60 conversations in SF later, here is everything I learned, and the hot takes most of the people I talked to would argue with.
Data, models, deployment, hardware, money, and where the bottleneck goes next. https://x.com/i/article/2120617534596820993
What 60 conversations taught me about robotics
Where this comes from
Four weeks ago I shut down my first company and knew nothing about robotics. Since then I have had more than 60 one-on-one conversations in San Francisco with early-stage robotics founders, researchers, mechanical engineers, people selling robot data, people buying it, and a few investors. At the peak it was more than 20 calls and meetings a week. Many of them were in person: I went to founders' offices to see what deployment actually looks like, spent time at Founders Inc. with the team and the founders building there, went to the Humanoid Hub conference and the events at Mission Robotics, joined hardware and robotics hackathons, and had dinners with robotics founders. I also bought a robot arm.
This is what I heard, what I believe now, and where I disagree with the people I talked to. I am not an engineer. I spent four years at Slush, most recently as CPO, and before that in finance and sales. Read everything here with that in mind. I changed my mind several times in these four weeks and expect to keep doing so.
A note on sourcing. Most of what is here comes from private conversations, so I do not name people.
Data
The hunger for data is at its peak. What people are hungry for is changing.
Raw egocentric footage is already saturated. First-person video of people doing everyday tasks, collected cheaply in China, India and Southeast Asia, still sells. It barely moves the models anymore. Almost everyone I talked to who sells or buys data said the same thing in different words: generic household and factory footage is low margin, the labs already have a lot of it, and more of it adds little. One investor-turned-advisor called it commodity tooling with a shelf life of months. One founder building a supply chain business avoids the whole category because he thinks it is done.
The middleman space is crowded. There are brokers, collectors, platforms and marketplaces, and most of them sell the same hours. I mapped the price chain for one Bangladesh garment-factory data pitch: a couple of dollars an hour to capture, ten at the first vendor, twenty once cleaned, sixty by the time a big lab pays for it. Everyone in that chain is trying to cut out everyone else. Even the people in that chain give the business a few months, maybe a year.
Demand is moving up the quality ladder. What labs are starting to ask for is data that is labeled, mapped, collected on known equipment and on specific robot hardware. Teleoperation data is the gold standard and it is clunky and expensive. Failures from robots already deployed in the field are the next most valuable thing, because they are the edge cases nobody can collect on purpose. Evals are often worth more than training data. One founder who sells to labs told me the right way to judge a dataset is how much it improves a downstream task, and that data a model has already absorbed is worth close to nothing. My bet is that this shift in what labs buy completes within six to twelve months. The data stage as a whole lasts longer. More on that below, under where the bottleneck goes next.
Labeling is the quiet bottleneck. Ask ten operators to pick up a spoon and you get ten ways of doing it. Without labels that say what happened in between, more demonstrations add noise, not skill. Two people told me independently that labeling, not collection, is where the real work is.
Touch and force is the next frontier. Video cannot tell a model how hard to squeeze. A model can know a glass is fragile and still have no idea how much force this particular hand should apply. I talked with people building hands, gloves and force-sensing rigs. The buyers of that data today are the labs, not anyone deploying robots in the field. That tells you how early it is.
Everyone is buying blind. Even the labs building VLA models (the models that turn vision and language into robot actions) do not know which data improves their models fastest. That is not a criticism. There is no internet for robot actions. Text models could train on everything ever written. For arms and joints there is no vast, cheap record of what happened when a joint moved. So the labs are guessing, and the honest data sellers know it. The smart ones have already moved from selling hours to selling tooling, labeling, quality control and evaluation.
What this means if you are thinking of a data startup. Selling hours of generic footage is a race to zero. The defensible places are narrow: data from jobs that are hard to access and expensive to staff, data tied to a specific embodiment, evaluation and quality tooling, or owning a vertical and collecting your own. A founder in Brazil doing agriculture data said VCs told him egocentric collection has near-zero entry barrier and they would only fund him if he owned the deployment. I think they are right.
Models
VLAs are winning today. I think general models win the decade.
Where things stand. The best real results right now come from VLA models that a team has fine-tuned and trained with reinforcement learning on its own data. Whether the base model is open-source or closed matters less than the work done on top of it. One founder running pick-and-pack facilities does exactly this and does not touch frontier research at all. It works. But the gains are incremental, each step up in task complexity seems to need around ten times more data, and a robot left alone on a single task still lasts seconds, not hours. A researcher who deploys robots for a living told me he uses deterministic classical planning in the field and ignores the foundation model hype, because learned policies cannot yet guarantee the precision his customers need.
