fan out to dozens of LLM calls across planning, tool-use, reflection, and re-planning. Third, the cost of being wrong went up — failed agent runs aren't free; they consume tokens, they corrupt state, and they need human triage.

  • Frontier inference: 30–40% higher per call vs 12 months ago.
  • Agentic call fan-out: 10–100× more inference per user task than chat.
  • Cost of error: failed runs cost tokens + cleanup + trust.
  • Net effect: production agents are routinely 2–5× the cost of the human they replaced.

§ Subscriber-only · 2 sections ahead

The rest of this dive is for subscribers.

Kelford Press deep dives are sourced, fact-checked, and free to read in part. The full piece — including 2 remaining sections and the takeaways — unlocks for subscribers from $4.99/month.

No tracking pixels. One-click cancel. Newsletter stays free either way.

§ Further reading

  1. 01Best AI models April 2026: ranked by benchmarks
  2. 02GPT-5.4 vs Gemini 3.1 Pro (2026): which AI wins?