The $47 billion bet that hasn't paid off — yet
In a little over six months in 2026, "physical AI" startups raised $47.4 billion in venture capital — nearly four times the amount invested in H2 2025. The number sounds staggering, but Amulya Vishwanath offers a simpler explanation for why the sector still hasn't experienced its inflection point: robotics is nowhere near its "ChatGPT moment." She says the industry is still in the "LEGOs era."
Vishwanath spent seven years leading product marketing for NVIDIA's robotics platforms before founding Techable VC, an early-stage venture firm focused on physical AI, materials science, and life sciences. Last Monday (Oct 13), in an interview with TechCrunch's Equity podcast, she laid out an uncomfortable argument against the current hype: despite billions being invested, the fundamental building blocks — actuators, robotic hands, control software, safety standards — are still being rebuilt piece by piece.
The comparison with the ChatGPT launch isn't accidental. In November 2022, a chatbot stunned the world not because it was technically perfect, but because it was legible: anyone, even without technical training, could talk to it and be impressed. In robotics, that legibility doesn't yet exist. An industrial arm that folds parts on an assembly line works, but it's not something a layperson would intuitively understand. A humanoid robot that walks in a lab looks impressive until it tries to pick up an apple.
The debate: general models vs. specialized components
Vishwanath's thesis doesn't exist in a vacuum. In the same Equity episode, General Intuition CEO Pim de Witte presented the opposite view: a single general-purpose foundation model would make most specialized robotics work obsolete. General Intuition, founded in 2014 and launched publicly in 2025, was valued at $2.3 billion in its Series A and trains models on billions of gameplay clips so AI agents learn to perceive, predict, and act in virtual and physical environments.
This is the same bet that made LLMs triumph: instead of building dozens of specialized systems for each task, train one single model capable of doing everything. The problem — and here Vishwanath's practical experience inside NVIDIA matters — is that robotics involves matter. A language model can write a perfect essay without ever worrying about the shear force of a hydraulic actuator. A computer vision model can classify images without ever dealing with the friction of a gripper trying to hold a deformable object.
Vishwanath describes the industry as being in a "LEGOs era," where each component needs to be developed, tested, and integrated individually. Cheaper actuators. Robotic hands with better proprioception. Software that links perception and action without prohibitive latency. Safety standards that define what a robot can and can't do near humans. None of these building blocks is ready.
What the numbers say
The investment math is unambiguous. According to Crunchbase data cited by the Wall Street Journal, the $47.4 billion in H1 2026 included a single Waymo round — $16 billion in a Series D at a $126 billion valuation — which accounted for roughly a third of the total. Even without it, the pace still doubled year-over-year. In 2025, global investment in robotics and "physical AI" reached $26 billion, up from $4.2 billion in 2019, per PitchBook.
But the numbers hide another detail: most of those dollars are going to companies that are already operating, not experimental startups. Waymo, Anduril ($5 billion Series D), SpaceX ($75 billion IPO) — these are mature companies. Vishwanath's bet with Techable VC, by contrast, is at a much earlier stage: actuators, sensors, certification. Startups that won't make headlines in year one, but without which robotics' "ChatGPT moment" is just a pitch-deck slide.
The gap between lab and the real world
Vin Patel, in independent analysis, points out that the crucial distinction between these two sides of the debate isn't technical — it's about reliability. What keeps robots from leaving the lab and entering a warehouse isn't raw capability — it's failure handling. Language systems have already earned user trust because a well-bounded model with guardrails produces consistent results. Robotics, on the other hand, deals with irreversible physical consequences. An error can't be undone with "retry prompt."
This is where the "yet" in Vishwanath's headline matters. She doesn't say robotics will never have its ChatGPT moment. She says it hasn't happened yet. And that betting on a general-purpose model as the catalyst is, at minimum, premature — because it ignores decades of hardware work that simply has no equivalent in the software world.
What comes next isn't a model race. It's a marathon of components.
Sources: TechCrunch, Value Add VC, Vin Patel
✓ Independent sources cross-checked and verified before publishing
