Interview

Reflection AI signs $1B compute deal with Nebius to train frontier open models, valued at $8B

Jul 14, 2026 with Ioannis Antonoglou

Key Points

  • Reflection AI signs $1 billion compute deal with Nebius to train frontier open-source models, valuing the company at $8 billion.
  • Co-founder Ioannis Antonoglou, a DeepMind founding engineer, argues that US open labs can compete without export-control workarounds by leveraging talent from closed frontier labs.
  • Reflection frames regulatory risk as geopolitical: compliance that slows domestic open-source development without equivalent international pressure could advantage rivals.

Reflection AI has signed a $1 billion compute deal with Nebius to fuel frontier open model training, with the company valued at $8 billion. Ioannis Antonoglou, co-founder and CTO, says the deal is straightforward: building at the frontier requires two things, exceptional talent and compute, and the Nebius partnership delivers the latter.

Antonoglou's background anchors the credibility claim. He was among DeepMind's founding engineers, working across DQN, AlphaGo, AlphaZero, and MuZero, and led RLHF for Gemini before leaving with co-founder Misha Laskin to start Reflection. The pitch is that many of the best people in the industry have joined them specifically to work on frontier open models.

Building Frontier open models requires two things — a lot of incredible talent, and we've been extremely privileged with the fact that many of the best people in the industry have actually joined us. At the same time, the other thing that you need is compute. So to this end, we actually signed this billion dollar deal with Nebius to just get the compute that we need.

Open vs. Chinese labs

On competition with Chinese open-source labs, Antonoglou acknowledges the picture is opaque. He says US labs benefit from direct proximity to frontier closed-lab talent, many of whom have joined Reflection. Chinese labs, he argues, may have bypassed export controls through third-party compute access and, according to accusations from US labs he stops short of confirming, have conducted distillation from proprietary models at industrial scale. His position is that competing without those shortcuts is sustainable: the closed frontier labs never cheated and still lead, so the know-how and talent pipeline that created them can do the same for open models.

Regulatory positioning

Demis Hasabis, Antonoglou's former colleague at DeepMind, is separately advocating for a US frontier AI standards body that would include open-source model testing. Antonoglou says Reflection will engage with whatever compliance is required, but frames the stakes clearly: frontier open models are the foundation on which US research institutions, developers, and early-stage startups build. Any regulatory structure that slows domestic open-source development without applying equivalent pressure internationally risks inadvertently advantaging geopolitical rivals.

The core bet is that closing the gap between open and closed frontier intelligence is achievable, and that the Nebius compute deal gives Reflection the fuel to get there.

Every deal, every interview. 5 minutes.

TBPN Digest delivers summaries of the latest fundraises, interviews and tech news from TBPN, every weekday.