Market — AI Data Stack

Verified claims and product-axis read for **AI Data Stack**. Every fact below is sourced; every product judgment traces back to underlying signals.

Top products (engine read)

Data Labeling Platform for AI / RLAIF — data labeling + preference data pipeline

**Opportunity:** Meta's $14.3B into Scale AI confirms data labeling / preference data is now strategic infra at the frontier-lab tier. Community is asking concrete operational questions ('audit-ready preference data', 'top tools for RLAIF'), suggesting the next layer below Scale (auditable / agent-loop / synthetic alternatives) is open.

Hottest single capital signal in the bucket — Meta-Scale deal validates labeled data as the bottleneck for frontier models. Scale is the incumbent; opportunity sits in auditability, RLAIF/agent-loop labeling, and synthetic data alternatives that the community is openly probing.

Vector Database — vector database

**Opportunity:** Massive 'how-to / which one / is it dying' query volume around vector DBs — indicates a confused but high-intent buyer market. Hyperscaler entry (S3 Vectors) is reframing the category from standalone DB to commodity feature; 1B-scale indexing and cost remain unresolved pain points.

Vector DB is the most discussed primitive in the ai-data-stack right now, but the conversation has shifted from 'what is it' to 'is it still a category' after AWS S3 Vectors launch. Hot but contested — incumbents (Pinecone/Weaviate/Milvus class) face commoditization pressure; in-process / lightweight forks (Zvec, LEANN, FerresDB) are emerging as the differentiated wedge.

RAG Infrastructure — retrieval-augmented generation pipeline

**Opportunity:** Production RAG cost ($2.4K/mo / 73% reduction story) and operational burden ('mostly infrastructure maintenance') are repeatedly surfacing — gap for managed RAG-ops platforms. Also a trust/leakage concern (LlamaIndex silent OpenAI fallback) signaling need for verifiable local/private RAG.

RAG is past hype peak into 'painful production' phase — the demand signal has shifted from build-it to operate-it/cut-cost-of-it. Big opportunity for RAG-ops, cost optimization, and verifiable-local stacks; existing tooling (LlamaIndex) is being audited.

Ontology + Graph Database Layer for LLMs — knowledge graph / ontology layer

**Opportunity:** Reaction against vector-only retrieval — demand for structured facts/ontologies layered on top of LLMs and vector search. Likely opportunity for graph+vector hybrid retrieval and ontology-as-a-service.

Emerging counter-trend to 'vector DB solves everything'. Still discussion-stage in this dataset, no committed companies visible, but recurring pairing of graph vs vector suggests hybrid retrieval is the next architectural debate in ai-data-stack.

Cross-cutting opportunities (industry read)

• **800G/1.6T Silicon Photonics & Co-Packaged Optics (CPO)** — Pluggable copper SerDes is at its reach limit for rack-scale GPU fabrics; 800G→1.6T and CPO are the only paths to keep up with NVLink/UALink rack densities.
• **AI-Optimized Ethernet Switching Fabric (RoCEv2, lossless)** — Hyperscalers want to break NVIDIA/Mellanox InfiniBand lock-in on GPU back-end fabrics; RoCEv2 + UEC-style lossless Ethernet is the consensus alternative but requires retimers, congestion control, and NCCL tuning to match IB.
• **Direct-to-Chip Liquid Cooling & CDU (100kW+/rack)** — Rack densities have crossed the 100kW threshold; air can no longer remove heat from H100/B100/B200 racks. Every colo + every cooling vendor needs a direct-to-chip or rear-door HX SKU on the shelf.

_See [Products](./products) for the full product list and [Strategy](./strategy) for forward-looking judgment._