Physical AI is having its moment. Robotics, autonomous vehicles, interactive video. Underneath all of it sits the same primitive: a world model that predicts how an environment responds to action. There's a lot of debate about world models. JEPA vs. diffusion vs. autoregressive. How sim-to-real actually transfers. Nobody agrees on how to evaluate physical understanding. So @taiuti & @_bschmidtchen at @reactorworld , @ashray_malhotra at @Adobe, and I decided to put the top minds in the field in one room. On June 18 in SF, we're hosting World Models SF Summit, likely the first summit dedicated entirely to world models. Researchers, founders, and practitioners from @GoogleDeepMind @OpenAI @Meta @nvidia @theworldlabs @Tesla @Apple @Microsoft @Roblox and many strong startups in the space have already signed up. A few seats left: luma.com/hkik7xtr

Qwen-trained probe pricing difficulty for @MistralAI & Phi means difficulty lives in the latent representations... not the solver. Excited to see @VmaxAI ship 🔥 Congrats @MavorParker💪

Mars rover ride Every frame generated in real time on @reactorworld's longlive-v2 well done @taiuti @_bschmidtchen 👏 just getting started.

The next era of animation won't be generated. It'll be explored. Yoki (予期) is my first film made with world models video models. The landscapes of Japan, rebuilt as worlds you move a camera through with @reactorworld. World models are the most underrated storytelling tech right now, and almost no one's using them for film. Animation made with @invideoOfficial.

Jensen at @Computex: "Text data plus compute gives you AI. Now that we have AI, compute is data (in physical AI)" LLMs trained on a corpus that already existed... whole internet of text. Robots have none. Nobody recorded the world in first-person, action-labeled frames. (Hence so many startups building here.) So you generate it. World models like @nvidia 's Cosmos spend compute to produce physics-accurate, action-conditioned rollouts from any perspective, then close the loop as the policy's own simulator. So... data becomes a compute problem in physical AI. All those world models need to compute somewhere. That's @reactorworld ... and the biggest infrastructure opportunity in physical AI today 🔥 cc @taiuti @_bschmidtchen

The unix terminal is the natural interface for agents to get work done on a computer but how well can agents actually use unix? Claude Code. Codex. Devin. Every frontier agent ships as a terminal tool. With unix-ctf, Vmax is using setters and solvers to measure Unix competence.

Amazing event last night. @SarvamAI & @LightspeedIndia (@MohapatraHemant @dkhare ) have pulled together an impressive bench of researchers from @OpenAI @GoogleDeepMind @Meta , @stripe @MistralAI and others. @pratykumar walked us through building sovereign AI out of India, and expanding into the Bay Area to tap the talent that's actually trained large models at scale. World models came up again and again in my conversations with researchers from the frontier labs. That's going to be the biggest space in AI 🚀 cc @reactorworld @taiuti @_bschmidtchen

Beautiful evening in SF @ a wonderful housefull event w/ @pratykumar of @SarvamAI talking about the company's plans, hiring for an SF office, scaling out compute, building v. large models, and global GTM. Inspiring and genuinely one of the clearest visions in the world of AI.

AI is moving from language models to world models. They need their own runtime. That's @reactorworld 🚀 LLM inference became a multi-$B category because batching, routing & GPU orchestration were too complex for app developers to own. World models will follow the same pattern... but technical bar is way higher. The workloads are stateful, temporal, heterogeneous. Video tokens. Latent states. Action loops. Streaming rollouts. Latency and compute profiles vary across models. Different category entirely. @taiuti @_bschmidtchen are building this from first principles. Both led technology on @Apple Vision Pro. @taiuti previously co-founded @LumaLabsAI as CTO.  Privilege to be on this journey with them from the start 🙏 Congrats to @reactorworld team on the $59M Series A led by stellar investors @lightspeedvp (@buckymoore @theamberyang) @WndrCoLLC @Amplify @Sky9Capital @FPVventures @AbstractVC ... and @awscloud on board as preferred cloud partner 💪  variety.com/2026/digital/new… Biggest infrastructure opportunity in physical AI today. Let's go! 🔥

Congrats to the @reactorworld (@taiuti @_bschmidtchen) & @emergentlabs (@mukundjha) teams on being named to @Redpoint annual list of the 100 most important infrastructure companies powering AI. Well earned 🔥    Proud to be an early investor in both!

