There is a quiet, expensive war being fought inside every AI agent. It is a battle of retrieval, where queries must be parsed across separate search clusters and vector databases, each pulling from its own copy of the data. The compute and complexity costs add up, a tax paid in latency and operational headaches. Infino AI, a new startup founded by engineers who built search infrastructure at Amazon and LinkedIn, is betting that war is unnecessary. Their proposition is disarmingly simple: one copy of the data, in a standard Apache Parquet file on object storage, that can handle full-text search, vector embeddings, and SQL all at once [Infino AI, about].
They recently emerged from stealth with a $7.5 million seed round led by Bessemer Venture Partners to make that bet [Ground.news, October 2026]. The goal is not to build another vector database, but to build what they call "retrieval for agents," a governed execution layer that could become the default way AI systems query their own knowledge.
The bet on a unified data surface
Infino's core technical wedge is consolidation. In a typical retrieval-augmented-generation (RAG) setup, a developer might maintain a separate Elasticsearch cluster for keyword lookups, a Pinecone or Weaviate instance for vector similarity, and a data warehouse for structured SQL queries. Infino proposes collapsing all three functions into a single Rust-based engine that operates directly on Parquet files in cloud object storage like Amazon S3 [Infino AI, about].
The promise is a 10x reduction in cost and complexity, according to the company's claims [Infino AI, homepage]. For developers building AI agents, particularly in sensitive environments like finance or healthcare, the appeal is a unified, auditable layer. Infino pitches itself as a "governed execution layer" that provides immutable lineage and centralized access controls, making it easier to deploy agentic search where compliance matters [Infino AI, blog].
A team built on distributed search
The founders bring a specific kind of credibility to this problem. The team includes Vinay Kakade, who led development of the AWS Elasticsearch Service and Amazon's internal distributed-search infrastructure, and Asif Makhani, former CTO of Handshake who worked on Amazon CloudSearch and LinkedIn Learning [FundraisingFox]. Co-founders Ekechi Nwokah and Murali Krishna round out a group with deep experience in search, machine learning, and large-scale systems at Google, LinkedIn, and Amazon [Infino AI, about].
This background is critical. Building a retrieval engine that is both performant and reliable on object storage is a distributed systems problem first and an AI problem second. The team's pedigree suggests they understand the scaling and durability challenges inherent in replacing dedicated search clusters.
| Founder | Previous Roles | Relevant Experience |
|---|---|---|
| Vinay Kakade | Lyft, Amazon, AWS | Led AWS Elasticsearch Service; built Amazon's distributed-search infra. |
| Asif Makhani | Handshake (CTO), Amazon, LinkedIn | Worked on Amazon CloudSearch, LinkedIn Learning. |
| Ekechi Nwokah | Migo, A9.com | Co-founded fintech Migo; background in security and search. |
| Murali Krishna | Amazon, AWS | Early work on AWS's first search services. |
| Table: Infino AI's founding team draws from a deep bench of search and cloud infrastructure experience [Infino AI, about][FundraisingFox]. |
The path to production and pricing
As an open-source project, Infino's adoption will hinge on developer experience and raw performance. The company has published benchmark figures, claiming ingestion rates of about 33,100 documents per second for full-text search and 62,200 docs/sec for vector embeddings on a 10-million-document table [Infino AI, blog]. These numbers, while impressive in a controlled test, are just the opening argument.
The commercial model meters usage based on the work performed during data ingestion, with a "write token" that follows bytes written [Infino AI, pricing]. This aligns cost with value, charging for the computational lift of indexing rather than for stored data. The real test will be whether teams running production AI workloads find the unified approach simpler and more cost-effective than stitching together best-of-breed point solutions.
Where the wheels could come off
The ambition to unify is also the primary risk. The incumbent stack of separate databases exists for a reason: each is optimized for a specific type of query. Asking one engine to be exceptional at keyword search, vector similarity, and complex SQL is asking a lot. Performance trade-offs are inevitable.
- The performance trap. While benchmarks show speed, they may not reflect the jagged profile of real production queries. A slowdown in hybrid search, which fuses results from different retrieval methods, could be a deal-breaker for latency-sensitive agents [Infino AI, agents].
- The ecosystem moat. Established vector databases and search platforms are not standing still. They are rapidly adding agent-centric features, governance tools, and tighter cloud integrations. Infino must out-innovate and out-execute well-funded incumbents.
- The open-source adoption curve. Success depends on developers choosing to rebuild their retrieval stack around a new primitive. This requires not just a better mousetrap, but compelling documentation, easy deployment, and a community that provides answers at 2 a.m.
The company's most plausible answer is that the complexity tax of the current fragmented stack is high enough that a good-enough unified solution wins on total cost of ownership, especially for teams that prioritize simplicity and security over peak performance in any single domain.
The next twelve months
With $7.5 million in seed funding, the immediate task is to move from a promising open-source project to a commercial service with clear enterprise traction. The next milestones will be less about technical benchmarks and more about named production deployments. The team needs to prove that their engine can handle the messy, unpredictable queries of real AI agents in the wild.
Investors like Bessemer's Lauri J. Moore are betting that the team's infrastructure chops can translate into a new category standard [Ground.news, October 2026]. The market is signaling a need for simplification; the average AI project's data stack has become a Rube Goldberg machine of APIs and sync jobs.
On the back of an envelope, the unit economics argument is straightforward. If a mid-sized engineering team spends $50,000 a month on separate vector database, search cluster, and cloud data warehouse costs for its agent operations, Infino's promise of a 10x cost reduction turns that into a $5,000 monthly bill. The savings alone could fund a small engineering squad. The real value, however, isn't just in the saved dollars but in the saved weeks of developer time not spent debugging why the vector store and the full-text index return different answers for the same question.
For Infino to succeed, it must become the obvious choice for developers who are tired of that debug session. Its incumbent to beat isn't a single company, but the entrenched habit of using three different tools for three kinds of search. It is betting that a single, well-engineered Parquet file can break that habit.
Sources
- [Infino AI, about] Infino | About | https://infino.ai/about/
- [Ground.news, October 2026] Pakistani-origin founders' startup Infino AI raises $7.5M seed | https://ground.news/article/pakistani-origin-founders-startup-infino-ai-raises-75m-seed
- [Infino AI, blog] Infino AI Blog | https://infino.ai/blog/
- [Infino AI, homepage] Infino AI Homepage | https://infino.ai/
- [Infino AI, agents] Infino AI Agents | https://infino.ai/agents/
- [Infino AI, pricing] Infino AI Pricing | https://infino.ai/pricing/
- [Infino AI, security] Infino AI Security | https://infino.ai/security/
- [FundraisingFox] Infino AI Company Profile | https://fundraisingfox.com/companies/infino