Nebius × NVIDIA: Physical AI Will Be This Year's Christmas Present

A round of applause for Nebius and NVIDIA for putting on a great event in Seoul with Bloom

A global AI hackathon just kicked off in Korea. With Nebius, 150 AI builders from all kinds of industries came together around physical AI. Korea has been a manufacturing powerhouse for a long time, and the opportunity this new wave of AI brings is right around the corner.

I don't think I'll forget the moment someone told me that this Christmas, kids might be unwrapping robots under the tree. Change is happening in the blink of an eye. Let's keep building with Nebius and keep pushing what's possible.

Huge thanks to Dhruv Diddi, Guan-ru Huang, Alice Lin, and Daniel Colaianni at Nebius for their support, and to Andy Lee of NVIDIA and Yubin Park of DIDEN Robotics for their inspiring stories. Thank you as well to our wonderful Bloom ambassadors Changwon Jeon, Minjae Lee, Jueon Kim, and Jungmin Kang, and to the whole Bloom community.

We wanted to beat Tokyo, at the very least

The venue was Umul, a cultural space in Hapjeong (a neighborhood in western Seoul). It's about 7,000 square feet, with an outdoor courtyard, an indoor second and third floor, and a rooftop, and we rented the whole thing for the night. Bloom usually meets in something like a lecture room, and we'd done an indoor event with another company just the day before. But this was the launch of a hackathon, so we picked a different kind of space.

It was a Friday evening and overlapped with the commute home, but people started arriving early for the 5 p.m. check-in. Everyone ate dinner first and said hello to whoever was sitting next to them, and we started at 5:30. When I asked who was at Bloom for the first time, about a quarter of the room raised their hands. I also saw plenty of faces from the night before.

Nebius is an AI infrastructure company listed on Nasdaq, and this was a global hackathon Nebius is running together with NVIDIA. It started in Tokyo two days earlier and goes through Da Nang, Seoul, Kuala Lumpur, Singapore, and Taipei, then London, Paris, Berlin, New York, Toronto, and San Francisco, 20 cities in all. In my opening, I said exactly how I'd felt while preparing.

As the organizer, I felt a little competitive. We have to do better than Tokyo, at least. We have to do better than London, at least.

I wanted ours to be the hippest of the 20 cities, with the biggest turnout and the most buzz. We even opened a Discord channel just for this hackathon.

Today is the starting line

Nobody was submitting anything that night. The hackathon runs online, and the deadline is 10 a.m. Pacific time on October 30, which is 2 a.m. on October 31 in Korea. Anyone can enter, even if they didn't come to the Seoul event, so we treated the night as a place to form teams and get comfortable with the tools. The Luma page said up front: no demos, no pitches.

To qualify, your project has to run on Nebius Token Factory or Nebius AI Cloud and use at least one NVIDIA open-source model. On the language side that means the Nemotron 3 family, and on the physical AI side it means Cosmos, GR00T, and Sonic. There are four tracks:

- Coding and agentic engineering: coding agents and developer tools that write, run, and test code in a Token Factory sandbox. - Apps and agents: apps or agents that someone will actually use. - Personal AI: an always-on personal assistant where you keep control of your own data. - Physical AI: agents that sense and act in the real world, through robotics, IoT, or on-device systems.

The prize pool is over $50,000. First place gets $20,000, second gets $10,000, and third gets $6,000, and the four track winners each get an NVIDIA Jetson Orin Nano. There's also $3,000 for the best use of Tavily, $500 each for 20 city winners, and $100 plus NVIDIA swag for ten people who leave useful product feedback. City prizes are for people who attended an offline event, so everyone there that night enters under Seoul. A single project can win one overall prize, or one track prize plus one bonus prize.

You submit a YouTube video no longer than three minutes, a public repo with an open-source license, a written description, and feedback on the Nebius and NVIDIA tools. Participants keep ownership of what they build. For the physical AI track, the video has to include at least one minute of real hardware or a real robot moving. Simulation footage alone doesn't count, and a lot of what was said on stage that night came back to this one rule.

Judging first checks the basic requirements, then weighs four things equally: technical implementation, design, potential impact, and quality of the idea. Judging runs December 1 to 15, and winners will be announced around January 11. As of September 12, 3,589 people had registered on Devpost.

Nebius builds data centers with NVIDIA

The first session was from Guan-ru Huang, a cloud solutions architect at Nebius. Guan-ru is from Taiwan, has been at Nebius for over a year, and covers all of Asia, including Japan and Korea.

The first thing Guan-ru brought out was gifts: $100 in credits for Token Factory, Nebius's inference platform, and 1,000 free credits for Tavily, which Nebius recently acquired. If you're already using ChatGPT or Google APIs, you can switch to open-source models right away through the same style of API endpoint. Guan-ru also passed along what he called insider info: usually one team represents each city, and Seoul is included.

