Nebius

The ultimate cloud for AI innovators. For GenAI open-source model endpoints, check out @nebiustf.

[Nebius @nebiusai]

"I'm going to give you a tagline: No tagline needed!" - @JensenHuang, CEO, @nvidia #NVIDIAGTC Berlin

[waqas @waqasmakhdum]

$ python -c "import this"

Thank you, Tim.

[Nebius @nebiusai]

How far has physical AI come? Marc Tuscher, Co-Founder & CTO at @SereactAI, discusses once-unimaginable capabilities and what it takes to put robots to work reliably.

Watch the full conversation from NEBIUS LIVE. https://t.co/7a1FdfiiJB

[Nebius @nebiusai]

3/ Data curation, simulation, synthetic data, policy training, evaluation and observability, all through one CLI.

3/ Apache 2.0. Issues and PRs welcome: https://t.co/ENA8fpWnUQ

[Nebius @nebiusai]

Nebius Physical AI is our open-source CLI and SDK for teaching robots on Nebius AI Cloud

For example:
🤖 a robot arm picking something up,
🧬 a humanoid walking,
⌛️ a quadruped running.

[Nebius @nebiusai]

2/ When you’re done, tear it down. Skipping a step leaves something billing, so npa cleanup prints the exact runbook for your machine, and npa destroy stays read-only until you add --yes.

[Nebius @nebiusai]

@guy_drt75 So happy to partner with you, @Photoroom.

[Nebius @nebiusai]

Marc Boroditsky, Chief Revenue Officer at Nebius, discusses the growing appetite for next-generation compute.

With a gigawatt of @nvidia Vera Rubin NVL72 capacity planned for next year, Nebius is already seeing strong demand.

Watch the full conversation from NEBIUS LIVE.https://t.co/a6tCxJ1M8m

[Nebius @nebiusai]

Watch the full interview here - https://t.co/cuuTLk9IsG

[Nebius @nebiusai]

Our SVP of Applied AI, Ryan Wang explains the importance of KV Cache management in enabling LLM Inferencing work efficiently at scale.

For starters, When an autoregressive LLM generates new text, it looks at all the previous tokens to predict the next one. And to do that, it turns each token into a (K) key (what it’s about) and a (V) value (what it carries).

Since the past keys and values do not change, the system stores them instead of recomputing them, And, That is KV cache.

[Nebius @nebiusai]

3/3 Also, that's why it is important to design your workload to maximize reuse. Especially when there is an app or a harness, avoid changing your prompt structure at the beginning of a query; add new instructions or tasks at the end instead to ensure you keep the existing cache intact and to keep inference costs low.

[Nebius @nebiusai]

We’re bringing AI inference to #KubeCon + CloudNativeCon North America! Join our session on Nov 10 to learn how web tools, vision models, and distributed inference come together in a live application and meet the team at our booth #133 in Hall B.

https://t.co/CpE0b9ZA7i

[Nebius @nebiusai]

We're live!

Build your KubeCon survival kit and steal our tips and tricks!

Join @kunalstwt and @devopsjacquie on Build With Nebius as we build a session planner together and share first-timer tips, food picks, and events to catch. https://t.co/FoR28jKpML

[Nebius @nebiusai]

“If you’re not paying attention, you’re missing an opportunity.”

Compute Exchange CEO & Silicon Data Founder @carmenli and Nebius CRO Marc Boroditsky explain why live spot pricing matters. When demand varies across GPUs, transparent prices help companies see which capacity fits their needs.

Watch the full conversation on NEBIUS LIVE at NVIDIA GTC Berlin. https://t.co/iLy84IjYzM

[Nebius @nebiusai]

Read more: https://t.co/OJoIlYjXrc

[Nebius @nebiusai]

Our team measured NUMA Affinity's impact on a DeepSeek V3 pretraining workload running with Megatron Bridge on 192 NVIDIA B300 GPUs in Managed Service for Kubernetes on Nebius AI Cloud.

Exclusive NUMA binding reduced mean training step time by 1.11% and increased model throughput per GPU by 1.13%.

In this post, we’ll explain how NUMA-aware process affinity works, compare it with the controls Slurm users already have, and walk through the measurement.

[Nebius Token Factory @nebiustf]

Up to 5× faster RL training steps, on the same GPU budgets.

Our internal auto-tuner optimizes the full RL loop: training, inference, and environment execution. It profiles each job, forecasts promising configuration changes, tests them in real runs, and follows the bottleneck as it moves.

Across 210 runs on NVIDIA B300 GPUs, we tested four dense and MoE models on math and agentic coding workloads. All eight jobs improved: 1.42-4.93x faster steps, with a median speed-up of about 3.5x.

One revealing result: nearly doubling inference-engine throughput in a benchmark didn’t shorten the RL step. Sandbox execution absorbed the gain.

That’s why we measure the complete job, alongside numerical checks and longer reward comparisons. GPU counts stayed fixed; some configurations added sandbox hosts.

We’ve shared the system, results, and five engineering lessons, including what didn’t work, in the article below: