How they use Nebius

Brave uses Nebius inference for Answer with AI in Brave Search. In its customer post, Nebius described Brave using low-level access to its compute for inference. Nebius describes a stack using vLLM, Terraform for provisioning and HAProxy for load balancing. Its customer collection reports 10–70B-parameter models and nearly 100% compute utilization. The May 2025 digest reported more than 11 million AI-generated answers daily.

At Inflection 2026, Nebius described Brave moving from self-hosted infrastructure to managed inference, with custom model quantisation and sharding. The presentation reported 16 million daily AI-summary queries at the June event. These figures describe query volumes at the time of the respective statements.

Brave’s privacy notice also identifies Nebius Token Factory as the service used by Ask Brave, which generates answers with source references. Brave says it enables Zero Data Retention, so Nebius does not store user prompts or responses after processing or use them for model training. The notice also says Nebius never sees users’ IP addresses.