Vector databases have develop into an indispensable technological foundation at the starting of the era of the modern generation of artificial intelligence.
However, what has modified over the last 12 months is that vectors, the numerical representations of data used by LLM, are increasingly becoming just one other type of data in various databases. Now, Amazon Web Services (AWS) is taking one other step forward in vector ubiquity with the general availability of Amazon S3 vectors.
Amazon S3 is an AWS cloud object storage service widely used by organizations of all sizes to store all kinds of data. Most often, S3 is also used as a core component of Data Lake and Lakehouse implementations. Amazon S3 Vectors now adds native vector storage and similarity search capabilities directly to S3 object storage. Instead of requiring a separate vector database, organizations can store and query embedded vectors in S3 for semantic search, augmented search generation (RAG) applications, and AI agent workflows without moving data to specialized infrastructure
The service was first made available in July and initially offered 50 million vectors in a single index. In the GA version, AWS dramatically increased this number to 2 billion vectors per index and to 20 trillion vectors per S3 storage bucket.
According to AWS, customers have created over 250,000 vector indexes and consumed over 40 billion vectors in the 4 months since the preview launched. Scaling up with the introduction of GA now allows organizations to consolidate entire vector datasets into single indexes slightly than fragmenting them across infrastructure. GA’s launch also shakes up the enterprise data landscape, providing a recent, production-ready approach to vectors that has the potential to disrupt the market for purpose-built vector databases.
Adding fuel to the fire of the competition, AWS claims that S3 Vector might help organizations “reduce the total cost of storing and viewing vectors by up to 90% compared to specialized vector database solutions.”
AWS positions S3 vectors as complementary, not competitive, to vector databases
While Amazon S3 vectors provide a powerful set of vector capabilities, the answer to whether or not they replace the need for a dedicated vector database is a bit nuanced and depends on who you ask.
Despite aggressive cost statements and dramatic improvements in scale, AWS positions S3 Vectors as a complementary storage tier, not a direct alternative for specialized vector databases.
“Customers choose whether to use S3 vectors or vector database based on their application’s latency requirements,” Mai-Lan Tomsen Bukovec, vp of technology at AWS, told VentureBeat.
Bukovec noted which you could think of it as “performance tiering” based on the organization’s application needs. She noted that if an application requires super-fast response times and low latency, a vector database like Amazon OpenSearch could be a good option.
“But for many types of operations, such as creating a semantic layer of understanding of existing data or extending an agent’s memory with much more context, S3 Vectors works great.”
The query of whether S3 and its low-cost cloud object storage will replace the database type is also nothing recent for data professionals. Bukovec made an analogy to the way enterprises use data lakes today.
“I expect vector storage to evolve similarly to tabular data in data lakes, where customers will continue to use transactional databases such as Amazon Aurora for specific types of workloads, while also using S3 for application storage and analytics as the performance profile works and they need S3 features such as durability, scalability, availability and cost-effectiveness as data grows.”
How customer demand and demands have shaped Amazon S3 Vector services
In the first few months of preview, AWS learned what real enterprise customers actually need and need from vector data storage.
“We received a lot of very positive feedback on the preview, and customers told us they wanted these capabilities, but at a much larger scale and with lower latency, so they could use S3 as their primary vector store for most of their rapidly growing vector storage,” Bukovec said.
In addition to the increased scale, query latency was reduced to roughly 100 milliseconds or less for frequent queries, and rare queries accomplished in lower than one second. AWS has increased the maximum number of search results per query from 30 to 100, and write performance now supports up to 1000 PUT transactions per second for single-vector updates.
Increasingly popular use cases include hybrid search, extending agent memory, and building a semantic layer on top of existing data.
Bukovec noted that one trial customer, March Networks, uses S3 vectors for large-scale video and photo evaluation.
“The vector storage economics and latency profile mean that March Networks can cost-effectively store billions of embedded vectors,” she said. “Our built-in integration with Amazon Bedrock means it makes it easy to incorporate vector storage into your generative AI and video workflows.”
