As AI companies shift from flat-rate plans to consumption-based billing and operating costs for AI agents surge, database firms are moving forward with infrastructure solutions that reduce token usage, The Register reported on June 24 local time.

Vector database company Pinecone recently launched Nexus. Nexus is a knowledge engine that organises an organisation's data structure and context in advance so AI agents do not repeatedly perform the same exploration tasks. Jeff Zhu (제프 주), Pinecone's vice president of product, said coding agents identify table structures and repeat exploration work each time they receive a question. He said they consume large amounts of tokens in the process even if they reach the right answer. He said Nexus handles such repetitive work in advance and stores reusable context.

According to The Register, IDC research director Devin Pratt (데빈 프랫) said Nexus reflects cost at the design stage rather than treating it as something to consider later. He praised its embedding of token-budget controls and usage tracking in the query layer.

Pratt's view is that the key challenge in agentic deployments has shifted from AI models to the surrounding data infrastructure.

According to an IDC survey, two data obstacles blocking the expansion of generative and agentic AI were security and compliance constraints and cost. Almost two-thirds of organisations operate more than 11 different database technologies, making fragmentation a serious problem.

TigerData, the developer of TimescaleDB, introduced Ghost, a database platform dedicated to AI agents. It instantly provides each agent with an independent PostgreSQL database and can be discarded after experiments, so it does not affect other agents and users.

Billing is based on the amount of computing time actually used rather than the number of databases. Ajay Kulkarni (아제이 쿨카르니), TigerData's co-founder and chief executive, explained, "Whether you have 1 database or 50, the cost is the same and you pay only for the computing time."

Large platform companies such as Snowflake, Oracle and Microsoft are also integrating similar features into their own stacks, The Register reported.