Astra changed the conversation. In September a general model, OpenAI's Astra, controlled two arms and put a block in a bowl 19 times out of 20. Since then research groups have run it across dozens of simulated and real robot tasks, and on many of them it matches or beats models built only for robotics. OpenAI has not said what it was trained on. Some people assume it saw no robot data at all. I do not buy that. One study found that Astra answers questions about a popular lab arm's geometry as if it had memorized that arm, which suggests robot data was in the mix. The interesting part is not what it was trained on but how it works. Astra does not output motor commands the way a VLA does. It looks at the camera, reasons about what to do, writes small programs, sends a target to the robot's own controller, checks the result and corrects. It works like a programmer, not like a reflex. The demo of the year did not come from a robotics company. That is the point.
This is a different approach to edge cases, not a faster version of the old one. The old way is to collect every edge case from the real world, one demonstration at a time. It is slow, expensive and never finished. The new way is a model that already understands physics, plus simulation to run through endless edge cases cheaply. Most practitioners I talked to do not buy this. They see simulation as a way to test a model, not to teach it. It cannot reproduce flickering lights or a wet floor, and closing the gap between simulation and the real world still needs a small set of real data. They may be right for the next two years. I do not think they are right for the next five. My bet is that simulation does not finish the last mile of data collection. It replaces most of the collection.
Astra is nowhere near the nines. Each attempt costs real money and takes long seconds to think. You cannot run a warehouse on that. So VLAs win the near term, and they are moving in the right direction. But VLAs are a few steps ahead on a slow road, and general models are on a fast one. I expect general models to catch up within a couple of years and would not be surprised if it is sooner. Several researchers already describe the architecture they want as a general model in charge, calling on a library of smaller skills, instead of one heavy VLA. A product manager at a robotics company told me some teams already consider VLAs outdated and are trying language models with traditional controllers underneath.
The reflex layer stays separate. When the brain is a general model, what stays specialized is the fast controller: the thing that tightens the grip the instant the glass starts slipping. A reasoning model cannot think fast enough for that. A flight controller runs at a thousand hertz, a VLA at one. That layer stays small, fast and specific to the hardware, and it may end up being plain control software behind a standard interface rather than a trained model at all.
What this means for robotics model companies. General models have taken over verticals one by one. Robotics is one of the next ones. Standalone robotics models will have to prove what they can do that a general model cannot, and that list is getting shorter. The counterargument is coding: frontier models are great programmers and specialized coding products still have big markets. Maybe robotics is big enough for both. But if I were raising for a robotics foundation model today, I would want a very clear answer to why a frontier lab cannot do this in six months.
Deployment
Deployment is the part nobody warned me about. It is where the money is made and lost, and it is almost entirely manual.
It is an operations business. The most memorable visit I made was to a company that deploys robots for customers. I booked thirty minutes and stayed two hours. What I saw was not research. It was engineers holding a battery at a fixed current for fifteen minutes to check it is stable. It was somebody standing next to a camera adjusting the angle by hand. Mundane tasks, but you cannot hand them to just anyone. It has to be the mechanical engineer or the founder. And the task is only half of it. The other half is winning the customer's trust so they let you come back.
A loose wire means someone gets on a plane. Say you deploy an arm doing a simple pick-and-place task in another state or country. A wire comes off, or the camera is slightly off. The people on the factory floor cannot fix it. They only know it stopped. So you go. One founder told me about a friend's company flying an engineer from California to Spain to swap a memory card, at a cost of several thousand dollars. There is no remote fix for a lot of this yet. One company selling a remote supervision tool for deployed robots signed sixteen pilots in four months. That is how common the problem is.
The prerequisite is the nines. For any deployment, not just pick-and-place, accuracy has to get as close to the nines as possible, meaning 99.9 percent and up, and that goes for the hardware and the model together. A robot that works almost every time still fails several times a day, and every failure eats the savings that justified the robot. Customers expect near-perfect reliability and better than human speed. The founder running pick-and-pack facilities said getting the nines is the whole game and that mapping the edge cases is what wins. A hardware engineer who interned on a well-known humanoid program said the same: field issues, edge cases and reliability past a few hundred units are the real challenge. Demos are easy. Production is hard.
Customers will not wait around while you learn. An engineer at a big lab doubted that any customer would tolerate a robot that works most of the time for long enough to collect useful failure data. He expects a CFO moment in a few years, like the autonomous driving shakeout, when someone asks what the ROI actually was. I think he is right and that most deployment startups have not priced it in.