When he asked who had heard of Nebius, not a single hand went up. Guan-ru laughed and said that was actually great. Nebius is already a public company on Nasdaq, so when someone at an event tells him they're a small Nebius shareholder, he tells them he's doing his best to maximize shareholder value. A lot of people know the name because the stock has done well in the US, but he wanted to stress something else. As a strategic partner of NVIDIA, Nebius builds data centers to NVIDIA's reference architecture and is vertically integrated, making its own hardware, data centers, and software stack. The company slide listed deals worth about $17.4 billion with Microsoft and up to $27 billion with Meta, a $2 billion strategic investment from NVIDIA, and a goal of more than 5GW of NVIDIA systems by 2030. After the US and Europe, Nebius is now expanding into Asia Pacific.

He put up a diagram of the layers of an end-to-end AI platform and said that if you understood every term on that slide, you could work at Nebius, so come find him afterward. They're hiring in Korea. He also introduced the Builder Program, which was launching that day. If you sign up, you get $25 in Token Factory credits and $25 in Tavily credits, and you can get Nebius certified for $1. The official page also lists partner credits from companies like LangSmith and Toloka. Everyone in the room already had $100, so he asked us to tell friends who couldn't make it.

You hired ML scientists, but you have no GPUs

Nebius describes itself as an AI-native cloud with the performance of a supercomputer and the flexibility of a hyperscaler. Guan-ru explained that line by splitting the industry into two groups. Hyperscalers like AWS, GCP, and Azure built their data centers for general-purpose workloads, not AI. On the other side, GPU rental companies lend you GPUs with no software stack. Nebius sits in between: it builds data centers for AI workloads from day one and puts its own cloud software on top.

When customers go looking for GPUs, the first thing they care about is time to market. The worst case for an AI startup is not being able to get capacity.

ML scientists panic. They've been hired, but there are no GPUs to train the model on.

Models today train for weeks or months at a time, so if the cluster drops out partway through, you start over. That's why customers look at stability, and at whether developers who know Kubernetes, Terraform, CLIs, and SDKs can work the way they're used to. You might start with 10 GPUs or a few hundred, but if the business grows you'll need thousands, so they also check whether they can grow inside the same cluster and the same data center. Guan-ru came back to this point more than once: Nebius has a full cloud stack, and bare-metal GPU rental is only one part of it.

He explained the hardware so non-developers could follow. He called the three ingredients of a GPU cloud, compute, storage, and networking, the fat, carbs, and protein of an infrastructure company. On the spec sheet, H100 and H200 are the older Hopper generation, and B200 and B300 are the newer Blackwell generation. Each generation has more GPU memory: 80GB on the H100, 180GB on the B200, and 279GB on the GB300. If you train a model with a lot of parameters on older GPUs, you have to split the model up, which means more maintenance and more developer time. That's why customers show up already knowing which GPU they want.

One VM usually holds eight GPUs. GPUs inside a node are linked with NVLink, and nodes are linked to each other with InfiniBand, so if you need 16 GPUs, you connect two eight-GPU nodes. Within Blackwell, the GB300 uses NVIDIA's own Grace CPU instead of the Intel CPU from the previous generation, so the CPU and GPU are more tightly coupled.

There are layers to how you use the GPUs, too. The easiest is serverless. You don't manage VMs or Kubernetes, you just push a container: one-off training runs go through Jobs, and inference goes through Endpoints. For teams that want to handle the infrastructure themselves, there's Soperator, an open-source operator that runs Slurm on top of Kubernetes. ML scientists today have been writing Slurm scripts since their PhD labs, while DevOps engineers coming from AWS know Terraform and Kubernetes but nothing about Slurm. Both of those people work at the same company, so Nebius gives them a Kubernetes cluster for DevOps and a Slurm cluster for the scientists.

There's also support for SkyPilot and Ray, which a lot of people use now, and managed Kubernetes. Storage comes in two kinds: a shared file system that every node can read from during distributed training, and S3-style object storage for data and training checkpoints.

Retraining the model on your own traffic

Next was Token Factory, the managed inference platform. Guan-ru asked if anyone knew Artificial Analysis, a third-party site that compares the speed and price of models and inference providers. Again, nobody did. On that site, Nebius ranks in the global top three for throughput on DeepSeek R1, gpt-oss-120B, and Kimi K2.5, and gets around 409 tokens per second on Kimi.

Right now, most developers either use a closed API or host models themselves. Closed APIs from OpenAI or Anthropic are the easiest at first, since the endpoints and docs are all there. But as your service grows, API costs eat into your margins, and performance may not be what you hoped. Self-hosting on your own GPU cluster means you need a dedicated DevOps team to run the inference cluster, and you may not be able to get as many GPUs as you need. The bigger problem, Guan-ru said, is that neither option gives you a way to keep improving the model with your own service's traffic.

Token Factory builds that loop into the product. You pick an open-source model and start running inference, export the traffic data to see what customers actually ask, and then post-train again. You deploy the resulting custom weights to a dedicated endpoint on whatever GPU type and region you want. Guan-ru called it a flywheel that keeps making the model better. Latency comes down through speculative decoding, cache-aware routing, and KV caching.

