Fortune just published a piece arguing AI models are choking on drunk data. The thesis is simple. Data labeling companies are flooding the market with low-quality training material in physical AI that would be robots, self-driving cars, systems that understand the real world is grinding to a halt because of it. They point to OpenAI shutting down Sora as evidence. Their world model couldn't grasp physics properly, so the whole thing collapsed under its own weight. I think this completely misframes what's actually happening. I'm L, I have a PhD in computer science, and I analyze AI developments to understand what's actually happening beneath all of the hype because here's what the drunk data narrative gets wrong. We didn't hit a data quality wall. We hit an infrastructure gap, and the difference matters enormously for what happens next. Let me explain what I mean. When people say AI, they're collapsing an insanely broad field into a single word. It's like saying medicine. There are a bazillion use cases and branches, and lumping them all together obscures more than it clarifies. Text-based AI, your ChatGPTs, Claude's, your large language models, could free write on existing human behavior. People were already writing online, billions of words sitting there in various languages ready to be scraped. Reddit posts, Wikipedia articles, news sites, books, research papers, you name it. The internet was a massive pre-existing training corpus that cost exactly zero dollars to generate because humans made it by just existing and expressing themselves.

Physical AI is a completely different problem. Think about what a robot actually needs to know to fold a shirt. It needs to understand how much force to apply so the fabric moves but doesn't tear, how gravity affects different materials. Silk drapes differently than denim, it's example. How the fabric's edges behave when you pinch them, what fold it even looks like from multiple angles. This is not information you can scrape from a Wikipedia article. You need sensor data. Robots need to understand how objects move through space. Autonomous vehicles need to know what happens when you brake on wet asphalt at 40 mph versus dry concrete versus gravel. World models need to grasp elasticity, friction, momentum, how light refracts through rain on a windshield at night, the entire physics engine of reality. And here's the thing, that data doesn't exist at scale yet because it requires active instrumentation. We need sensors. Temperature sensors measuring thermal expansion of materials, motion sensors tracking three-dimensional movement, pressure sensors detecting force distribution when a robotic gripper touches something, cameras with depth perception calculating distances, accelerometers measuring changes in velocity, gyroscopes tracking orientation in space. You need to actually measure the physical world in real-time, continuously, across millions of different scenarios and edge cases, and we are just not there yet.

Now, before you think that I'm being pessimistic, let me show you what's actually going on beneath the surface. The IoT sensor market, that's Internet of Things, basically any physical sensor that connects to a network and sends data somewhere, was valued at 18.34 billion dollars in 2024, projected to hit 422.13 billion dollars by 2034, growing at a compound annual rate of about 37%. That is enormous growth. As of the end of 2025, there are about 21.1 billion IoT devices deployed worldwide. For context, that's roughly three sensors for every person on the planet, we're talking about temperature monitors in buildings, wearables tracking your heart rate, motion sensors optimizing factory floors, pressure sensors in industrial equipment, humidity sensors in pharmaceutical storage, light sensors adjusting building illumination, stuff like that. In North America alone, the IoT sensor market is expected to grow from 8.52 billion dollars in 2025 to 136.06 billion dollars by 2034. The fastest growing segment is motion sensors, expanding at 15.8% annually, driven by smart buildings trying to understand how spaces are actually being used so they can optimize energy and reduce waste. This is real infrastructure being deployed at scale.

But here's the critical question. Why isn't it growing even faster? If sensor data is the bottleneck for the next wave of AI, why aren't we seeing Manhattan project levels of investment in sensor deployment? I think the bottleneck isn't technical, it's incentive alignment. Let me break down what I mean by that. When a technology gets deployed at massive scale, it's because someone sees a reason to pay for it. With text-based AI, the deployment was basically free. The internet already existed. Companies like OpenAI and Google just needed to build a scraping infrastructure and training systems, but IoT sensor deployment requires upfront capital investment with delayed returns. You need to manufacture sensors, install them, maintain them, build the data infrastructure to collect and process what they're measuring. And here's the problem, IoT sensor deployment isn't sexy like large language models. Just think about it from a funding perspective. If you're a venture capitalist in 2024, would you rather invest in A, a company building conversational AI that can write code and have philosophical discussions with viral demo potential and clear consumer appeal or B, a company installing temperature and humidity sensors in commercial buildings to optimize HVAC systems. One of these makes headlines, one of these gets TechCrunch articles and Twitter buzz, the other is infrastructure, critically important but kind of boring. You can't have a conversation with your building's HVAC system. There is no viral demo where a pressure sensor writes you a poem. So, IoT gets less funding, less hype, less attention, even though it is arguably more foundational for the next wave of AI development than yet another chatbot.

