Turbopuffer, a startup that built its reputation on fast vector search for AI applications, pulled off a quiet architectural about-face this fall. On September 30, the company announced that its long-standing primary feature—the vector index—would become secondary, repositioned beneath a new general-purpose search engine designed to handle full-text queries, aggregations, and attribute filters alongside embeddings.
It's a pivot that says as much about where the database market is heading as it does about Turbopuffer's own ambitions. The company insists the redesign will preserve vector-search speed while unlocking performance gains elsewhere, but the move also signals a tacit acknowledgment: pure-play vector databases may no longer be enough.
Dan Harrison, an engineer at Turbopuffer, sketched out the rationale in a blog post. The old architecture, he explained, treated approximate-nearest-neighbor vector indexing as the storage key—a choice that created headaches as customers loaded documents with multiple embeddings. Storage bloated. Writes cascaded through cluster rebalancing. Block sizes that worked well for vectors choked on simpler queries. "We're in the process of moving to a new primary index," Harrison wrote, "and making ANN 'just another' secondary index."
Version 3 of the engine hit a 100% continuous-integration pass rate on September 5, though the company readily admits performance still trails its production system. One early win: batched reads accelerated full-text search roughly ninefold in September 8 tests, according to an October 2 update on the development tracker. The company says it will release comparative benchmarks in the coming weeks, pitting v3 against the current engine across vector search, hybrid queries, and attribute-based ordering. The nightly suite runs on Google Cloud's us-central1 region, using Cohere Wiki embeddings for vector workloads and MS MARCO for text and hybrid tests.
The shift mirrors a wider convergence across the search and database landscape. Elasticsearch folded hybrid retrieval into ES|QL earlier this year. Milvus unveiled Storage v3—a columnar object-storage layer it calls "Vector Lakebase"—over the summer. ByteDance published research in August describing UBASE, an internal platform that evolved from text search into what the company terms "unified AI search," blending vector retrieval, lexical matching, and predicate filtering in a single engine. Vespa, meanwhile, has leaned into hybrid top-k operators and reciprocal-rank-fusion reranking. Even PostgreSQL is seeing integrated vector-plus-SQL engines emerge from academic labs.
When Turbopuffer's announcement hit Hacker News on October 1, several commenters noted the company now finds itself in a different competitive set, one that includes Elasticsearch and Vespa rather than specialized vector stores alone.
The customer roster offers a glimpse of the scale Turbopuffer is operating at, though some figures remain hard to pin down. Cursor, Notion, Anthropic, Linear, Ramp, and Superhuman all appear on the client list. Mickey Liu, an engineer at Notion, said at a Data Council talk that his company has indexed more than 10 billion chunks and writes roughly 15 billion embeddings monthly, with peak query latencies hovering between 50 and 70 milliseconds. Cursor reportedly stores over 1 trillion code chunks on the platform, according to a June LinkedIn post, and claims a 95% cost reduction after migrating. TELUS uses Turbopuffer to index more than 25,000 AI copilots.
The company's limits page cites some eye-watering numbers: 100 billion vectors in a single namespace occupying 200 terabytes, more than 1 trillion documents globally spanning over 3 petabytes, and peak throughput exceeding 10 million writes per second at 32 gigabytes per second. Turbopuffer declined to clarify whether those represent simultaneous loads or separate high-water marks.
Simon Hørup Eskildsen, a Shopify infrastructure alumnus, and Justine Li founded Turbopuffer in 2023. The startup has raised a modest sum in primary capital—well under typical seed amounts for enterprise infrastructure plays, according to statements cited by the research firm Sacra. Lachy Groom led an initial investment, and Thrive Capital joined a seed round, per Crunchbase and company disclosures. A June LinkedIn post indicated the company reached a nine-figure annual run rate earlier in the year and operates with a 34-person team, according to company statements.
Pricing cuts have been aggressive: the company slashed query costs by as much as 94% in February 2026 and dropped its minimum monthly plan from $64 to $16 in June 2026.
The technical foundation rests on object storage rather than traditional compute-plus-database clusters. "You could turn off all the servers… we would not lose any data because we are… all in on object storage," Eskildsen said in a March podcast with Latent Space. Amazon S3's shift to strong consistency in December 2020 made the architecture viable, he noted.
Harrison said detailed v3 benchmarks and a performance deep-dive on the batched-read optimization should arrive in the next few weeks. Whether the new engine delivers on its dual promise—faster non-vector queries without sacrificing vector speed—will determine if Turbopuffer's gamble pays off, or if the company simply traded one set of constraints for another.
