E Merfalen

about two years ago we started a small experiment: what if you could run the best open models without touching a single GPU? it was called Nebius AI Studio. a few models, a playground, an api. Then inference ate the world. Agents, coding tools, apps serving millions of people, all of them hitting a model thousands of times a second. So the experiment grew up. It became @nebiustf, and less than a year after launch it's approaching a multi-hundred-million-dollar run rate. in Q3 alone, contract value grew 17x quarter on quarter. honestly super pumped about what the team shipped this quarter: - new models live on release day, like Kimi K3 and GLM 5.3 - custom speculator training: a small model tuned to your traffic drafts the next tokens and the big model just checks them, so answers come back faster and cost less - First ai cloud to adopt Nvidia Groq 3 LPX, to boost token generation on Vera Rubin - Top tier on artificial analysis for the latest Deepseek, Kimi and GLM models and the best part is who's building on it: - @appliedcompute and @cursive_ai post-trained on our cloud, then moved straight into inference without leaving - @hark_labs runs its computer-use agent on it, where every bit of lag is a frozen screen - @Prosusgroup serves open models across a portfolio of about 1 billion users training is a project. Inference is a bill that arrives every second, forever. that's the part we're building for.