And here's the other problem. People won't opt into ubiquitous monitoring without seeing clear, immediate benefits. Think about it from a user perspective. If I told you, "Hey, we want to put sensors all over your house to monitor temperature, occupancy, energy usage, movement patterns." Your first question would be, "Why? What's in it for me?" And my answer is, "Well, eventually this data might help train better AI models." You tell me to get lost, rightfully so. The value proposition has to be direct and immediate, not abstract and future-oriented. But what if instead I said, "These sensors will cut your electricity bill by 28% by automatically optimizing when your heating and cooling run?" Now we're talking, right? It's not hypothetical. A university deployed smart building sensors across their campus and documented a 28% reduction campus energy use intensity, equivalent to 310,000 US dollars in annual utility savings. The system also identified bearing degradation on six air handling units, preventing failures estimated to have cost 280,000 in emergency repairs.

Here's how that actually works because the mechanism is super important. Traditional buildings run HVAC systems on fixed schedules. Heat turns on at 6:00 a.m., off at 6:00 p.m. regardless of whether anyone is actually in the building. With IoT sensors, you're measuring occupancy in real time using motion and CO2 sensors. The system essentially learns conference room 3B is only used Tuesday and Thursday afternoons, so it doesn't heat that room for the rest of the week. Fair enough. Temperature sensors detect when a space is already at the target temperature, so the system doesn't waste extra energy over shooting. The building becomes responsive instead of running on dumb timers. The energy savings are real and immediate. Your utility bill goes down this month. That is the value proposition that gets people to opt in. Or take health care. Google and Fitbit launched Device Connect, which allows hospitals to integrate wearable data directly into electronic health records. The pitch is not, "Please help us train better AI." The pitch is, "Enable real-time cardiac monitoring so your doctor can catch arrhythmias before they become emergencies." Studies show that physician monitoring of Fitbit data embedded in routine primary care visits is logistically feasible, and patients are actively interested in sharing exercise and health data with their providers when it leads to personalized recommendations and early intervention. This is how you build the infrastructure, not by making grand proclamations about AI that understands the physical world, but by deploying narrow, high-value applications that justify the cost of instrumentation.

Each successful deployment creates a data flywheel. Sensors get deployed because they solve a specific problem. They generate data as a byproduct. That data enables better optimization and new insights, which justifies deploying more sensors in adjacent areas, which generates more data, and the cycle compounds. Let's talk about what happened to OpenAI's Sora for a second, because I think the Fortune piece gets the diagnostics exactly backwards. Sora was OpenAI's text-to-video model. The idea was basically you type a prompt like a golden retriever playing in the snow, what a cutie, and Sora generates a realistic video of that scenario. But it was not just a video generator, it was supposed to be a world model, a system that understands physics well enough to simulate how objects move, how light behaves, how materials interact. They launched it in late 2025 with a big splash, got Disney to commit to a billion-dollar partnership where users could generate videos featuring Mickey Mouse or Marvel characters, positioned it as the future of content creation. Six months later, they shut it down.