Trust, access and insurance. More than one founder said the real bottleneck is not the technology but getting in the door. A company with tens of thousands of robots deployed told one of them they simply had no bandwidth to work together. Insurance is underdiscussed: one deployment company said a major reinsurer had no idea how to price robotics and floated a flat monthly number per robot. Lab robots are still rare in US labs, and the two-person team I met building one said insurance was their biggest cost.
Robotics scales linearly. Because of all this, deployment goes one site at a time, not exponentially like software. You cannot vibe-code a robot arm into a factory the way you can ship a SaaS tool and send a Stripe link. Someone has to build the facility, install the arms, change how the employees work, and come back when it breaks. What is possible in two years will be mind-blowing. The world itself will not look that different. One person who has worked on autonomy for years said he does not expect robots to be unboxed and safe without integration work for at least a decade. The upside of that: integration skills and customer relationships are a real moat, which is the one moat I did not have in my last company.
Hardware and supply chain
Software is already a commodity. Models will be next. What is left is hardware, and hardware is where I now spend most of my attention.
Hands are the hardest problem. Nearly every researcher I met named manipulation as the thing that is not solved. Many humanoid demos have no hands at all, or plastic hands that do not move. Commercial dexterous hands are expensive, fragile and break often, which discourages the research that would make them better. Sensing is still mostly cameras, while real dexterity depends on touch. Tactile skin, force sensing and even smell came up as open gaps. One mechanical engineer from automotive thinks hands and arms become commodities eventually and that the next frontier is forward simulation of physics on the robot itself, not path planning.
Actuators and batteries, not AI, limit most robots today. One founder put it bluntly: make the vision model fifty percent faster and the robot gets about ten percent faster, because the physical system is the slow part. Battery life caps drones and warehouse robots at a few hours, and adding batteries adds weight. One solo founder has a hydrogen fuel cell proposal in with the US Navy to push that to a day. The existing actuator suppliers, legacy industrial firms on one end and hobby servo makers on the other, are a poor fit for new robotics companies, so a few startups are building actuators from scratch.
Compute on the robot is a coming bottleneck. If the brain is a frontier-size model, it needs more memory and compute than fits on a robot today, even heavily compressed. Models with a few billion parameters are already hard to run on edge boards, and compression costs precision. One founder I met is building an optical chip for exactly this. Expect the chip on the robot to matter as much as the chip in the data center.
China is years ahead on hardware, and the gap is density. A prototype takes hours in Shenzhen and weeks or months in the West. The reason is not one factory. It is thousands of specialized suppliers next to each other who all feed each other. That ecosystem is very hard to build from scratch. A friend who was in Shenzhen last week was told he could have a custom PCB back in a day or two. In the US I hear weeks.
The US is pushing back, and that will shape where robots get built. One founder who sells into that market estimates that most US robotics startups are now chasing Department of Defense contracts, which means dropping Chinese parts, and there are only a handful of made-in-USA alternatives. The FCC has started restricting certain foreign-made robots and domestic-content rules are tightening. Certification and sourcing can stretch a launch to a year or two. One founder is building a compliant parts marketplace just to cut that time. None of this changes how fast robots get smart. It changes who is allowed to sell them where.
Hardware is also the moat. Several founders told me the same thing unprompted: do not do a pure software play in robotics. Robotics founders get pitched twenty tools a week, a big lab could copy your software in days, and the data space is crowded. Hardware is hard to replicate, slow to build and proprietary. Combined with customer relationships, it is the only moat I have found that holds.
Longer term, I think printing wins. If China leads because of supplier density, the end game is a factory that can print almost any part, at which point geography matters much less. More than one person is already imagining it. When people are imagining it, it is a matter of time.
Form factors and verticals
The world is built for humans, so a lot of people conclude the robot should be shaped like one. I think that is the wrong conclusion for the next few years.
Humanoids are a gimmick outside entertainment, for now. Almost nobody I talked to who actually deploys robots is building a humanoid. A deployment company that sells a robot operating system is hardware-agnostic and openly bearish on legs; they run quadrupeds and wheeled bases with arms, and expect investment to move that way. The humanoid startups I heard about are doing dancing and fighting shows while quietly collecting teleoperation data for later. I was told of one accelerator-backed company that sold a cheap humanoid and was gone within the year. Entertainment is a real business and I am not dismissing it. Expecting humanoids to do anything else at scale in the next couple of years is unreasonable.
