Eighteen months ago the sophisticated move was breadth. You assembled your own stack: a tool for prompts, a tool for evals, a tool for vector search, a tool for orchestration, a tool for observability. Today that same stack is a liability, because most of those tools are being absorbed, from above by the model providers and from below by the platforms your data already lives on. This is the ordinary physics of a maturing market, and it is worth naming plainly, with the receipts, so you can plan around it instead of renewing into it.
Squeezed from above
The model providers spent 2025 turning the surrounding tools into features. OpenAI’s own year-in-review for developers lists, as standard kit, an Evals API for evaluation, built-in tracing and guardrails in its Agents SDK, higher-level orchestration through AgentKit, and hosted retrieval through File search and vector stores. Each of those was, a year earlier, a paid product from a startup. Anthropic is pulling the same plumbing inward: its work on advanced tool use added a Tool Search Tool that lets a model discover tools on demand instead of loading them all up front, lifting one model from 79.5 to 88.1 percent on its tool-use evaluations, and its Model Context Protocol has become a cross-vendor way to wire tools in at the model layer. When the provider ships evals, tracing, and retrieval in the box, a standalone version has to be dramatically better to justify a separate contract, a separate login, and a separate thing to break.
Squeezed from below
At the same time, the platforms beneath the model are pulling capabilities in. The clearest casualty is the standalone vector database. Amazon shipped native pgvector support in its managed Aurora Postgres in April 2025, which means the database you already run can do semantic search and retrieval without a separate vector store. The flagship independent, Pinecone, was reported to be exploring a sale by August 2025 after losing a marquee customer in Notion, this for a company that had raised $100 million at a $750 million valuation two years earlier. The sale was never confirmed and Pinecone remained independent into mid-2026, so read it as a pressure signal, not a closed outcome. Operators are voting with their architectures: the team at Confident AI wrote up replacing Pinecone with pgvector precisely to stop running a second database just for search. The same absorption is hitting observability. In January 2026 Snowflake agreed to buy the observability platform Observe, reported at around $1 billion and its largest acquisition ever, and it closed the deal that February. Its chief executive framed the point exactly: bringing observability “directly into the Snowflake AI Data Cloud.” Months earlier, Databricks had bought the serverless-Postgres company Neon, noting that more than 80 percent of the databases provisioned on Neon were being created automatically by AI agents, not people.
The question for any tool in your stack is no longer “is it good?” It is “is it good enough to outlive the platform shipping the same thing in the box?”
This is absorption, not collapse
It would be easy to over-read all this as the death of the category, and that would be wrong. The money is still pouring into infrastructure: Menlo Ventures pegged 2025 enterprise generative-AI spend at $37 billion, more than triple the prior year, with $18 billion of it going to the infrastructure layer. The standalone tools are not dying; they are being subsumed, and the strongest survive at the demanding end. pgvector covers the common case, but it is still CPU-bound and strains at billions of vectors and the tightest latency budgets, which is exactly where a dedicated vector engine keeps earning its contract. The shift is not “point tools are over.” It is that the default home for a capability is moving into a layer you already pay for, so the point tool now has to justify itself against free.
What to do if you bought the sprawl
If you assembled a best-of-breed stack in 2025, you were not wrong. You were early, and early has a cost. Three moves now. Audit for overlap: anything your model provider or cloud does natively today is a renewal you can probably drop. Watch the seams: consolidation breaks integrations, and the tool that quietly stops being maintained is more dangerous than the one that shuts down loudly. And favor the boring core, the parts of the stack that own your data and your evaluation, the things that are yours rather than a thin wrapper around a model API. Menlo’s same survey found 76 percent of AI use cases are now bought rather than built, up from just over half a year earlier. The sprawl was a phase, not a destination. The teams that travel lightest from here treat their tooling as something to subtract from, not only add to.