According to reporting from the Wall Street Journal and TechCrunch, Sora was burning through roughly $1 million per day in compute costs. Some estimates put it even higher. One analysis suggests around $15 million per day when you factor in all the infrastructure costs. User growth peaked at around a million people, then collapsed to fewer than 500,000. Revenue from in-app purchases totaled an estimated $2.1 million over the app's entire lifetime. Let me just put that into perspective. If Sora was burning a million dollars per day and only generated $2.1 million in total revenue, it would take 2 days of operating cost to consume the entire lifetime revenue. The unit economics were catastrophically bad and OpenAI pulled the plug. Now, here's where it gets interesting. The Fortune article frames this as a junk data problem. Their world model lacked sufficient understanding of physics because the training data was apparently low quality, and the implication is basically if only they had better data labeling, better simulated scenarios, Sora would have worked. I think that scenario is missing the forest for the trees. OpenAI tried to simulate their way to understanding the physical world. They fed the model synthetic data, computer-generated scenarios of object falling, bouncing, interacting, stuff like that. They tried to teach it physics by showing it endless variations of simulated interactions, and fundamentally it didn't work, not because the data was junk, but because simulations cannot replicate the complexity of reality.

Let me explain why simulation breaks down in these cases. When you simulate physics in a computer, you're always using approximations. You define coefficients for friction, elasticity, air resistance, but reality is a lot messier. A piece of fabric doesn't have one coefficient of friction. It depends on humidity, how worn the material is, the microscopic texture of the surface it's sliding across, whether there's dust or oil present, stuff like that. A robotics training guide notes, I'm quoting, "A microscopic surface variation creating friction can make the real world difficult for robots well trained in simulation." Or think about water, another example. Simulating fluid dynamics is computationally expensive and still doesn't capture edge cases, like surface [snorts] tension behaving differently at temperature boundaries, or how soap molecules change flow patterns. You can get close, but close isn't good enough when you're trying to train a robot to wash dishes or a world model to generate realistic videos of rain. This is called the sim-to-real gap. You can train a robot in simulation to walk perfectly on flat ground, but the moment it encounters real-world irregularities, slightly uneven floors, unexpected friction changes, vibrations from nearby machinery, it stumbles. The simulation doesn't prepare it for the long tail of reality.

You know what does work? Real-world data collection at scale. Tesla has logged over 8 billion miles driven on full self-driving supervised as of February 2026, their fleet generates 160 billion daily video frames from over 2 million vehicles. Let me put that into perspective for a second. 8 billion miles. If you drove 24 hours a day at 60 mph, it would take you over 15,000 years to accumulate that much driving experience. Tesla's fleet did it in a few years, so that is a remarkable achievement. Every single one of their vehicles is a rolling sensor platform continuously feeding data back about how the real world behaves. What does a pedestrian look like in fog? How do shadows affect object detection at sunset? What happens when someone drives the wrong way down a one-way street? What do lane markings look like when they're faded or covered in snow or painted on old asphalt versus new asphalt? Tesla knows because it happened and their fleet recorded it. Here's how their data flywheel actually works. Step one, fleet collection. Every Tesla on the road with cameras enabled is collecting video and telemetry. Most of this data is discarded locally. You don't need to upload footage of routine highway driving in clear weather because Tesla already has millions of examples of that. Step two, shadow mode. Tesla runs two systems simultaneously. The production autopilot system that's actually driving and experimental models running in shadow mode making predictions but not controlling the car. When the shadow model disagrees with the production model or when it's uncertain about something, that scenario gets flagged. Step three, targeted retrieval. Tesla sends a request to specific vehicles. Next time you encounter a scenario matching these specific parameters, upload the footage. This is how they collect edge cases at scale. They're not randomly sampling stuff. They're actively hunting for the scenarios where their models are uncertain or wrong. Step four, a super important one, human annotation. The tricky scenarios get sent to human labelers who annotate what's actually happening. That's a pedestrian partially occluded by a mailbox. That's a motorcycle lane splitting. That's a debris on the road, not a stationary vehicle. Step five, model training, the coolest part. The new label data goes into training the next version of the model, which then gets deployed over the air to the entire fleet. Step six, continuous validation. As the new model runs, Tesla monitors its real-world performance and safety metrics across millions of miles. If it performs worse, they roll back. If it performs better, it becomes the new production baseline.

