Reduces cost of vector + full text storage and search by 90% or more
Manages over two trillion documents
Handles more than 10 million writes/second and 10,000 queries/second
turbopuffer, an Ottawa-based start-up, is a serverless vector and full-text search engine built on Google Cloud to help AI businesses search every byte of data without the high costs and complexity of traditional search architectures.
turbopuffer, an Ottawa-based start-up, is a serverless vector and full-text search engine built on Google Cloud to help AI businesses search every byte of data without the high costs and complexity of traditional search architectures.
Connecting AI to massive amounts of data
With Google Cloud, our customers can connect AI to massive amounts of data while saving millions.
Simon Hørup Eskildsen
Co-Founder and CEO, turbopuffer
Generative AI has the potential to transform businesses, but only when AI can access the right data. That's where semantic search engines come in.
These systems store data as high-dimensional numerical representations called vectors, enabling semantic search and Retrieval-Augmented Generation (RAG), both key components of generative AI applications. The challenge? Traditional search engines use storage architectures with elevated price points, forcing you to choose between indexing all your data and keeping costs low.
turbopuffer turns this model on its head. Its serverless search engine uses a storage architecture that is 90% less expensive than traditional approaches. Great news for AI innovators, including Cursor, Notion, and Linear, who use the technology to handle massive volumes of data at scale.
How did the turbopuffer team solve this challenge? Most search engines rely on high-performance memory that serves data with incredible speed, but at considerable storage costs. Instead, turbopuffer designed a serverless vector and full-text search database built from the ground up using object storage on the cloud. As co-founder and CEO Simon Hørup Eskildsen evaluated available platforms, Google Cloud stood out.
"When we looked at how to store and access petabytes of data at the lowest possible cost, Google Cloud was not just a strategic preference, it was the only engineering option for building a system that matched our customers' performance needs," he says.
Low-cost infrastructure, elite performance
turbopuffer's architecture is a masterclass in efficiency and simplicity, proving that you don't need complicated infrastructure to achieve elite performance. The company's unique architecture takes advantage of Google Cloud Storage features like a Compare-and-Swap API that ensures data integrity and consistency in distributed systems. This allows turbopuffer to deliver fast, strongly consistent query speeds on massive datasets without the prohibitive cost associated with traditional memory.
"Instead of an arrangement where multiple servers coordinate through a separate database to manage access to cloud storage, turbopuffer's only dependency is Google Cloud Storage itself," says Eskildsen.
Alongside Google Cloud Storage, Google Kubernetes Engine (GKE) simplifies containerized application management and Google Compute Engine launches virtual machines on demand.
The latest generation of Google Cloud servers powered by Google's ARM-based C4A compute nodes, which offer market-leading price performance, were another draw. "By combining ultra-low-cost storage with optimized, stateless compute, we can offer a solution that is fundamentally cheaper to operate than our competitors," says Eskildsen.
As a result, turbopuffer operates at a phenomenal scale, managing over two trillion documents and counting. The system is optimized for the unique demands of AI search workloads: high write throughput, cheap storage, and fast vector and full-text search queries. The search engine routinely handles 10 million writes per second and serves over 10,000 queries per second. Cold queries which read directly from object storage take less than half a second, but that drops to under 10 milliseconds for hot documents in the cache.
Choosing Google Cloud was a matter of engineering necessity. It means that we can offer a vector database that is both affordable and performant.
Simon Hørup Eskildsen
Co-Founder and CEO, turbopuffer
Powering the next wave of AI innovation
With Google Cloud, we've solved a technical challenge that makes searching every byte more affordable for organizations building the new AI economy.
Simon Hørup Eskildsen
Co-Founder and CEO, turbopuffer
By reinventing search from first principles, turbopuffer enables both innovative start-ups and large enterprises to scale new AI features fast and pass on operational and financial efficiencies to their customers.
The results speak for themselves. Notion, the ground-breaking popular workspace platform, manages a vast amount of internal company information, primarily text, for millions of users and hundreds of thousands of businesses. Since migrating to turbopuffer, the business has seen an 80% reduction in costs, enabling it to remove per-user AI charges.
Coding disruptor Cursor was another early adopter. turbopuffer powers code retrieval features in Cursor to populate the context window with code searches when appropriate. Since switching its vector database to turbopuffer, Cursor has saved 95% on data storage and retrieval, while supporting more than one million writes per second for faster migrations and indexing. This move also stabilized their per-user costs, ensuring they can continue to scale and use AI resources without the fear of ballooning expenses.
Another client, the software development platform Linear, was drawn to turbopuffer for its ability to ingest hundreds of millions of full-text search documents and vectors without having to think about machine types. Although cost wasn't the primary motivation for the move, the business has since achieved 70% savings in data storage and retrieval.
A partnership for growth
Beyond providing infrastructure, Google Cloud's comprehensive support was a key factor in turbopuffer's success and its ability to deliver game-changing results for its clients.
Eskildsen recalls that as the business scaled, it quickly ran into a CPU bottleneck. Despite being a small start-up at the time, the Google Cloud team provided crucial support. "Product Managers (PMs) at Google consistently provided critical information on upcoming features and strategic advice on what would be beneficial for the company's future," Eskildsen says.
This partnership was vital in helping turbopuffer overcome its early scaling obstacles. Since its launch, the platform has evolved from a simple vector search tool into a full-fledged search engine, adding capabilities like full-text search, aggregations, and various other query types.
Google Cloud remains at the core of the business. "We didn't know that this architecture would scale as far as it has, but it has only increased our conviction in doing absolutely everything on Google Cloud Storage with no other databases involved," says Eskildsen.
Most of all, turbopuffer's success sends a powerful message for the entire industry. The future of AI isn't just about bigger models and more data, it's about smarter, more efficient infrastructure that makes the power of AI accessible to everyone.
By building its platform on a foundation of low-cost, high-performance Google Cloud services, turbopuffer solved a major technical challenge and created a new economic model for AI that will benefit potentially thousands of start-ups.
turbopuffer is a serverless vector and full-text search engine built from first principles on object storage. It is designed to be a fast, scalable, and cost-efficient database, especially for AI applications.
Industries: Technology, Startups
Location: Canada
Products: Google Cloud, Google Cloud Storage, Compute Engine, Google Kubernetes Engine (GKE)
