Nebius

AI is shortening production timelines. According to Alex Mashrabov, that is only stage one. The bigger shift begins when AI changes how original entertainment is developed. Social platforms can serve as testing grounds for AI-native concepts, giving creators real-time signals about what audiences want and which ideas have the potential to become larger franchises. Alex predicts that AI-native intellectual property will emerge by the end of 2026. His bigger prediction is that the next Netflix-scale media company will be AI-first. On the Nebius for Startups podcast, the Higgsfield AI CEO and co-founder explains what is already driving this shift and where he believes the media industry goes next. Watch the full episode: https://nebius.com/podcast

Congratulations to Higgsfield AI on its $400 million Series B, led by DST Global, at a $5.4 billion valuation. The numbers are remarkable. Just 16 months after launch, Higgsfield has reached $700 million in annualized revenue. We’re proud to support the infrastructure behind Higgsfield’s training and inference as the team scales this new creative medium globally. Congratulations to the entire Higgsfield team. The next chapter is going to be a big one. 🎬   

Before Ray Summit kicks off, the tokens and drinks are on us. Nebius and LangChain are co-hosting an intimate happy hour in San Francisco for people shipping models to production. No agenda. No slides. Just good conversation with people who get it. Plus, an early preview of Ryan Hanrui Wang's Ray Summit lightning talk. Only 20 spots remain. Request to join: https://luma.com/hjpck0r3 #RaySummit #AIInfrastructure #GenerativeAI

🏆 The results are in! Meet the winners of the Nebius Serverless AI Builders Challenge. Thirty-five qualifying projects showed what’s possible when builders apply serverless AI to real-world problems across AI & ML, healthcare, scientific AI, and robotics. Alongside the podium winners, seven special award categories recognized outstanding work in areas including robotics, agents, healthcare, student research, and CPU-based applications. Every project shipped with a public repository and technical write-up, so you can explore the methods, clone the code, and run the projects yourself.  Congratulations to all the winners! Learn more about the winning projects in our blog:  https://lnkd.in/gpkvTUMw #ServerlessAI #AIEngineering #MachineLearning 

When does your coding agent need its own NVIDIA H200, and when should it just call an API? 🤔 That’s one question we answer in the first issue of Build with Nebius, our new monthly newsletter for people building AI products and infrastructure. What's inside this edition? → How a $320 fine-tune led to a public benchmark for evaluating medical-text judges. → What it took to serve Kimi K3 on day zero and increase single-stream throughput by 60%. → A practical framework for choosing between a dedicated GPU and a hosted inference API. → A clonable competitive-research agent built with LangChain, Tavily, and Token Factory. Read the first issue and subscribe to Build with Nebius for more monthly developer insights. 📨 #BuildwithNebius #DeveloperNews

Your team is arguing fine-tuning versus RAG. A vendor just quoted you a per-token price. Someone says the agent needs tool access before it's useful. Technical decisions should follow business needs, not lead them. The Nebius AI Leader Certification covers the fundamentals behind those decisions: AI basics and building AI products with models accessed through providers like Nebius Token Factory. Passing means you're able to evaluate AI options and explain the trade-off with confidence and expertise. Nebius Academy graduate Aleksandr Dzhumurat put it this way: “I can now walk colleagues through the real choices in plain terms: fine-tuning vs. RAG, why an LLM needs tools to fetch live data, when structured output beats a free-form answer.” 60 minutes, 20–30 multiple-choice questions, online whenever suits you, no prerequisites. Pass and you get a verifiable certificate, a Credly badge, and free credits to build with. $1 during early bird, $100 after. https://lnkd.in/g8E_9a5Q

Nebius was built on the conviction that AI should stay competitive and diverse, rather than consolidate around a handful of models controlled by a handful of companies. In his first post since joining as SVP Research, Matthew Zeiler makes the case that the future of AI won't be decided by whether open weights exist, but by whether the system around them makes them competitive in production. The post is also a record of what that looks like in practice: Soperator, our server designs at OCP, SWE-Rebench, the Agents Blueprint, day-zero support for new open models, and our own console agent Echo running on open models. The whole AI stack moves faster in the open. https://lnkd.in/d3dv2i-r

Everyone at Nebius seems to tell a similar story about their first weeks, Mashrur Haider 🕶️ included. “You join and start contributing from day one. Your first assignment is usually something complex and genuinely impactful. You are trusted to take ownership, learn fast, and rise to the challenge. Once you do, the next opportunity follows.” That kind of leap wasn’t new for him. Before Nebius, he founded a biotech startup, where he developed the instincts he’d bring to the current role: learning quickly, reasoning from first principles, and taking calculated chances where required.  "Making decisions with suboptimal information," as he puts it. The Nebius culture matched that startup pace. With ambitious targets in an engineer-led team, people discuss challenges openly, tackle work head-on, and participate proactively to reach solutions faster. For Mashrur, that means leading products across the model lifecycle, from post-training to dedicated inference. "If you want to work on state-of-the-art technology, competing alongside the top 0.01% in the world, this is one of the few places to go." Find your role with us: https://lnkd.in/e_U8Q43N #BuildersatNebius 

Benchmarking speed alone isn't enough. Production inference needs to be fast *and* accurate. We're excited to see Artificial Analysis launch the Endpoint Accuracy Index, helping bring more transparency to how well inference providers preserve model quality. It's great to see the conversation expanding beyond latency to include correctness as a first-class metric.