77% of FM-dependent products lock onto three vendors

Of the catalog's non-FM rows, 140 explicitly name a foundation model in their llm-lock or runtime-dependency cells — about 16% of the non-FM population. Within that 140, 108 (77.1%) depend on OpenAI, Anthropic, or Google. The long tail is sparse: Qwen 16, Mistral 12, Cohere 8, DeepSeek 6, Jamba 3, Grok 2, Nova 2.

The substrate-dependency-risk implication is that any TOS change, pricing change, or output-policy change at OpenAI, Anthropic, or Google ripples into roughly four-fifths of the named-FM dependency surface in the catalog. The other 32 FM-dependent rows are distributed across seven labs — meaning the diversification tail is too thin to absorb concentrated-vendor disruption.

The strategic read: a vendor positioning as the "second source" for FM-dependent agent products (multi-model routing, model-agnostic memory, vendor-neutral abstractions) is selling into a market that has 108 willing buyers whenever single-vendor-risk anxiety spikes. Today the market mostly pretends single-vendor risk isn't there.