Narrow, hyper-specific robots are what actually ship. Welding robots. Baggage handling. Pipe bolt removal. Turning off breakers. Bell pepper harvesting. Lab sample handling. A desk arm that holds your monitor and follows your face. These sound small and they are, but they are the robots with customers. Single-task robots are what deploy, and they tend to come from teams of ten or fewer who raised single-digit millions. The pattern is: pick one task, own the whole stack from hardware to deployment, get to the nines, then expand to the next task in the same building.
The home vacuum is still the best business in robotics. It came out more than twenty years ago and nothing has beaten it. Robotaxis are on the streets and it is still unclear whether they will ever be a good business. Surgical robots are precise and still fighting adoption after ten years. Two separate people told me there is little value in home robots beyond the vacuum. That is a sobering fact for an industry raising this much money.
Defense is where a lot of the money is. Militaries want tanks, trucks, drones and robot dogs that drive themselves, and the budgets are enormous. Low-volume, high-value hardware is the one place where building hardware outside China clearly makes sense. Two European defense drone companies became unicorns in two years. The civilian drone market, by contrast, is crowded and owned by one Chinese company. I am avoiding defense myself because sales cycles are slow, but if you want to raise in the US right now, it is the easiest story to tell.
Dangerous, dirty and underwater work is underexplored. Several people pointed me toward hazardous industries: chemical plants, oil and gas, mines, rust removal, climbing robots, subsea cable repair. The logic is simple. The work is expensive, it is hard to staff, nobody wants to do it, and legislation is thin. One experienced founder's advice on data was to target the highest paid jobs per hour, like underwater divers, because that data stays valuable even if world models eat everything else. It is not a coincidence that I ended up underwater.
Fleet management is a real but early opportunity. Factories will run robots from several vendors with different models, and someone needs to orchestrate them, track failures and predict part breakage. Two different people independently proposed this as the startup they would build. The catch is that real deployed fleets are still tiny, so the timeline is far out.
Money
Investor interest in robotics is as high as I have seen for any sector. A lot of that money is going into things that will not matter in two years.
Many investors still judge robotics with SaaS logic. They ask about ARR. Hardware revenue looks nothing like that early on: it is lumpy, it comes with installation and service, and the first customers are pilots. An engineer at a big lab and I agreed on the fix: do not take money from investors who cannot see that. There are enough who can.
The data gold rush is a short-term play. It is surprising how many VC-backed startups are betting on the hunger for data growing for years. My read is that it grows for a while and then falls off a cliff as labs become self-sufficient and simulation improves. The market itself is a strange shape. The buyers are AI labs with no revenue, paying with money they raised. So it is venture money paying venture money. One advisor said a few hundred people in the world buy robotics data at all. That is not a market, that is a mailing list.
Deployment itself is becoming the product. This surprised me most. Deployment is so hard that some companies and some investors treat it as the end product. VCs now back companies that deploy for the sake of deploying, to collect edge cases and failures, either to sell to labs or to hoard for when better models arrive. More than one startup has pivoted from automating a task to just deploying and selling the data. One was building robot hands for a cloud kitchen. Another built lab robots and now mostly films humans doing the lab work. I understand the logic. I also think it is a trap, because it means your customer is a lab with a budget that can disappear, not a business with a problem.
Some investors will not fund anything that contradicts the big labs. Several people told me they got passed on because their thesis competed with what OpenAI, Anthropic or Google might do next. Real-world deployment, by contrast, appeals to almost everyone, because nobody thinks a frontier lab is going to install arms in a fish factory in Finland.
Size the problem before you build the robot. The best advice I got on money came from a founder whose earlier hardware company struggled because the problem was real but too small for venture capital. Market size dictates the team, the resources and how sophisticated the robot needs to be. Comparable narrow robotics companies raised high single digits to low teens of millions with teams of about ten. If your problem cannot support that, build a small hardware business instead and do not pretend otherwise.
Where VCs actually want to be right now. From what I heard: robotic hands and tactile sensing, robotics brains, defense autonomy, and vertical deployment with owned customer relationships. Where they do not want to be: generic egocentric data collection, humanoid data, and picks-and-shovels software sold to robotics startups.
Where the bottleneck goes next
The bottleneck in robotics moves. My best guess at the order, and what each stage means for anyone building now:
1. Data, today. Everyone is collecting and buying, mostly blind. This stage ends when labs have what they need from the real world and the rest comes from simulation. I think that is one to two years away.
