Trends›Models commoditise, value moves up
Models commoditise, value moves up
Raw LLM capability commoditises, so durable value capture moves up to the scaffolding, agent and application layer and down to chips, but because frontier labs are climbing into applications themselves the app layer is contested rather than free for independents.
74 confidence, computed 23 Oct 2026
- Economic
- High weight
What rests on thismarkets it underwrites or threatens
Evidencethe schedule the figure draws
The evidence, newest first
- Jul 2026
for · observable Private source - Jul 2026
for · primary Seventh near-parity frontier model in a fortnight (GPT-5.6 Sol/Terra/Luna at $5/$2.50/$1, Grok 4.5, Gemini 3.5 Flash, Meta's first paid model, Cognition sub-$2 coder, Thinking Machines' Inkling at 975B/41B active under Apache 2.0, then Kimi K3). The Once and Future Kimi (State of the Future) - Jul 2026
for · primary OpenAI researcher argues raw model capability commoditises (compute is a commodity, frontier models trained on substantially the same internet 'feel eerily similar'), so durable value moves up specifically to DATA — the exclusive, manually-collected corpora on which labs diverge (OpenAI… A Stargate for Data — will depue (OpenAI) - Jun 2026
against · claim Frontier labs are defending margin via usage-based metered pricing — OpenAI doubled GPT-5.4→5.5 token prices, Microsoft raised Copilot 3x-9x, and Anthropic pushes large customers to Enterprise ($20/seat + metered API rates, no token allowance) where a Claude Max task that… JPMAM — frontier token pricing & open-model migration - Jun 2026
for · observable 27 of 51 evaluated models are open-weight, with GLM-5.2 the top open model at Intelligence Index 51 — the open-weight frontier is closing on proprietary models. 27 of 51 evaluated models now open-weight (ArtificialAnalysis) - Jun 2026
for · observable Industry token prices fell ~80% 2025→2026; GPT-4's ~$30/$60 (2023) collapsed to ~$2.50/$15 (GPT-5.4), DeepSeek V4 Flash at $0.14/$0.28 — raw capability commoditising fast. LLM API pricing 2026 — token prices fell ~80% YoY (TLDL) - Jun 2026
for · primary Enterprise AI budgets grew from $1.2M annually in 2024 to $7M in 2026, but spending increasingly flows to inference and application infrastructure rather than foundation models. Foundation Model Valuations Near $1 Trillion But History Says Infrastructure Builders Rarely Win - Jun 2026
against · observable Inference consumes ~23% of revenue at scaling-stage AI companies, locking app-layer gross margins ~30 points below SaaS norms (~52% in 2026) — value capture at the app layer real but structurally thinner. AI agent economics: token tax locks margins ~30pts below SaaS (TechTimes) - May 2026
for · observable Xiaomi cut MiMo-V2.5 API prices up to 99% and DeepSeek holds V4-Pro at ~a quarter of prior cost — aggressive commoditisation of raw model access. China AI price war: Xiaomi cuts API costs up to 99% - May 2026
for · claim Cursor reached ~$500M+ ARR ~2.5 years post-launch — durable independent app-layer value achievable even as labs ship competing coding agents. Cursor ~$500M+ ARR ~2.5 years post-launch - Apr 2026
against · observable Cursor operated at negative gross margins until recently and one $100M-ARR app voluntarily cedes ~6 points of margin to incremental inference cost — app-layer value capture fragile. App-layer margins fragile (SaaStrAI) - Apr 2026
for · observable Anthropic reached a >$30B run-rate with Claude Code alone ~$1B annualized within six months — frontier labs themselves capturing the coding/agent application layer. Anthropic overtakes OpenAI, >$30B run-rate; Claude Code ~$1B - Apr 2026
against · primary Claude Code generated $2.5B ARR by February 2026, accounting for more than half of Anthropic's enterprise revenue, proving frontier labs can establish durable moats through application-layer vertical integration. The Foundation Model Threat: When Your AI Infrastructure Provider Becomes Your Competitor - Apr 2026
for · primary Performance gap between proprietary frontier models and open-source alternatives has narrowed to less than 5% on many core tasks, while inference costs remain 10x higher. The LLM Moat Is Collapsing: Why Your Frontier Model Strategy Is Already Dead
What would move itthe measures watched
- watch$/token price for frontier-capable models. Steep continued declines confirm raw-capability commoditisation; price floors holding would weaken it
- watchFrontier vs open-weight capability gap (Intelligence Index / SWE-bench). Narrowing gap = capability commoditising; widening gap = frontier moat reasserting
- watchFrontier-lab application revenue (Claude Code, ChatGPT, agents). Labs scaling app revenue = app layer contested by model owners; flat = third-party apps safer
- watchApp-layer / agent gross margins ('token tax'). Margins stuck well below SaaS norms = value not durably captured at app layer; expansion = durability
- watchIndependent app-startup ARR durability vs lab encroachment. Startups holding/growing ARR in lab-adjacent categories = durable value; churn to labs = contested