This is the data moat, not necessarily better algorithms, not bigger compute clusters, owning the sensors and having a built-in distribution network of millions of cars continuously generating training data. And here's the thing that should make you pause. No generalist AI lab can compete with that without building their own sensor infrastructure. Waymo has sensors, but they're on maybe a few thousand vehicles operating in limited geographies. Cruise had sensors before they shut down operations. Traditional automakers are trying to build sensor fleets, but they're years behind Tesla in terms of data accumulation and software integration. OpenAI, Google's AI division, Anthropic, they don't own cars. They don't have sensor infrastructure deployed in the physical world at scale anywhere. So, what do you do if you're a generalist lab trying to build physical AI without sensor infrastructure? Well, you kind of have to simulate. Let me show you why that keeps failing. OpenAI tried to simulate the physical world and burned a million dollars a day doing it, which is an extraordinarily expensive way to discover that friction exists. But here's why they tried in the first place. Simulation is scalable in a way real-world data collection isn't right now. If you need 10,000 examples of a robot picking up a coffee mug, you can generate those in simulation in a few hours. In the real world, you need to physically set up the scenario, run the robot 10,000 times, and hope nothing breaks. It's slow and very expensive.

But, simulation has a fundamental problem. It can only teach the model what you programmed into the physics engine. If you did not account for how coffee mug handles can be slightly oily from manufacturing residue, or how ceramic behaves differently at different temperatures, your simulation will just not capture that. MIT Technology Review reports that virtual simulations can train robots to perform acrobatics, but not how to grasp and move objects because simulations struggle to model physics with perfect accuracy. So, what's actually happening in robotics, right? Companies are pivoting to real-world data collection at scale, even though it's expensive and slow. They have to. There are now training centers in China where people wear exoskeletons and VR hardware while doing the same repetitive tasks hundreds of times per day. Gig workers in Nigeria, Argentina, and India are filming themselves doing chores at home. Scale AI and other data companies are recruiting armies of people to record themselves folding laundry, washing dishes, organizing shelves, capturing the messy, varied reality of how humans actually manipulate objects. Investors poured over $6 billion into humanoid robots in 2025, and a huge chunk of that money is going towards data collection infrastructure, not algorithms, not compute, data, the holy grail of computer science.

The pattern is clear, isn't it? Real-world data beats simulation every single time. And real-world data requires real-world sensors. So, if simulation doesn't work, and you need real-world sensors, how do you actually get them deployed at scale? Look at what's working right now. In manufacturing, the industrial IoT segment is growing at 37.5% annually. Why? Because factories are deploying sensors for predictive maintenance. Instead of replacing parts on a fixed schedule, sensors monitor vibration, temperature, and acoustic signatures to predict when a machine is about to fail. Deployments demonstrate 25% maintenance cost savings and 70% downtime avoidance. That's real money saved immediately, which justifies the sensor investment immediately. As a byproduct, they're also generating massive data sets about how industrial equipment actually behaves under stress.

As a next use case, let's look at health care. Over 300 million wearable devices will be in use worldwide by 2026. Most of them fitness trackers and smartwatches with heart rate monitors, accelerometers, and increasingly blood oxygen sensors. Google and Fitbit's device connect integration allows this data to flow into electronic health records. Patients opt in because they get continuous health monitoring and early intervention. And as a byproduct, health care systems are accumulating data sets about human physiology at unprecedented scale. Next one, let's look at smart cities. Smart cities represent 26% of the global IoT market. That is a huge chunk. Over 600 cities worldwide have implemented IoT solutions for traffic management, water distribution, and waste collection. IoT powered traffic lights and sensors are operational in 145 US cities, reducing congestion by an average of 12%. Cities deploy these sensors because they reduce infrastructure costs and improve quality of life. As a byproduct, they're generating real-time data about urban dynamics that could eventually train AI systems managing city infrastructure.

