Reflection AI raises $2B to be America’s open frontier AI lab, challenging DeepSeek (techcrunch.com)

Reflection, a year-old startup founded by ex-DeepMind researchers Misha Laskin and Ioannis Antonoglou, closed a $2 billion funding round at an $8 billion valuation — a 15x jump in seven months — and is pitching itself as an “open frontier” AI lab to rival closed outfits like OpenAI/Anthropic and mirror China’s DeepSeek. With ~60 engineers and researchers recruited from DeepMind and OpenAI and backing from investors including Nvidia, Sequoia and GIC, Reflection says it has built a large-scale training stack that can run Mixture-of-Experts (MoE) models and plans to release a frontier LLM next year trained on “tens of trillions” of tokens. The founders highlight prior wins in autonomous coding and reinforcement learning as proof they can scale agentic reasoning outside big tech. Technically and strategically, Reflection’s approach mixes openness and proprietary control: it intends to publish model weights for public use while keeping datasets and full training pipelines private, and will monetize via enterprise deployments and “sovereign AI” contracts. That model could broaden access to frontier architectures (MoEs, RL platforms) while preserving a compute/infrastructure moat — a compromise between pure open-source and closed-lab models. If successful, Reflection could shift competitive dynamics by making frontier weights available to researchers and enterprises in the U.S. and allies, addressing geopolitical and commercial concerns about reliance on non-Western models.