There are plenty of inference platforms, but few have their own AI cloud, so elsewhere you can't choose the GPU type or the serving region. For companies with security and compliance requirements, that choice matters. There are more than 60 open-source models to choose from, with more every week, and going from PoC to production takes a few weeks. The example on the slide was Revolut: two weeks from first contact, through a PoC, to processing billions of tokens a day.

No model was trained on what happened two seconds ago

The last piece was Tavily. If you ask ChatGPT what the score is in a game right now, the model has to pull the answer from the web.

No model can be trained on that data for things that just happened two seconds ago.

But the web was built for people to look at, so whether it's a search results page or raw data from an API, it's hard for an agent to read as is. Tavily is the layer between agents and the web. It offers search, which returns a list of relevant URLs; extract, which pulls the content of a URL into something like Markdown; and crawl, which follows the links between URLs. The official site also lists research, for deep research, and map, which walks through the URLs on a site.

There were three problems to solve. The web isn't a static database, since everyone is posting something new every two seconds, so accuracy is hard. Even if you get thousands of results, how relevant they are to the question is a separate problem. And if you let an agent dig through hundreds of pages, results come back slowly and users leave. Fetching and parsing pages every time a query comes in takes several seconds per request, so Tavily searches an index it has already fetched and chunked, and returns results in under a second.

What ties all three together is freshness. If a web developer asks why the latest React upgrade isn't working and gets back docs or Stack Overflow questions from two or three years ago, that's useless. So Tavily runs a refresh scheduler. It combines four signals into one score: how often a page changes, when it was last updated, how often it's searched, and how important the site is. That score decides when to crawl the site again. Tavily isn't its own model. It's a layer that sits on top of models like ChatGPT, Mistral, and Anthropic's, so it fits anywhere you need to pull information from the web, whether that's a coding agent, CRM data enrichment, or company research.

By the end of this talk, everyone's a roboticist

The second session was Dhruv Diddi, who leads the physical AI ecosystem at Nebius. Dhruv has worked at Google, Turo, and YouTube, and owns the Physical AI Workbench, which Nebius put together from NVIDIA Cosmos, Isaac Sim, and Isaac GR00T. He opened by saying Seoul is a manufacturing hub, so he wanted to bring GPUs, robots, and this technology here. He asked the room about their backgrounds first. Most came from traditional software, some from AI-native startups, and a few were roboticists who already knew physical AI inside out.

My goal is that by the end of this talk, everyone becomes a roboticist. Well, that won't really happen, but at least you'll be comfortable with the terms in robotics and physical AI and able to hold a real conversation about them.

Physical AI means any AI you deploy into the real, physical world. It could be an embedded system or something that makes household objects smarter, but the form with the most value is robotics. The slide put digital AI and physical AI side by side. Digital AI outputs tokens on a screen, so a hallucination is just wrong text and the cost of an error is zero. Physical AI outputs motor torque that goes straight into the real world, so a hallucination becomes a physical collision and the cost of an error can be fatal. A single instruction like "don't drop the glass" eventually has to turn into joint angles and torque.

You can't make mistakes in the physical world. This isn't about dropping a glass of water. If you're not precise, the results can be truly catastrophic.

He boiled down the terms we'd keep hearing that night into three. The two ways of training robots, imitation learning and reinforcement learning. Evaluation, which you can think of as the robot version of a test suite. And foundation models. Robotics people spend most of their time on simulation training and simulation evaluation, because before you put a model on a robot and burn out the motors, you need to check whether it overheats or makes noise. One line on the slide said that 95 percent in the lab still means 50 failures a day at scale. Foundation models here also differ from LLMs: they deal with motor actions, trajectories, and sensor predictions. NVIDIA's Cosmos, GR00T, and Sonic fall into this category.

The history slide was titled "76 years trapped in abstraction." It started with Turing's imitation game and rule-based AI in the 1950s, went through the statistical era of learning from pixels from 1986 to AlexNet in 2012, and then the generative era after the transformer in 2017, when AI could reason and create in language but still lived inside a screen. Since 2023, world models like Cosmos and action models like GR00T have been giving AI degrees of freedom in the real world. Dhruv said the next 75 years will be the era of physical AI, and added that this wasn't his line. Jensen Huang said it himself.

Three computers that don't line up

The one thing Dhruv said you have to understand about physical AI is the three-body problem. The factory where you collect data and train, the matrix where you run simulations, and the edge where the model actually runs are all different computers. The factory is a DGX-class GPU cluster built from H100s or H200s. The matrix is an RTX-class server that uses Omniverse, Isaac Lab, and Cosmos to make digital twins and synthetic data. The edge is a Jetson Thor-class onboard computer that has to react within milliseconds. The conditions on these three computers don't match each other, and that's why robots aren't everywhere yet.

Sometimes data collection doesn't match the simulation. Sometimes the simulation doesn't match deployment. Sometimes nothing works, and sometimes everything works. The cool videos you see online are the times everything worked.

Every model in robotics falls into perception, reasoning, or action: a model that takes in sensory data, a model that reasons about what to do next from that information, and a model that turns that reasoning into motor trajectories that move the robot. In the slide's example, a VLM notices the glass is at the edge, a world model predicts that on this trajectory the glass will fall and break, and a VLA corrects the grip at up to 1000Hz. Dhruv said robots sound very complicated, but they aren't. You just need to understand the frame: collect data, train, and deploy the trained model.