Do you notice the pattern? Every deployment has an immediate domain-specific value proposition that justifies the infrastructure investment, the AI training potential is a secondary benefit, not the primary driver. This is what I mean by incentive alignment. You can't convince cities to deploy sensors by saying, "Please, this will help train better AI models someday." You convince them by saying, "This will reduce traffic congestion and save your residents 15 minutes per commute." The data accumulation happens as a consequence of solving real problems. Now, I know that some of you are thinking, "Al, this sounds like a surveillance dystopia, sensors everywhere monitoring everything feeding data into AI systems." I totally get it, and I think this tension is definitely worth taking seriously. There is a version of ubiquitous sensor deployment that's extractive and coercive. Sensors tracking your movements to sell behavioral data to advertisers, workplace sensors monitoring bathroom breaks to squeeze more productivity out of workers, insurance companies demanding access to your health data to adjust premiums. That's monitoring of people, treating humans as data sources to extract value from.

But, there is another version, my friends. Sensors that reduce your energy bill, and you can see the savings on your utility statement every month. Health monitors that catch cardiac arrhythmias before they become heart attacks, building sensors that optimize air quality and reduce respiratory illnesses. That is monitoring for people, technology deployed to improve human flourishing, the best kind of technology. The difference isn't the sensors themselves, it's who controls them, who benefits from the data, and whether participation is genuinely optional with clear value exchange. I personally think the path forward requires transparency. You should always know what sensors exist in your environment, what they're measuring, and where the data goes. Opt-in with value exchange. Participation should be voluntary, and the benefits should flow primarily to the people being monitored, not to some third random party. Privacy by design is a must. Data should be anonymized and aggregated whenever possible. Processing should happen locally before data ever leaves the device. You can measure building occupancy with motion sensors without identifying specific individuals. There's methods for that. Human-centric deployment, the question shouldn't be what AI applications could this data enable? It should instead be what problems can we solve for people that happen to generate useful data as a byproduct. Let's feed two birds with one stone, right? These are not just philosophical niceties. They're pragmatic requirements for getting people to actually opt in at scale. And by the way, we already generate massive amounts of data just by existing. Heat signatures from our bodies, how we use appliances, movement through space, electricity consumption patterns, we're just not capturing a system- atically. The question is not whether data generation happens, it already does. The question is whether we instrument it in a way that benefit the people generating it.

Let me bring this back to where we started. The Fortune article argues that AI is choking on junk data, that companies like Scale AI are producing low-quality labeled data that's dragging down model performance. The author, who's by the way a co-founder of a computer vision company, claims the bottleneck is data quality, that we need better annotation, better filtering, better curation. I think this narrative is a red herring. The problem is not that we're producing bad data, the problem is that we're not producing enough of the right kind of data because the infrastructure to capture it doesn't exist at scale yet. Think about it. Scale AI can label images all day long. They can annotate bounding boxes around objects, classifying scenarios, describe what's happening in a frame, but they cannot create data about how a robot gripper feels when it touches different materials. They cannot generate sensor readings about how a car suspension responds to potholes at different speeds. They cannot produce thermal imaging data about how buildings lose heat through windows in sub-zero temperatures. That data requires physical sensors deployed out there in the real world measuring actual interactions. And the reason that infrastructure does not exist at scale yet is not technical. We know how to build sensors. We know how to deploy IoT networks. We know how to process and transmit data. The reason is as always economic and social. IoT deployment isn't as sexy as LLMs, so it sadly gets less funding and attention. People won't opt into sensor networks without seeing clear, tangible, immediate benefits. The business models for sensor deployment are domain-specific and long-term, which makes them less attractive to venture capital looking for explosive growth.