Vector database vendors highlight performance gaps
Vendors of specialized vector databases highlight significant performance gaps between their offerings and AWS’s storage-centric approach.
Purpose-built vector database providers incl ConeWeaviate, Qdrant, and Chroma have established production deployments with advanced indexing algorithms, real-time updates, and purpose-built query optimization for latency-sensitive workloads.
Pinecone, for example, does not see Amazon S3 Vectors as a competitive challenge to its vector database.
“Prior to the initial launch of Amazon S3 Vectors, we were properly briefed on the project and did not believe that the cost-performance ratio was directly competitive at mass scale,” Jeff Zhu, vp of product at Pinecone, told VentureBeat. “This is especially true now for our dedicated read nodes, where, for example, our major e-commerce customer recently benchmarked a recommendation use case against 1.4B vectors and achieved 5.7k QPS at 26ms p50 and 60ms p99.”
Analysts are divided on the future of vector databases
The launch reignites the debate on whether vector search stays a standalone product category or becomes a feature enabled by major cloud platforms through storage integration.
“It has been clear for some time that vector is a feature, not a product,” Corey Quinn, chief cloud economist at The Duckbill Group, said in a statement. message on X (formerly Twitter) in response to an inquiry from VentureBeat. “That’s all he says now; the rest will follow.”
Constellation Research analyst Holger Mueller also sees Amazon S3 Vectors as a competitive threat to independent vector database providers.
“Now we go back to the vector providers to make sure they have the edge and are better,” Mueller told VentureBeat. “Packages always win in enterprise software.”
Mueller also highlighted the advantage of AWS’s approach to eliminating data movement. He noted that vectors are a tool that allows LLM to understand enterprise data. The real challenge is creating vectors, which involves how and how often data is moved. By adding vector support to S3, where large amounts of enterprise data are already stored, you’ll be able to solve the problem of data movement.
“CxO likes this approach because no data transfer is needed to create the vectors,” Mueller said.
Gartner’s distinguished vp, analyst Ed Anderson, sees AWS growing with recent services, but he doesn’t expect it to spell the end of vector databases. He noted that organizations using S3 for object storage can increase S3 usage and possibly eliminate the need for dedicated vendor databases. This will increase value for S3 customers while increasing their dependence on S3 storage.
Even with AWS’s growth potential, vector databases are still needed, at least for now.
“Amazon S3 Vectors will be valuable to customers, but it will not eliminate the need for vector databases, especially when use cases require low-latency, high-performance data services,” Anderson told VentureBeat.
AWS itself appears to share this complementary view, while signaling continued performance improvements.
“We are just getting started in terms of the scale and performance of S3 Vectors,” Bukovec said. “Just as we improved read and write performance in S3 for everything from video files to Parquet files, we will do the same for vectors.”
What does this mean for businesses
In addition to the debate over whether vector databases will survive as standalone products, enterprise architects face immediate decisions about how to implement vector storage for production AI workloads.
The performance tiering structure provides a clearer decision path for enterprise architects evaluating vector storage options.
S3 Vectors works for workloads that tolerate 100 ms latency: semantic search on large document collections, agent storage systems, batch evaluation for vector embeddings, and background RAG context retrieval. The economics develop into compelling at scale for organizations that have already invested in AWS infrastructure.
Specialized vector databases are still vital for latency-sensitive use cases: real-time suggestion engines, high-throughput searches supporting 1000’s of simultaneous queries, interactive applications where users wait for results synchronously, and workloads where consistency in performance outweighs cost.
For organizations running each types of workloads, the hybrid approach reflects how enterprises already use data lakes, deploying specialized vector databases for performance-critical queries while using S3 Vectors for large-scale data storage and less time-sensitive operations.
The key query is not whether to replace existing infrastructure, but how to design vector memory at different levels of performance based on workload requirements.