He explained the model names with the same frame. An LLM takes text and gives you text. A VLM takes an image and text, so if you ask whether there's a cat in this photo, it answers yes or no. A VLA takes an image and text and gives you a robot action.

A VLA doesn't answer with words like yes, no, or hello. It answers with what the motors should do, given this scene and this instruction.

A $200 robot arm went around the room

The data robots learn from comes from three places. The most scalable is video. Dhruv said it would be great if robots could learn by watching YouTube, and a lot of people are working on exactly that. The catch is that phone videos carry no force or joint information. Simulation lets a GPU cluster generate millions of episodes overnight, but the physics is approximate, so it drifts from reality. Teleoperation, where a person controls the robot directly, pairs actions and states perfectly, but you only get 5 to 50 episodes an hour. The slide said we've already used up the internet's text, and the kind of data robots need, with joint angles, force, and camera frames all lined up, doesn't exist at scale. That's when Dhruv brought out an SO101 robot arm for teleoperation.

This is an open-source robot arm. You can buy it online for about $200. It's 3D printed, so you can fix it yourself, break it, and source your own motors. It's the most scalable robot you can get right now.

You move the leader arm and the white follower arm copies it. You collect thousands of episodes this way through teleoperation and train on those trajectories. Dhruv passed the arm into the audience, said he just hoped nobody would take it home, and it went from hand to hand. The easiest way to tell imitation learning and reinforcement learning apart is whether the robot interacts with an object.

All those videos online of robots dancing and doing workouts are reinforcement learning, because they're not interacting with objects. If you want a robot to handle a cup of water or move a chip, you use imitation learning.

The principle on the slide was "imitate first, reward later." Imitation learning copies human demonstrations. It's fast and sample-efficient and gets you to a starting point of about 80 percent success in a few days, but it can never do better than the human. After that, reinforcement learning optimizes for reward through trial and error in massively parallel simulation, grinding down the rare failures imitation learning can't reach. You also have to randomize textures, friction, and lighting during training so the model holds up when the real world varies.

All of this is complicated: the GPU setup, the training methods, the know-how for the robotics tools. The field is new, so the tools are scattered. That's why Nebius bundled the workflows into one open-source repo. It includes a training workflow that takes as few as five episodes, augments the rest, and produces a model; a data augmentation workflow; and a workflow for evaluating robot models, which normally takes a lot of time and know-how. A CLI called npa connects your Nebius project and storage and calls LeRobot, Isaac Lab, Cosmos, GR00T, and FiftyOne, and you can deploy any model with three commands.

With the Workbench CLI, you can literally just tell an agent to do it. The agent sets up the GPU cluster and runs the training. We designed it so agents can use it easily, so people who aren't comfortable with CLIs or have never done robotics can jump in.

He had prepared a demo of a model trained to pick up a Nebius marker, but we ran short on time, so it moved to after the talk. The example left on the slide was GR00T, pretrained on more than 20,000 hours of human video and robot demonstrations, fine-tuned for the SO101 for about 2,000 steps, then carrying out the instruction "pick up the red block." A deployment slide he didn't get to on stage had this point too: if inference is even 50 to 100 milliseconds late, a bipedal robot can fall over, so intelligence has to be physically split between the cloud and the edge. A large VLA plans on a slow cycle, and a small policy inside the robot handles balance and contact at 1000Hz.

The model maker, the field deployer, and the infrastructure provider on one stage

At 6:15 I moderated a panel. With me were Andy Lee, who leads physical AI in Asia Pacific at NVIDIA along with Inception, NVIDIA's program for startups and VCs; Dhruv, who had just presented; and Yubin Park, CBO of DIDEN Robotics and also a Bloom member. DIDEN Robotics builds quadruped robots with magnetic feet that climb walls and ceilings of steel structures to weld and inspect. They passed a field trial with Samsung Heavy Industries and have started delivering units, and they're co-developing a welding robot for the inside of ship blocks with HD Korea Shipbuilding & Offshore Engineering. So we had a company that makes the models, a company that supplies the infrastructure to run them, and a company that actually puts robots into shipyards, all in one row.

The first question was what physical AI they'd gotten their hands on lately. Dhruv said he'd just given the same talk in Tokyo and Singapore was next, and that he owns about 35 open-source robots, not counting the big ones. These days he's playing with OpenAI's Astra model, and there's a lot of talk online about it showing emergent properties that make robot tasks easier. Andy used the 3D printer at home to have Astra make life hacks that save him time, and recently bought a MicroDuck and has been tinkering with it. Yubin's answer was short. DIDEN builds its own robots, so the last robot Yubin touched was the spider robot they'd brought along with a colleague for that day's presentation.

When I asked Andy to explain Cosmos and GR00T in plain language, he said this.

Cosmos is a world foundation model. It tries to understand the world we live in. It perceives, predicts, and generates video data so the model can learn about the world around it. GR00T is a robot foundation model. It turns what it learned about the world into real action, in other words a policy. So when it sees a ball flying at it, it can lift a leg and block it.

