Frontier AI startup Reflection AI, founded in 2024 by two former Google DeepMind researchers, has raised $2 billion in fresh funding at an $8 billion valuation — a fifteenfold increase from its $545 million valuation just seven months ago.
The massive round, backed by major investors including Nvidia, DST, Sequoia, Lightspeed, B Capital, CRV, Citi, GIC, Eric Schmidt, and Eric Yuan, positions Reflection AI as a direct challenger to both closed-source labs like OpenAI and Anthropic, and Chinese frontier AI developers such as DeepSeek.
A new contender in the global AI race
Reflection AI was founded by Misha Laskin, who led reward modeling for DeepMind’s Gemini project, and Ioannis Antonoglou, co-creator of AlphaGo, the first AI system to defeat a world champion in the game of Go.
The company’s mission is clear: to build open frontier AI models that rival the capabilities of the world’s leading proprietary systems, while ensuring technological sovereignty for the U.S. and its allies.
“DeepSeek and Qwen are our wake-up call,” said Laskin, CEO of Reflection AI. “If we don’t act, the global standard of intelligence will be built elsewhere — and not by America. We want to ensure that open, powerful, and transparent AI is led by the West.”
Advertisement
From autonomous coding to general agentic reasoning
Originally focused on autonomous coding agents, Reflection AI has now expanded its vision to general agentic reasoning — developing large-scale, Mixture-of-Experts (MoE) language models that match the complexity and scale of frontier LLMs.
The startup has built its own AI training stack and secured massive compute infrastructure, with plans to release its first frontier model in early 2026. The system will be trained on tens of trillions of tokens — rivaling DeepSeek and OpenAI’s largest models.
“We’ve built what many thought only top labs could create: a large-scale LLM and reinforcement learning platform capable of training massive MoE systems,” Reflection AI wrote in a company post.
The team now numbers around 60 researchers and engineers, mostly drawn from DeepMind and OpenAI, spanning AI architecture, reinforcement learning, and large-scale infrastructure.
Building America’s “open” AI ecosystem
Reflection AI’s approach to openness mirrors that of Meta’s Llama and Mistral, emphasizing access over full transparency. The company plans to release model weights publicly while keeping training datasets and infrastructure proprietary.
“The most impactful thing is access to the model weights,” Laskin explained. “That’s what enables innovation — anyone can experiment, fine-tune, and deploy.”
This balance forms the foundation of Reflection AI’s business model. Researchers and developers will have free access to its open models, while enterprises and governments will pay for deployment, customization, and sovereign AI infrastructure.
“Enterprises want ownership — they want to run AI on their infrastructure, optimize costs, and customize it for their workloads,” said Laskin. “That’s the market we’re serving.”
A growing open AI movement
The announcement has drawn praise across the U.S. tech ecosystem.
David Sacks, the White House AI & Crypto Advisor, commented:
“It’s great to see more open-source AI models from American labs. A meaningful segment of the global market wants the cost, control, and flexibility open AI offers. The U.S. should lead this category.”
Clem Delangue, CEO of Hugging Face, added:
“This is great news for open AI in the U.S. The next challenge will be to maintain high velocity in sharing models and datasets — the hallmark of truly open innovation.”
About Reflection AI
Reflection AI is building America’s first open frontier AI lab, developing large-scale foundation models using Mixture-of-Experts and reinforcement learning architectures. Its mission is to democratize access to frontier intelligence while ensuring technological sovereignty for enterprises and nations.
The company has raised $2 billion in Series B funding from leading investors, valuing it at $8 billion. Its first frontier model is expected to launch in early 2026.
