Qdrant Cloud Adds GPU Indexing, Multi-AZ Clusters and Audit Logs

What Qdrant shipped

On April 28, 2026, Berlin-based Qdrant added three features to its managed vector database service, Qdrant Cloud: GPU-accelerated indexing, Multi-AZ clusters and audit logging.24 All three were available to customers the day they were announced.24 SiliconANGLE, TechTarget, BigDATAwire and Qdrant's own blog all described the release the same way: it is meant to make the product ready for production on three fronts, performance, availability and compliance, as retrieval-augmented generation (RAG) and AI agents become central to how companies run.212226

None of the three features is a new idea. The interesting part is what the bundle says about where vector databases now compete. Speed on benchmarks is no longer enough. Buyers want reliability and audit controls of the kind they expect from a traditional database, and specialists like Qdrant are being judged by those standards.

GPUs move from inference to index building

The headline feature uses GPUs to build HNSW (Hierarchical Navigable Small World) indexes, the graph structures that let a vector database run fast similarity searches without comparing a query to every stored vector.21 CEO Andre Zayarni summed it up by saying GPUs are useful for indexing, not just model inference. He also pointed out that Qdrant's open-source engine has supported GPU-built HNSW indexes since version 1.13.24

Qdrant's own benchmarks put index builds on dedicated cloud GPUs at up to four times faster.24 Each node in a GPU-enabled cluster gets its own GPU. The setup requires at least 16 GiB of RAM per node and runs on NVIDIA T4 cards.26 Customers can also add GPUs to existing clusters to handle bursts of heavy indexing. For now the feature is only available on AWS, with other clouds promised later.24

The speed figures don't line up across Qdrant's own announcements. When the company introduced GPU indexing in its open-source engine in early 2025, it said builds could be up to 10x faster. The cloud launch says up to 4x.24 The two numbers probably reflect different hardware and test conditions, but buyers should not take either as a guarantee. The T4 is an older inference card, so the cloud figure is likely tied to that choice of hardware.

Rivals also claim bigger gains. One summary of the launch notes that Milvus has reported index builds up to 17x faster in some configurations, and that Qdrant's GPU work is built on the vendor-neutral Vulkan API rather than tied to NVIDIA.25 That is a fair point about avoiding lock-in. Still, the cloud product as shipped runs on NVIDIA T4s on AWS.26 For now, hardware neutrality matters more to people running the open-source engine themselves than to cloud customers.

Why faster indexing matters for streaming workloads

The main reason to speed up indexing is data that changes all the time. Qdrant said the feature targets write-heavy workloads such as dynamic content catalogs, real-time recommendation systems and agent applications.26 Bastian Hofmann, Qdrant's head of product, told TechTarget that as datasets and write volumes have grown, CPU-only indexing no longer keeps up for some workloads. Moving GPU indexing into the cloud service means customers don't have to run GPU infrastructure themselves.22

The same theme runs through Qdrant's open-source releases. Version 1.17, released in February, let users avoid creating large unoptimized segments under heavy write load. It also added "delayed fan-outs," which send a query to a second replica if the first one is slow, to cut worst-case latency.20 In May, version 1.18.1 reportedly began rejecting asynchronous upserts with the wrong vector dimension before they reach the write-ahead log. Previously those errors failed silently later on, during replication.16 Taken together, these changes suggest Qdrant expects its database to sit inside continuous data pipelines rather than being loaded in occasional batches.

Other vendors are moving the same way. MongoDB Atlas previewed automated embedding in May, which re-embeds fields when documents change so teams don't need a separate sync pipeline.16 Across the market, keeping vectors fresh is becoming as important a selling point as query speed.

Availability without failover

Multi-AZ clusters copy data across three availability zones in one cloud region. Qdrant stresses that this is ongoing cross-zone replication, not failover. If a zone goes down, the surviving zones keep handling reads and writes with no delay and nothing for the customer to do.24 The option is only offered on the Premium Multi-AZ tier, which carries uptime commitments of up to 99.95%.24 One limitation appears only in Qdrant's own documentation: Multi-AZ has to be chosen when a cluster is created and can't be switched on later.26 That matters for teams that want to start small and add redundancy as they grow.