Every frontier lab is hitting the same wall: human-generated training data doesn't scale ✋ Most teams respond with bigger RLHF pipelines and more human annotators. @VmaxAI is taking a fundamentally different approach. Their new PopuLoRA paper introduces co-evolving populations of LoRA adapters performing asymmetric self-play on a frozen base model. They're doing evolutionary operations directly in weight space Teams working on self-improvement without human data dependency will matter a lot. Talk to @MavorParker to learn more. They are hiring too!

GPUs melting 🫠

Real-time World Models are the next AI frontier. Today, we @reactorworld are taking the first step towards this reality: our early preview lets you experience worlds generated in real-time, running on our global low-latency infrastructure. Try it now: reactor.inc/

Real-time World Models are the next AI frontier. Today, we're taking the first step towards this reality: our early preview lets you experience worlds generated in real-time, running on our global low-latency infrastructure. Try it now: reactor.inc/

The missing layer in the physical AI stack is starting to emerge: reactor.inc/ LLMs became useful to developers only after the serving layer matured. Inference providers abstracted away GPU orchestration, batching, routing, scaling, and model-specific deployment. This led to several multi-$B startups 💰 World models need their own version of that layer. But this is not just “LLM inference with video” 🚫 World models are stateful, temporal, and heterogeneous. They involve video tokens, latent states, action loops, rollouts, streaming generation, multimodal inputs, and very different latency/compute profiles across models. The next generation of physical AI experiences will need a runtime that can orchestrate these models, preserve state, manage latency, hide backend complexity, and expose a clean developer surface. The foundation models for physical AI are being built. Now the experience layer needs its execution engine. @reactorworld is starting to reveal that layer today... Check out the demo here 👀 @taiuti @_bschmidtchen

World models ​might need more compute than LLMs.. and LLMs already triggered one of the largest infrastructure buildout in history 🏭 LLMs learn the structure of language. World models learn the structure of causality: how objects collide, how fluids flow, how crowds behave, how a scene changes after an action. This changes the compute math: Text is sparse. Video is dense, temporal, redundant, and expensive to curate. A book can fit in hundreds of kilobytes. A minute of useful video can be hundreds of megabytes before you even ask whether it contains the right objects, actions, camera angles, contacts, failures, and edge cases. The trick is compression. World models do not predict every pixel in a 4K frame. They compress video into latent representations and learn to predict how those latent states evolve under perturbations: actions, camera motion, text prompts, or environmental change. Even with this compression, the compute requirements are staggering. NVIDIA’s Cosmos world-model trained on 20​M hrs of video ​w​ith 10,000 H100 GPUs for roughly three months​ (10,000×90×24≈21.6​M GPU-hours​). At GPU-hour rates, that is ~ $100M of training compute. Owning comparable capacity outright would be a $400M capex decision 💸 But compute is not the only constraint... Data may be the harder one. The internet gave LLMs ​m​assive corpus of text. Robotics has no equivalent internet-scale corpus of action-conditioned experience. ​Passive video is not enough.​ That data has to be generated through teleoperation, real robots, simulation, synthetic worlds, or some combination of all four. And this is just training. Inference may be an even bigger bottleneck. A robot does not just answer a prompt. It has to plan, simulate possible futures, handle uncertainty, and act safely in real time. World models shift part of that burden into learned representations. Instead of hand-coding every corner of physics, ​world models amortize some of that complexity into a neural network. The ​"stochastic messiness of reality​"* gets baked into the weights. That is the real shift​... LLMs taught machines to read the internet. World models will teach machines to operate in reality. The compute buildout for that may be much larger than people expect. * I read this somewhere and liked how it captured the challenges of the real world.

Exciting! This is the rocketship to join 🚀