A perfect score in simulation, and the robot still slips in the shipyard

When I asked what's hardest about teaching a robot something, Dhruv laughed that Andy had taken the easy question. His answer was dealing with the physical world. Training in simulation has come a long way, and GR00T and Cosmos are basically state of the art. But as earlier physical AI startups like self-driving companies learned, you get a perfect score in simulation, decide that's enough, and only after deploying in the real world realize you have to break the problem back down into autonomy levels like L1 and L2. For some things, you don't even understand why they're complicated until much later. Dhruv named deployment as the hardest stage, and said that for deployment to get smooth, there just has to be more of it.

I asked Yubin how their robots actually work. Yubin laughed that it was a very long question, then split the problem in two. One part is getting around. A ship hull is a three-dimensional ferromagnetic surface, so the robot has to move vertically, horizontally, and on the ceiling, and to get from point A to point B they use model predictive control or reinforcement learning based control. The other part is tool use. The industrial arm on the robot swaps between welding, angle grinding, shot blasting, and painting tools, and each tool is controlled differently, so they're working out how to use VLA models to control how the robot welds and paints steel plates.

Then, building on what the Nebius team had said, I asked what breaks first when you put a robot into the real world after plenty of simulation.

When we first built our reinforcement learning simulator, we made a very ideal environment, with flat floors and clean angles. But a real shipyard has disturbances. When you weld, small metal fragments scatter on the floor, and if the robot steps on them, it can slip and lose control.

That's just a very simple example, and there are countless variables that can go wrong. Whether it's the NVIDIA simulator or MuJoCo, you can train the algorithm and test everything, and it still always breaks in the real world. So to iterate on the system quickly, you first need a real environment and a real robot, and you keep running it there.

Without the real world, you cannot build any kind of solutions to it.

I asked Andy which is harder in physical AI, the AI or the machine. His answer was 50/50, for sure. AI models are getting exponentially better, like GPT-6 Astra and the just-released Fable, but the moment you deploy in the real world, variables you couldn't account for pour in.

The lighting conditions are different, the environment is different, something can go wrong. Some of the lights might not be on. So hardware has to catch up with AI, and that's exactly the stage we're at right now.

Physical AI forces you to think more efficiently

When I asked whether more compute makes a team better, Dhruv pointed the other way. Of course more data, post-training, and good GPUs help, and he got a laugh by saying that's thanks to the NVIDIA engine. But the physical world forces you to specialize. If you train on garbage data or data that's spread too wide, the weights never settle properly, so right now things work best when you narrow down to one kind of lighting, one kind of robot, and one environment. GR00T models also work well when you follow the instructions in the paper exactly.

Physical AI forces you to think more efficiently. It's not the paradigm where you throw a ton of internet data into a big model and make it work.

In fact, physical AI models are on the small side, around 2 billion or 7 billion parameters, because they're multimodal and can be trained more densely. A text model only gets its resolution by stacking text embeddings, while a multimodal model has more space to fill, so he expects a lot of clever new ways of using GPUs. This was where it hit me again that software and physical AI start from different places.

Korea's edge is being close to the factory

Here I asked a question that wasn't on my list. Software data looks about the same anywhere, but physical AI is hard because you need detailed data from the field, and Korea is known as a hub for manufacturing and factories. Did he have advice for builders here? Dhruv said there's clearly a big advantage in being close to manufacturing sites and being able to get into those spaces. Korean startups are already using that advantage, as with the shipbuilding example, and he also mentioned RLWRLD, a Korean company building foundation models for hand manipulation.

The biggest advantage comes from being able to control the manufacturing environment directly. Because of that, Korea gets ahead on data collection and will probably be able to deploy models earlier than other countries.

I asked Andy why NVIDIA cares about Korea. Until this event I'd had no connection to NVIDIA, so I didn't even know how big NVIDIA's Korea team was. Andy said Korea is a very strategic place for NVIDIA. Manufacturing is strong, including semiconductors and cars, and the physical AI ecosystem is growing fast. By the International Federation of Robotics' count, Korea has 1,220 industrial robots per 10,000 manufacturing workers, first in the world and more than nine times the global average of 132. At the same time, the population is shrinking and it's getting harder to find people to work on the floor, so the demand from manufacturing is clear. Out of that, startups like RLWRLD and Motif Technologies are emerging. There's also OpenGraph Labs, which collects tactile data from fingertips, and SpaceAI, which builds robot training data by attaching physical properties like touch and force to objects that only appear in video. SpaceAI was the only Korean company selected for the APAC AI accelerator that NVIDIA runs with a Taiwanese accelerator. Andy said there are a lot of hidden gems in this region.

When Andy asked for a show of hands from anyone who'd heard of NVIDIA Inception, not many went up. Inception is a program for startups and VCs, and any company under ten years old can join for free. It comes with chances to learn and build with NVIDIA products like Nemotron, Cosmos, OpenUSD, and Omniverse, Nebius cloud credits, and introductions to US and Korean VCs for funding. NVIDIA's Korea team is here and can connect you all the way to the US and Europe, which is a real advantage for Korean startups.

