The ChatGPT Moment for Robotics Is Here — And It Changes Everything for Founders
· Nia
The ChatGPT Moment for Robotics Is Here — And It Changes Everything for Founders
When ChatGPT launched in late 2022, it didn't just create a new product category — it created a permission structure. Suddenly, thousands of founders had permission to build things that previously seemed like science fiction. AI wasn't just for PhD researchers anymore. It was a tool anyone could wield.
We're about to see the same thing happen for the physical world.
A Robot That Can Screw In a Lightbulb
This week, WIRED's Will Knight visited Eka, a startup in Cambridge, Massachusetts, and watched their robot do something no commercial robot arm has ever done: screw in a lightbulb. Not in a lab demo with perfect conditions — with random objects thrown in front of it, the robot adapting in real time, gently manipulating objects it had never seen before.
Knight, who has covered robotics for over a decade, called it the most natural robot movement he's ever witnessed. The machine would swoop down, nip gently at objects a few times to get a grip, and lift them with an eerie fluidity. When Knight tried to take his keys back from the robot, it resisted for a moment before letting go and immediately looking for something else to pick up.
This isn't incremental progress. This is a phase change.
Why This Is Different from Every Other Robot Demo
I've seen a hundred robot demos. Most of them involve carefully controlled environments, objects placed at precise angles, and engineers nervously watching from behind laptops. Eka's demo is different because:
1. It generalizes. The robot handles objects it was never trained on — earplugs, hairbrushes, key rings with plush attachments. This isn't hard-coded behavior for specific shapes. It's learned dexterity.
2. It's founded by the right people. Pulkit Agrawal is an MIT professor. Tuomas Haarnoja is ex-Google DeepMind robotics. These aren't people overhyping incremental work — they're researchers who understand what "solving dexterity" actually means.
3. They claim it's a scaling problem now. According to the founders, they're "halfway there" to full dexterity, and the remaining challenge is scaling up their approach — not fundamental research breakthroughs. If true, this mirrors the LLM trajectory: once you have the right architecture, you scale.
The Trillion-Dollar Hand Problem
"Trillions of dollars flow through the human hand," Agrawal told WIRED. He's not exaggerating.
Think about every job that requires manipulation: warehouse picking, manufacturing assembly, food preparation, elder care, construction, agriculture. The total addressable market for a robot that can handle arbitrary objects with human-like dexterity is genuinely in the trillions.
But here's what matters for founders: you don't need to be Eka to capture value from this moment.
The Entrepreneurial Opportunity Stack
Every major platform shift creates layers of opportunity. The iPhone didn't just make Apple rich — it created the app economy, the gig economy, mobile payments, and thousands of adjacent businesses. Physical AI will do the same.
Layer 1: The Robots Themselves
Companies like Eka, Figure, and 1X are building the hardware and foundation models. This is capital-intensive, requires deep technical talent, and has long development cycles. If you're not already in this game with $50M+ in funding, you're probably not starting here.
Layer 2: Robot Operating Systems and Middleware
Every robot needs software to manage tasks, handle edge cases, and integrate with existing workflows. This is the "picks and shovels" layer — less glamorous but potentially more profitable per dollar invested. Think: task orchestration for robot fleets, quality control systems, human-robot handoff protocols.
Layer 3: Vertical Applications
This is where most founders should be looking. Once robots can manipulate objects reliably, specific industries need domain-specific solutions:
- Warehouse automation that doesn't require rebuilding the entire facility
- Food service robotics for restaurants that can't afford full kitchen redesigns
- Agricultural harvesting for crops that currently require human pickers
- Elder care assistance — not replacing caregivers, but giving them robot hands for lifting and physical tasks
Layer 4: Data and Training Infrastructure
Robots need to learn. Companies that build simulation environments, collect manipulation data, or create training pipelines will be essential infrastructure. This is analogous to the data labeling and training pipeline companies that rode the LLM wave.
Layer 5: Safety, Compliance, and Insurance
Robots operating in physical spaces near humans need certification, insurance, and safety monitoring. This is a massively underserved market that will explode as deployment scales.
Lessons from the LLM Wave (Applied to Physical AI)
Having watched the LLM startup wave closely, here are patterns founders should apply:
Don't compete on the model. Eka, Figure, and the other foundation players will build the best general-purpose manipulation models. Your job is to find a specific domain where you can fine-tune, add domain knowledge, and deliver 10x more value than a general solution.
Move fast on integration. The startups that won in the LLM era weren't always the most technically sophisticated — they were the ones that integrated fastest into existing workflows. The same will be true for robotics. The company that gets their robot working in an actual warehouse in 6 months will beat the one spending 2 years perfecting the algorithm.
Wedge in with a single task. Amazon didn't start by selling everything. Find one manipulation task in one vertical where the economics clearly work, dominate it, and expand from there.
Own the data flywheel. Every robot deployment generates manipulation data. The companies that structure their deployment to capture, clean, and learn from this data will compound their advantage over time.
The Timing Question
Is it too early? Founders always ask this. Here's my framework:
- If Eka's founders are right that dexterity is now a scaling problem, we're 18-36 months from commercially viable general manipulation.
- If you start building vertical applications NOW, you'll have domain expertise and customer relationships by the time the hardware is ready.
- The companies that started building LLM applications in early 2022 (before ChatGPT) captured massive first-mover advantages. The same window exists right now for physical AI.
What AI Spending Tells Us
The New York Times reported this week that AI capital expenditure has hit a new record "with no end in sight." Big Tech companies are pouring hundreds of billions into AI infrastructure. A significant and growing portion of that is going into physical AI and robotics.
This isn't speculation money — it's companies like Amazon, Google, and Tesla making long-term bets on the physical world being automated. When this much capital flows into a space, the ecosystem opportunities multiply.
The Bottom Line
We're at the beginning of a platform shift as significant as mobile or cloud computing. The difference is that this one operates in the physical world, which means the opportunities are larger and the competition is (currently) thinner.
If you're a founder thinking about what to build next, physical AI should be at the top of your list. Not building robots — building the applications, infrastructure, and services that make robots useful.
The ChatGPT moment for the physical world isn't coming. It's here. The question is: what are you going to build on top of it?
Eka is based in Cambridge, MA. Their work builds on reinforcement learning approaches developed at MIT and DeepMind. If you're exploring the physical AI space, the research coming out of Eka, Figure, 1X, and Covariant is required reading.