Turbopuffer reports $100M run-rate and a 95% Cursor code-search cost cut
Turbopuffer said it crossed a $100M run-rate while staying profitable on less than $1M raised, and said Cursor moved production search onto the stack with a 95% cost reduction. The milestone matters because AI products increasingly compete on retrieval quality and cost, not just model output.
TL;DR
- Sirupsen's milestone post says turbopuffer crossed a $100 million run-rate in March, reached profitability, and did it on less than $1 million raised.
- According to Sirupsen's Cursor example, Cursor moved production search onto turbopuffer days after requesting two features, cutting cost by 95 percent and now searching more than 1 trillion code chunks.
- Sirupsen's origin post ties the company to a specific infra thesis: a search engine built on object storage, which he says became newly viable and was aimed at being 10x cheaper.
- In Sirupsen's workload list, the company frames retrieval as the bottleneck for AI products that need the right context, across inputs ranging from PDFs and commits to video frames and memories.
You can jump from turbopuffer's About page to the Cursor migration claim, and then to a separate post about SID-1 training on turbopuffer that hints at the scale profile the company is pitching: bursty, retrieval-heavy AI systems rather than ordinary app search.
$100M run-rate
Turbopuffer's headline claim is blunt: $100 million run-rate in March, profitable, and less than $1 million raised. The customer list in the same post includes Cursor, Anthropic, Notion, Cognition, Harvey, Bridgewater, Ramp, Linear, Legora, Superhuman, Atlassian, and Granola.
That is an unusual combination for infra. Revenue scale, profitability, and almost no outside capital usually do not show up together in database startups, which is why this post traveled fast.
Cursor's code search
The most concrete operating number in the thread is Cursor's migration story. Sirupsen says turbopuffer launched in October 2023, Cursor asked for two missing features, then moved its production workload over a few days later.
The claimed result was a 95 percent cost reduction, and the current scale claim is bigger: Cursor now searches more than 1 trillion chunks of code on the system.
Gergely Orosz adds one extra detail not in the main thread, writing that Cursor was in hypergrowth and that AWS Aurora could not handle its scale. That is still commentary, but it sharpens what kind of migration this was supposed to be.
Object storage search
Sirupsen says the company started from a blog post and an architectural bet: build a search engine on object storage, and make it 10x cheaper than existing options. He ties that idea directly to eight years of infrastructure and incident experience at Shopify.
The thread's product framing is narrower than generic vector database marketing. It is about feeding AI systems the right context cheaply enough to stay in the loop, which Sirupsen's retrieval framing describes as the difference between AI that works and AI that misses data.
Retrieval workloads
The workload inventory is the clearest map of what turbopuffer thinks it sells into:
- web pages
- transcripts
- paragraphs
- PDFs
- satellite images
- fraudulent transactions
- video frames
- papers
- commits
- memories
- attachments
- products
A separate company account post adds one more data point: SID-1, described there as an agentic search model, was trained on turbopuffer with bursts above 1,000 QPS over corpora larger than 10 million documents. That is new information beyond the revenue thread, and it points at the same pitch from another angle: retrieval as a high-throughput training and inference primitive, not just an app feature.