So, the actual bottleneck is incentive alignment in deployment strategy. We don't need better data labeling companies. We need to make IoT deployment compelling enough that people actually want it. We need Fitbit integrating with NHS so patients get better cardiac monitoring. We need smart buildings that visibly reduce energy costs. We need predictive maintenance systems that prevent expensive equipment failures. We need domain-specific wins that justify the infrastructure investment. And once that infrastructure exists, once the world is actually instrumented and generating real-world data at scale, then the generalist AI labs can build the world models we're all dreaming of. Not by simulating reality, but by learning from it. So, what does this mean for the next 5 to 10 years of physical AI development? I think we're entering a bifurcation. Path one, domain-specific winners, companies that control sensors infrastructure in high value verticals will dominate those domains. Tesla in autonomous driving because they have the fleet data mode, medical device companies in health monitoring because they own the wearable ecosystem and healthcare integrations, industrial automation companies in manufacturing because they've deployed sensors across factory floors. These companies will build AI systems that work extremely well in their specific domain, but don't generalize beyond it. Tesla's driving AI won't help you fold laundry. Wearable health monitors will not really navigate a warehouse. The second path I'm thinking of is infrastructure build-out and then generalization. I personally think this one is more likely. Generalist labs like OpenAI, Google, Anthropic can succeed in building world models and physical AI systems, but only after the instrumentation layer existed a scale where they can access the data either through partnerships or because sensor data becomes a commodity that multiple

players can access. The sequence has to be first deploy sensors through domain specific applications with immediate value propositions. This is happening now. The IoT market is growing. Second, let that data accumulate and prove its worth. People see benefits, adoption expands, network effects kick in. Third, data access models emerge. Maybe through partnerships like Google Fitbit, maybe through data marketplaces, maybe through open data sets similar to how ImageNet enabled computer vision breakthroughs. Fourth, generalist labs train world models on aggregated real-world data from multiple domains. Only at this point can you build AI that generally understands physical reality across contexts. We're somewhere between step one and step two right now. The companies that recognize the sequencing, that understand you cannot skip straight to generalist physical AI without the infrastructure layer, are the ones that will win. Everyone else is going to burn money trying to simulate their way to understanding reality because here is the thing about reality. It is undefeated. You can simulate friction, but eventually you kind of have to test your robot on a real surface and discover that hey, carpet behaves differently than your simulation predicted. You can generate synthetic driving scenarios a dime a dozen, but eventually you actually need to know what really happens when a car cuts you off in traffic during a snowstorm. Tesla has 8 billion miles of people teaching their cars what the real world looks like including all the edge cases and weird situations that no simulator would even think to generate. Everyone else has hopes, dreams, and a really good physics engine. I know which one I would bet.

If you're building in this space, here's what I think you should pay attention to. The companies that will win in physical AI are not necessarily the ones with the best models. They're the ones that control the sensors in high volume domains. If you're trying to build a robotics company, you need a data collection strategy that doesn't depend on simulation, which means either deploying your own sensors or partnering with companies that already have them. If you're working on autonomous systems, you need access to real-world edge cases at scale, which means you need fleet data or equivalent. And if you're an investor or just somebody trying to understand where all of this is going, watch where the sensor infrastructure is being deployed. Healthcare, energy, manufacturing, smart cities, those are the domains where real-world data is accumulating right now. That is where the next generation of physical AI will be trained. The IoT sensor market is projected to grow from 18 billion to 422 billion dollars over the next decade. That is not just hardware sales, that is the infrastructure layer being built that will enable the physical AI applications everyone is so excited about today. The scaling hypothesis that worked for language models, just feed them more data and they get smarter, Turned out to be actually white until it wasn't. Quality matters now, but quality in this context doesn't mean better labeled images from a data annotation company. It means real-world sensor data that captures physical interactions as they actually happen with all the messy complexity that simulations cannot replicate. We are not in a drunk data crisis. We are in an infrastructure buildout phase. The narrative should shift from we need better data labeling to we need more sensors deployed in high-value applications that justify their cost while generating training data as a byproduct. And the companies that recognize that first, that understand the path to physical AI goes through domain-specific sensor deployment, not through better simulation, are exactly the companies that will build the AI systems that actually work in the world.

I'm developing a course in two tiers for non-technical professionals where AI is explained simply, by the way, breaking down these kinds of technical mechanisms and infrastructure questions into frameworks you can actually use. If you'd like to take two minutes to answer some questions which directly shape what the course covers, the link is down below. And if you want to dig into why Gen Z people are actually hating AI and actively sabotaging it, check out the video on your screen right now. It's actually pretty wild how things are unfolding right now. Thank you so much for watching. Subscribe, and I'll see you all on the next one.