Audit trails for autonomous systems

Audit logging records every operation made through the Qdrant API, including queries, upserts, deletes, collection management and snapshots. Each entry is a structured JSON record showing which user and API key made the call, when, against which collection, and whether it was allowed or denied.24 It is available on every paid cluster. Logs rotate automatically and can be pulled through an endpoint into security information and event management (SIEM) tools.26

Qdrant explicitly ties the feature to AI agents. When an autonomous system acts on data it retrieved, the log shows which service queried which collection, and whether it was authorized to.26 Hofmann said both Multi-AZ and audit logging came straight from customer requests, because vector search now sits in the critical path of their applications.22 The feature has since been extended beyond the managed cloud: Qdrant's private-cloud operator added audit-log configuration in May.23

Analysts agree on one thing

The analysts quoted in the coverage broadly agreed on how much each feature matters. IDC's Devin Pratt called the release a step toward making Qdrant Cloud more production-ready. He said faster indexing is the part that sets it apart, while high availability and audit logs are fast becoming things enterprises simply expect.22 Kevin Petrie of BARC U.S. agreed that the features help Qdrant stand out, but warned that vector search on its own has not been enough for RAG pipelines. In his view, more retrieval methods would separate Qdrant from rivals more than the infrastructure features do.22

This article reads the release the same way. Multi-AZ and audit logs keep Qdrant level with competitors, and GPU indexing is the one feature that could put it ahead. The launch fits a pattern that goes back years. In March 2025 Qdrant added single sign-on, role-based access control, a cloud API and Prometheus monitoring, and in July 2025 it added built-in embedding inference.28

What the Postgres comparison shows

All of this plays out against a challenge from Postgres, which this coverage treats as the bigger competitive story. Comparison guides now mostly agree on a simple rule: teams already running Postgres should use the pgvector extension and switch to a dedicated engine like Qdrant only when vector count, query volume or filtering complexity outgrow it.14 The reason is cost of operations, not just speed. With pgvector, vectors share transactions, joins and backups with the rest of the application's data.10

The benchmark data is contested. A widely shared result had Postgres with Timescale's pgvectorscale extension handling 471 queries per second at 99% recall on 50 million vectors, against 41.47 for Qdrant.25 One guide went so far as to say "Postgres won."5 Other tests show the reverse at smaller scales, with Qdrant about two to three times faster on median latency at a million vectors.4 An August test of Qdrant 1.19 against pgvector 0.8.2 found pgvector matching or beating Qdrant on per-query latency at small scale.7 One comparison concluded that no neutral, current, like-for-like benchmark of the two exists.1 That is probably the most accurate summary.

There is more agreement on where Qdrant is clearly ahead: heavily filtered search, especially multi-tenant setups where every query is limited to one customer's data.10 Teams moving off pgvector also had a security reason to pay attention this year. A May patch for CVE-2026-3172, a buffer overflow that risked exposing data across tables, had to be applied urgently.16 One benchmark noted that Qdrant ships with authentication turned off by default, so neither option is safe without proper configuration.7

What has happened since

The engine has kept improving since April. Version 1.19 added a 4-bit TurboQuant storage format that keeps no full-precision copy of vectors, cutting storage sharply, along with unified memory tiers and tenant-specific BM25 statistics.13 In September, Qdrant and Vultr released FineWeb-10B, an open benchmark of about 10 billion vectors and 24.47 TB, together with the Supernova benchmarking framework.1415 By October, the private-cloud operator was adding global quota enforcement and fixing version parsing for GPU images.23

The April cloud launch was mostly about running operations at scale. The open-source releases since then have been about compression and scale. Both are a response to the same pressure: Postgres keeps getting better at vector search, so Qdrant has to show it wins clearly once workloads grow past a certain size. Its 10-billion-vector benchmark is an attempt to change where the comparison happens, toward the scale where its approach is strongest.