The ground-truth bill is the part of this you can check with arithmetic. Exact top-1,000 results for 100,000 queries against 10.07 billion vectors is 1.007e15 distance computations [16], which is where Qdrant's "more than one quadrillion" figure lands [8]. That arithmetic tells you the method: no index and no approximation, one exhaustive pass over the corpus per query, on GPUs [10]. It is the only way a recall number means anything, because recall is measured against the exact answer. Vultr supplied the machines and ran the embedding pipeline, and Hugging Face donated hosting plus a storage grant [9]. Those costs sit with Vultr and Hugging Face, not with whoever tries to reproduce the run.
Read the dataset card before the announcement. A 768-dimensional float32 vector is 3,072 bytes, so 10.07 billion of them would be 30.9 TB, and Qdrant reports 24.47 TB of vector data [17][4]. So the 10.07 billion counts dense and sparse together, roughly one dense and one sparse representation per source document [3], or the dense side is stored in something narrower than fp32. Either reading changes what "ten billion" means when you size the hardware to hold it.
The same care applies to the query count. The extra 19,953 queries on the card are 16.6% of the total load [18], and the two filtered sets, which are the ones that exercise the interaction between the index and the predicate, are 8.3% [19]. A recall figure only means something once it names its query set.
For a FineWeb-10B latency number to predict your latency, three things have to hold: your embeddings are 768-dimensional and come from a model with geometry similar to gte-multilingual-base [3], your filter selectivity resembles the text- and structured-filtered sets [7], and your corpus is about as static as a web crawl snapshot. Qdrant concedes the dataset will age as embedding models and retrieval techniques change [21]. The write path is where I would push. Qdrant's own argument is that production systems must index changing data and hold tail latency across billions of records [14], and the harness tracks ingestion throughput and p50/p95/p99 latency as metrics [11]; whether it tracks them concurrently is the thing to confirm in the YAML before you quote a p99.
The leverage sits in one line of the release: the load and stress tools support Milvus and Elasticsearch alongside Qdrant [11]. Workloads are YAML files split across distributed workers with SkyPilot provisioning [10]. That makes "run this config and publish the p99" a reasonable thing to put in an evaluation plan, rather than a favour. Qdrant has an obvious interest in the answer. Runtimewire reads the release as a bid to set a standard aligned with Qdrant's own engineering strengths [13], funded in part by the $50 million Series B that AVP led on 12 March [12]. That interest doesn't make the test dishonest, it makes it a test worth running yourself, at roughly 100 times the 100-million ceiling where Qdrant says most vector benchmarks stop [5][20].
