"The foundation model layer will be a low-margin utility within 36 months."
Is the foundation model layer structurally commoditising, or are frontier labs sustaining defensible margins?
Inference economics overwhelmingly support the commoditisation thesis on price-per-token. However, frontier capability frontier appears to be diverging — not converging — across reasoning, agentic, and multimodal benchmarks. Best read: the median-quality model layer is commoditising fast; the frontier remains highly differentiated.
GPT-4-class inference cost dropped ~99% in 24 months.
— Epoch AI Pricing Tracker 2026
Open weights (Llama 4, DeepSeek R3) match closed frontier on 70% of enterprise benchmarks.
— HELM Leaderboard May 2026
Frontier reasoning benchmark gap (GPQA-Diamond) widened from 4pp to 11pp in 2025.
— GPQA Leaderboard
Enterprise contract concentration at OpenAI + Anthropic increased to 71% of measured spend.
— Battery Ventures State of Cloud 2026
The capability ceiling is moving faster than the commodity floor — both are true.
Commoditisation is the median story; the tails are widening.
- State of AI Report 2025— Nathan Benaich
- Epoch AI Trends in ML Compute— Epoch AI
Hoffman is directionally correct but the timeline ('36 months') is aggressive for the frontier tier. Below frontier: already happening.
Confidence reflects strength of evidence weighted by source quality.
For the Geeks · sources & method▾
Structured expert elicitation (IDEA protocol)
4 independent domain experts, Delphi-style two rounds, aggregated via geometric mean of probabilities.
n = 4 experts · 2 rounds
- PitchBook — Global Venture Report Q1 2026 ↗Stage-by-stage deal counts, median round size, dry powder.
- Epoch AI — Notable Models DB ↗Training FLOPs, parameter counts, release cadence.
- Harmonic.ai — Founder GraphFounder pedigree, prior exits, team formation signals.
Last agent run · 2026-06-08