Semantic caching is an empty market

Across the v6 catalog's 100-product priority cohort, semantic caching shows up as a real shipping feature in exactly one place: LangChain's SemanticCache. Every other entry that claims caching is either token-level prompt caching (Anthropic, OpenAI), CDN-style response caching, or vector-store dedup — none of which is what "semantic cache" actually means in the agent-cost-economics literature.

The opportunity is structural. The architecture is well-understood (embed prompt → kNN against a cache of prior prompt+response pairs → return cached completion when similarity > threshold). The wins are large (10-100× cost reduction on repetitive workloads, ~10× latency reduction). And nobody else has shipped it as a vendor-managed layer with cross-model similarity. A semantic cache as a service — sold by Anthropic, OpenAI, or as a third-party gateway like Helicone but cache-first — has no direct competitor in the priority cohort.

The empty-market signal is reinforced by the cost-economics analysis showing prompt caching is the only cache technique with broad adoption — and prompt caching only deduplicates exact-prefix repeats. Semantic caching catches paraphrased and reformulated repeats, which is where the long tail of LLM cost lives.