Speculum AI

Minima AI, Inc. (mnma.ai) is a US-based LLM inference optimization startup founded by ex-Atlassian AI engineer Sergii Kozyrev and AI systems architect Davyd Maiboroda. It builds a compression and serving stack spanning model weights, KV-cache and custom inference kernels, offered as an on-prem optimization runtime (mnma Engine), an OpenAI-compatible hosted inference API, and an enterprise AI agent product.

RadixArk, founded in 2025 in Palo Alto, California, USA, develops AI systems around the SGLang inference engine and Miles reinforcement learning framework, improving model inference and training efficiency by reducing compute and memory requirements.

Infinity, founded in August 2025 by former Google Brain researcher and AGI House co-founder Jeremy Nixon, builds a universal inference software layer for heterogeneous AI chips. Its autonomous agent Ignition automatically generates, tests and optimizes the low-level compute kernels that determine how efficiently a chip runs AI models — positioned as a CUDA alternative — and already works with chipmakers such as d-Matrix with several million dollars in ARR.

Loophole Labs (founded 2021, New York) builds virtualization infrastructure (Substrate hypervisor, Architect) for zero-downtime live migration and hibernate/wake of stateful Kubernetes workloads, plus AGX, a microVM sandbox platform for AI agents. Backed by Amplify Partners, Cervin Ventures and Peak XV; acquired by LiveKit on Sept 24, 2026.

Inference.net is a San Francisco-based AI inference infrastructure company. It began in early 2024 as Kuzco, a Solana-based decentralized network aggregating idle GPU compute for LLM inference, and rebranded in mid-2024. It now offers an LLM lifecycle product suite (Deploy, Observe, Trace, Train, Evaluate) that lets AI-native teams run open-source and custom models at lower cost than proprietary APIs.

Autoloops provides inference infrastructure for voice-agent builders: serverless APIs and on-demand GPU clusters serving open-source LLMs such as Gemma at realtime speed. Its proprietary Hanoi engine targets low-latency voice workloads, claiming 2–2.5x lower p50/p95 latency than DeepInfra Turbo and tuned vLLM at high concurrency. Public case study: voice-AI testing platform Cekura. Pitched at GMI Cloud's SCALE accelerator Cohort 2 Demo Day (2026-10-23).