Intelligence refinement. Rent early, build custom at scale.
Why it matters
Foundation models are the smelters, expensive, few can operate at scale. Once refined, the gold is a commodity, which is why model providers need to move up the chain.
The Smelter & Refinery
Raw ore becomes pure gold through smelting. In AI: foundation, specialized, and reasoning models refine raw data into intelligence. Refining is expensive and only a few can do it at scale, but once refined, the gold is a commodity.
The 5 sublayers
L2a
Foundation & Multimodal Models
Large pre-trained generalists, GPT, Claude, Gemini, Llama, and vision-language-action and video models (Sora, Veo) that span text, image, audio, and motion
L2b
Specialized & Fine-Tuned Models
Domain-tuned, distilled, and PEFT/LoRA-adapted models for specific verticals or tasks (BloombergGPT, Med-PaLM, Codestral)
L2c
Embedding & Retrieval
Vector representations, search indices, reranking, and RAG infrastructure
L2d
Model Routing & Composition
Selecting, chaining, ensembling, or mixture-of-experts routing across multiple models per task to balance cost, latency, and quality
L2e
Reasoning & World Models
Extended chain-of-thought, planning, and multi-step inference, plus predictive world models (V-JEPA, Genie, Sora-as-simulator) that let agents and robots imagine outcomes before acting
, Layer diagnostic card · SCOI v1
Is a company really at L2?
The smelter, foundation, specialized, and reasoning models that refine data into general intelligence.
Inclusion tests · include if ALL
Trains models from scratch (or substantively post-trains with proprietary L1).
Owns model weights and can ship without third-party model licenses.
Compute spend is the dominant cost line.
Exclusion tests · exclude if ANY
Calls a closed-source API and fine-tunes prompts. That is L7, not L2.
Distills or wraps another lab's open weights with no novel training.
RAG over a model you don't own, L2c at most, usually L5c.
The L2 removal test
Remove your in-house model and substitute the best public foundation model. If the product is unchanged, you are not at L2.
Economic work this layer does
Converts raw data + compute into generalized capability that downstream layers can rent per token.
Canonical examples
OpenAI
Trains frontier models; charges per token; absorbs L7 wrappers structurally.
Anthropic
Frontier models plus L3 trust posture for regulated buyers.
Google DeepMind
Frontier models tied to L0/L4 distribution, a fortress, not a pure L2.
Anti-examples · look-alikes that fail
Most 'foundation model' startups
Fine-tunes on someone else's base. L2b at best, no frontier compute.
Open-source distillers
Weights ship to everyone, by Law I, no margin lasts.