1. Models. Once data stops being the constraint, the question is whether a general model or a specialized one runs the robot, and how fast the general ones close the gap. I think they close it within a couple of years. The reflex layer stays separate either way.
1. Deployment. Even when the model is good enough, somebody has to understand the customer's workflow, install the hardware, build the evals, diagnose the failures and come back when the wire falls out. This is where the industry is going to spend the second half of the decade. It is an operations business, it scales linearly, and it is where the proprietary data gets created.
1. Hardware. When simulation and general models deliver, hardware becomes the limit: hands, actuators, batteries, tactile sensors, and the chip and memory on the robot. Whoever can iterate hardware fastest wins, which today means being near Shenzhen or building your own ecosystem.
1. After hardware. When hardware is a solved, sold product, we are living in utopia. I do not know the timeline. Neither does anyone I talked to.
Two things people living in the current stage get wrong. They assume the bottleneck they are solving is permanent, so they build a company around data collection or a specialized model as if the constraint will hold. And they fail to see how fast the stages move. What was possible a year ago looks like a different world from today. The same will be true a year from now.
Hot takes
The thing that surprised me most is not on any of the lists above. People in the same field live in completely different realities. A data broker, a hands researcher, a deployment founder and a VC describe the same industry in four incompatible ways, and each is right about the part they can see. The most useful thing I did was talk to all four in the same week and notice who was selling me their own position.
With that caveat, here is what I believe that most of them would argue with:
1. The most profitable humanoid deployment in the next two years will be in the sex industry. Everyone laughs, and then nobody can name a humanoid business that is closer to profitability. Outside entertainment, humanoids are a gimmick until the hands work.
1. Simulation replaces most data collection, not the last mile of it. Most practitioners told me the opposite. I think they are describing the next two years and I am describing the next five.
1. The brain of the robot will be a general model, not a vision model. VLAs are ahead on a slow road. General models are on a fast one. The only thing that stays specialized is the reflex layer, and that may end up being plain control software.
1. Fighting the frontier labs is a bad idea. Build on them, orchestrate them, or stay out of their way. Several companies I met have already shifted from building one robot brain to orchestrating many models from many labs. That is the right move.
1. The home vacuum will still be the best robotics business in 2028. The demos will get better every month. The kitchen will look the same.
1. Most of the egocentric data startups raising today will not exist in three years. The ones that survive will have become tooling, evaluation or vertical deployment companies, and several already know it.
1. You cannot build a robotics company without touching the hardware yourself, yet. One day the stack will be standard enough that a commercial founder can buy and integrate it. Today you have to learn it. I bought an arm for that reason.
1. China's hardware lead is about density, and the US will not close it by onshoring. It will close it, if at all, when printing replaces supplier ecosystems. That is a decade away and it changes the geography of everything.
What I am doing with this
I am building an underwater robot that cleans the hulls of cargo ships while they are in the water.
Today that job is done by divers, or by remote-controlled machines with a person at the controls. It is dangerous, slow and expensive, and ship owners pay for it anyway, because a fouled hull burns more fuel and sends the ship to dry dock more often. Regulation is pushing shipping to use less antifouling paint, and the alternative to paint is brushing the hull regularly. I served in the Finnish Navy and I have owned and sailed boats my whole adult life, so I have scraped hulls myself. The people I have talked to in the industry confirm there is paying demand.
It is a narrow, dirty task in a place with few competitors, which is exactly what the last four weeks told me to pick. It also leads somewhere bigger: a new kind of harbor service around keeping ships clean, fuel-efficient and out of dry dock. The first prototype will scan a hull, not clean it. Hardware sourcing is undecided. I am going to Shenzhen in January to find out how fast a prototype can really move.
If you want to break into robotics. You do not need a robotics degree, but you need to be willing to touch the hardware. Almost every player in the field has open-sourced something. Standardized kits like Hugging Face's SO-101 arm let you collect your own data, fine-tune a model and feel the pain yourself for a few hundred dollars. The robotics meetups in San Francisco have been running weekly for years and are generous with what they know. Hackathons are the fastest way in: I built Edgecase Miner at Encord's Long Horizon hackathon before I even had an arm of my own, and learned more in that weekend than in a month of reading. Ask your dumb questions in person.
The most valuable skill I found. Being wrong fast. I changed my view of this space several times in a month, and every time it was because I asked a dumb question. The people falling behind in this field are the ones protecting their ego and what they already know.