The Nebius site also has a call for the Nebius Physical AI Awards for physical AI startups. It's open to teams that have raised $1 million or more and have a product actually running, with applications due October 21. Each category winner gets $150,000 in Nebius compute credits and presents at the December summit in San Francisco.

Do hardware

I asked Yubin what had been hard about starting the company, and what they'd do better if they started over. There was no hesitation.

In physical AI, you have to do hardware. If you use someone else's hardware, there's a clear limit to how far you can optimize the system as a product.

Yubin said that if you're a really good software developer, find a hardware co-founder. If you're only optimizing the software layer, it's too easy for a customer to pull your software out and swap in someone else's. If you take Chinese hardware off the shelf, you lose any edge over a team that's better at software, and the better language models get, the easier software is to replace.

How do you stay attached to the customer? It's simple. Do both hardware and software, vertically integrate everything, and sell it as a package.

The last question was the one piece of advice they'd give someone starting to build today. Dhruv said to start with the resources that are already out there, including the builder program. With the right tools you can make magic, and you don't have to lean on AI for everything; you can mix MCP in with it. His yardstick was ROI. Focus on what's impressive and high-impact for your team, and on what your team brings to this competition.

I would recommend to start small. Rome wasn't built in a day. What you build tonight might not be the best product, and you might not get the result you expected. Don't aim for the best model in one day. Set small goals with small steps, measure whether you actually hit them, and move forward from there.

The mic went back to Yubin to close.

I've never won a hackathon, so I can't give great advice. But like I said earlier, software is pretty easy to build these days. So why not try to stand out with hardware? That's what I think.

The most valuable data comes from deployment

After the panel, the hackathon began. For people who came late, someone from Nebius went over how to get credits and what to build again on Discord. There were two kinds of credits, cloud credits and Token Factory credits, $100 each. You claimed them through a QR code posted on Discord, but only people who had registered on Luma could get the compute credits. They said they wanted people to use compute responsibly.

We asked everyone to form teams first. Teams that had formed and decided what to build could come up at 8 to introduce themselves, and the first ones would get Nebius T-shirts. The condition was that you couldn't just say anything.

Meanwhile, Dhruv kept talking about data with people who came up with questions. Besides the teleoperation he'd shown, there's a method where a person holds a gripper-shaped device in their hand and does the task. That's easier than mapping human hand motion to a gripper and then to a robot, so gripper data is cheaper to collect. Collecting robot-to-robot directly is worth more, and the most valuable of all is data collected while the model is deployed. If you have a model automating a task, even clumsily, you have workers watch it and correct it on the spot, and the data that builds up is called DAgger data. Dhruv said that's the most effective data with the highest ROI.

After he came off stage, Dhruv told me one more thing. This Christmas, kids might be unwrapping robots under the tree. He said San Francisco already feels that way. When you think about the robot arm that went around the room that night, a $200 open-source arm you can 3D print yourself, it doesn't sound that far off. It struck me that hardware is turning into something anyone can print out.

Another gap Dhruv pointed to was embodiment. What you train on one robot doesn't carry over to another. There's no way yet to make cross-embodiment work across different robots, and Dhruv saw exactly that gap as the opportunity to innovate. Unlike AI, the physical side is still open ground with so much left to do, so we need thousands more roboticists, he said.

We said no demos, no pitches, and over ten teams took the mic

At 8, people who had formed teams and settled on ideas started coming up one by one. The Luma page said no demos and no pitches, but once we said it was fine to come up solo, more than ten teams took the mic. Dhruv gave most of the feedback on the spot. I'm only describing the presenters by what they do.

The first presenter said they wanted to come up and get feedback directly, and that the idea wasn't final, so it was fine to tear it apart. They didn't think building one more agent was any kind of differentiator, and wanted to solve the social problems that come up when the economic actors are AI agents instead of people. Say a hotel booking agent checks prices at all one million hotels in the world, books every one of them, deliberately waits 23 hours, and then cancels them all. The hotels have to respond to useless requests and hold rooms, and with just tens of thousands of agents like that, booking systems and the economy grind to a halt.

A buyer agent costs nothing. Anyone can build one. But a seller agent has to quote prices, negotiate, and actually take the booking.

What the presenter wanted to build was a commitment gateway that helps the seller side so requests aren't free. It tells the seller whether this buyer is the type that checks a million listings to book one, or the type that looks at five or twenty and books one. Without hard-capping the number of calls, it would sort real buyers from real sellers and help the agent-to-agent economy run well. Dhruv said it was a very valid problem, and that a lot of ideas are coming up these days about how agents deal with each other.

Features that are easy for AI end up being easy for people

A team building a Kubernetes-based GPU management platform started in English and switched to Korean. They said we might laugh if they said they do something similar to Nebius. They don't provide infrastructure directly, but they build software that sets it up and helps people make good use of their GPU servers. GPU infrastructure has always been hard to handle unless you're an engineer, so they'd focused on a UI that's easy for people, but using AI changed their thinking. In the past, a person had to research and also know how to make slides, and now AI makes the first draft. They figured infrastructure setup could work the same way. For example, when a professor sets up a lab environment for 30 students, the infrastructure work usually goes to a teaching assistant, and they want AI to do it instead.

