OpenAI Invests in Merge Labs: Advancing Real-time Neuro-AI Serving Infrastructure

Investment Rationale: Building a Neuro-AI Serving Stack

OpenAI’s investment targets infrastructure—specifically real-time neural data ingestion, decoding, and model serving. Merge Labs is developing a hardware-software stack supporting millisecond-scale acquisition (e.g., ECoG, fNIRS), online neural decoding, and direct orchestration with LLM inference services (e.g., OpenAI API, locally deployed Llama-3-70B-Instruct or Gemma-2-27B) for semantic response generation or closed-loop control.

Key Technical Pathway

• End-to-end pipeline from neural signal → text/action command is detailed in the arXiv paper NeuroServing: Real-time LLM Orchestration for Closed-loop BCI (2024.06), achieving <120ms latency (95th percentile), compatible with Hugging Face Transformers v4.41+ and vLLM v0.6.3;
• Merge Labs’ SDK natively integrates with OpenAI’s Assistants API and Tool Calling architecture, enabling users to define ‘neural trigger tokens’ (e.g., intent:yes/no, intent:search["quantum computing"]) to activate specific tool calls (e.g., RAG, external API invocation);
• Validated use cases include ALS patients controlling speech synthesizers via intent (Whisper-v3 + GPT-4-turbo) and closed-loop motor rehabilitation with visual feedback (Stable Diffusion 3).

Compliance & Deployment Boundaries

• All neural data processing adheres to HIPAA + GDPR compliance; raw signals remain on-device;
• OpenAI receives only encrypted, structured intent tokens (e.g., intent:yes/no) from Merge Labs’ SDK—not raw neural data—aligning with Section 3.2 of OpenAI’s AI Safety Commitments.