I think the features that are easy for AI to use will end up being the features that are easy for people to use.

Dhruv said GPU orchestration isn't something to see as competition. The field is still early, especially at the level of interactive UI, and there's a lot to optimize, so think of it as growing the pie and definitely go for it.

Someone who runs a YouTube channel said none of the editing tools did what they wanted, so they built their own. After using it on their own channel, they heard from a lot of people who wanted it too, and realized many people want to put out content but can't publish videos because editing is hard. The goal is to help those people so more good content gets out into the world.

Someone who wanted to build personal AI used Jarvis from Iron Man as the example. The most basic piece is reminders. It came from something that had just happened to them. They had to complete a resident registration check (a government check on where you live), missed the deadline, and now someone was about to come to their home. If an AI had looked at their address and reminded them ahead of time, it would already be done. So they sketched a system that finds the dates you need to keep track of, summarizes them, and tells you which website to go to. Government offices and others put information out, and the AI gathers it and notifies the people who need it. Events posted on Luma that match your interests would come in as notifications so you can sign up right away. It's a platform meant to help both the side sending the information and the side receiving it.

Not a duckling, but the egg the duck came from

Someone who introduced themselves as a maker came up with samples they'd built. They have 12 3D printers in their office, and ten years ago they built ten 3D printers themselves through the RepRap project. They worked as an English teacher, then became a civil servant and ran a makerspace for three years, got bored, quit, and started a company. The office has thousands of parts, and when they think of something they want to make, they can print a prototype in two hours. They were also candid about why they make a lot of educational products.

Between a product and a prototype, education is the only thing that makes money.

The proposal was an egg. They'd been heavily inspired by MicroDuck, the open-source duck robot from Hugging Face and Pollen Robotics, and wanted to make the egg the duck came from instead of its duckling. Using reaction torque, the way satellites control their orientation, you can make an egg-shaped robot roll around on its own. With a light inside and a face on it, it could be cute, they said. They pictured a robot that roams around the house, small and cute enough that you'd want one, for under $200. They didn't know yet what the egg would do, so they asked for ideas, and since they didn't know much about GPUs, they were looking for a partner to handle the tech.

Dhruv said he loves open-source ideas built with 3D printing, MicroDuck especially, and that an egg is a good form factor. What we always want from robots is interaction, so he suggested adding actuation to an egg like a small Reachy Mini you can talk to. Voice is good, but it would be even better if the egg could tap things to make sounds or control other objects.

I think the real value of a robot being in the physical world is that it can move something.

It started with putting arms on it and got as far as how fun it would be to have an egg with just legs walking around. The egg that didn't know what it would do started growing arms and legs right there.

Pressing a button is a hard task for a robot

One team wanted to use button pressing to work on the differences between robots. The challenges they listed were cross lighting, so it works when the lighting changes; cross environment, so it works when the environment changes; and how to match the physics of one motor to another. Dhruv's first piece of advice was not to invent a new task. Models are already optimized for the tasks in existing benchmarks, so look at what models do best and build a version fine-tuned for a different robot.

Then the conversation turned to how hard pressing a button actually is. There's timing for when you press and when you release. Press too hard and the button breaks. Let go too early and nothing happens.

Most of these models judge by sight, not touch. They don't know whether they're touching the button. They only know whether the color of the robot arm is close to the button or far from it.

That's why in robotics, pressing a button is a task that takes a lot of dexterity. He suggested starting with easier things like picking up, and with soft objects that are less likely to break, like T-shirts, squishy toys, or socks. The presenter asked back: what if they put a picture like a QR code on the button so the robot knows up, down, and orientation? Dhruv said now the computer vision tricks were coming out. A QR code would work, or you could mark where the robot should go with a box. Be a little clever and make it work reliably, he said.

A researcher at a company that helps write radiology reports asked about training costs. Training at scale takes a lot of resources, data is scarce to begin with, and it needs to be accurate across many dimensions, which it often isn't. Starting from the question of whether you have to pay a lot to train every time, they wanted to research training small language models or multimodal models quickly and delivering them, customized, to relatively small groups of users. Their question was what to focus on, since more and more of the research process can be automated.

Dhruv said the opportunity is in data augmentation. Physical AI in particular is short on data, so if you have 100 episodes, you want to turn them into 1,000. For language models, you write synonymous sentences that teach the same concept with slightly different words. For vision, you zoom an image in 10 percent, zoom it out 10 percent, crop it, or tilt it 5 to 10 degrees to multiply the data tenfold, then fine-tune. People already do this in practice, and he said to focus on the highest-ROI work in research too. He pointed to Oumi as an open-source library for this kind of work.

I don't have feedback for you, I'm a fan

A neurosurgeon is building an autopilot for spine surgery full time. Their goal is to conquer every surgery in the world with robotics, and they said that just as data was the bottleneck for digital AI, physical AI models need to collect as much data as possible. The problem is that data is hard to get in an operating room. So they're working with a company that makes spine dummy models to collect the positioning doctors already do in surgery, and they've already built a tip fitted with a tracking system. They were looking for someone in computer vision to join them.

I don't have feedback for you. I'm a fan, actually.

Dhruv was impressed that they'd picked a problem that is extremely hard both technically and in terms of impact, and had already made progress with a device they built with their own hands. He said they'd already talked through how to improve it with more sensors.

A team building an AI autobiography service for older adults is running it as a program at several senior welfare centers. When they meet older adults, introduce AI, and teach them how to use their phones, it's harder for them to follow than you'd expect, and even when apps like ChatGPT are installed, almost nobody has actually used them. So they plan to train a model specialized in Korean, make it into a small local SLM, and put it in a small device that older adults can talk to in daily life. Older adults won't be coding or solving math problems, so if you focus on Korean, a cheap, small model is enough. They named the product AI, so you call out "AI" the way you'd call out "agiya" (baby) in Korean.

I think the big opportunity in AI is access to intelligence. Older adults are exactly the ones who need intelligence they can use without buttons or typing.

Dhruv said he loves projects with big social impact. There's a big market for Korean-specialized models, but not enough Korean data, so it would be good to open-source the data they collect. If they deploy on-device, it can help older adults 24 hours a day, offline, without depending on the internet, and post-training can happen in the cloud.

You need a foreign language to make friends, and friends to learn the language

As the presentations were winding down, someone who said they'd missed their chance earlier took the mic and said we didn't have to listen. They're building a voice messenger without language barriers. They want to get close to people from other countries, but the language barrier is so hard to get past, and there's a chicken-and-egg problem.

I wanted to make foreign friends, but you have to learn the language to make friends. And I wanted to learn the language, but you need foreign friends to learn it.

There are plenty of language exchange services, but none really solved this, and other services that connect you with foreigners end up hitting the language barrier too, so people give up. The proposal was to remove the language problem itself so you can just focus on hanging out with friends from other countries. When they looked into why existing services couldn't do real-time interpretation, they found that earlier AI couldn't meet the tricky demands of making friends, and there were UI and UX problems too. Starting this year, AI has gotten good enough to solve all of that, and they realized this business has only just become possible.

When they spoke Korean into the interpretation app they'd built, real-time English and Japanese translations streamed out. They had people use it alongside HelloTalk voice rooms, and the response was good. A large share of users went into the voice rooms with the translator on. Plenty of people use it offline or for work, and someone even got a job with it. There are about seven situations where people need interpretation, and they plan to go after them one at a time, starting with users of K-Friends (a Korean app for meeting foreign friends). No service covers both real-time interpretation and global socializing, so the plan is to start with interpretation and grow into an AI-native messenger and social network. They were looking for teammates too.

Using AI to do what can only happen offline

One piece of advice I got from the Nebius team after the event has stuck with me. If you're an offline community, hold on to what can only happen offline. Physical AI fits that exactly. Anything that ends on a screen can happen anywhere, but building a robot, touching it, and breaking it takes being in the same room.

Inside our team, we also tried applying what Yubin said to Bloom. Software is easy to scale, but now it's also far too easy to replace. Hardware is hard to scale, but also hard to replace. For Bloom, the hardware is people. We're in an era where anyone can use and control AI, but people still don't do what you want them to, and that's why a room full of people is worth something. People come to see people instead of a screen, which is why we hear that coming to Bloom feels like a breath of fresh air.

The barrier to entry has also dropped a lot. The SO101 that went around the room is a $200 open-source arm, and MicroDuck, from Pollen Robotics (which Hugging Face acquired), is taking preorders at $399. It's a bipedal robot, 25 cm tall, 800 grams, with 15 motors, and its SDK, simulation, and reinforcement learning stack are all open under Apache 2.0, so you can retrain all seven of its built-in moves. Developers loved it, preorders backed up, and I heard that if you order now, you'll wait until the end of the year. It's not just robots, either. A $359 Linux board built to keep agents running all the time is also taking preorders.

That's why, when Dhruv talked about Christmas, it didn't sound like a metaphor for the distant future. So our team decided to try two things. One is small offline sessions where people assemble and train open-source robots like the SO101 or MicroDuck themselves. The other is a regular physical AI meetup connected to Nebius and NVIDIA. We also talked about Bloom maybe being the one to bring robots like these into Korea. The point is to use AI to actually do the things that can only happen offline.

These two probably won't be the end of it. As the team building AI for older adults showed that night, if AI becomes something every generation needs, then Bloom has to reach beyond people who work in AI. That came up inside our team. I'm honestly not sure yet if that's right. But that night made one thing clear to me: some things only happen away from the screen.

I don't know how many of the teams that took the mic that night will actually submit by the 2 a.m. deadline on October 31. If today's robots can score perfectly in simulation and still slip on a single metal fragment, seven weeks might be a short time. Still, I'm pretty curious what the maker with the homemade samples and the doctor with the spine dummy model will bring back by then.

Originally written in Korean on Brunch, September 12, 2026.