Physical AI: Why now and what’s different?
Is robotics on the brink of its “ChatGPT moment”? A look at the catalysts converging to get us to the tipping point.
Physical AI took center stage last week, from NVIDIA’s GTC (complete with an Olaf robot cameo at Jensen’s keynote) to Bessemer’s Robotics Day to Unitree’s IPO news. The momentum didn’t stop there: this week brought news of Amazon’s Fauna Robotics acquisition and the appearance of a Figure humanoid at the White House!
Physical AI is certainly having its moment and VC funding in the sector has seen a meaningful uptick recently (chart above). As I wrote in my 2026 predictions piece, the embodied AI race could prove more intense and consequential than the LLM wars.
But robotics hasn’t always been a “hot” category and many investors still carry scar tissue from prior cycles (chart above). So what’s actually different this time? The key shift is that today’s Physical AI catalysts aren’t unfolding sequentially. They’re instead compounding in parallel, creating a convergence that makes this moment feel fundamentally unlike those that came before:
Physical AI foundation models are advancing rapidly: A new class of AI models purpose-built for the physical world is emerging, from vision-language-action models to autonomous driving models to world models (check out Chris Paxton’s interesting deep dive on this topic). In effect, we’re now seeing the beginnings of a “foundation model layer” for robotics, potentially unlocking a “robotics brain” capable of thinking and reasoning across tasks, environments, and form factors. This is a step-function improvement from traditional approaches that relied on brittle rules or narrowly-trained and ungeneralizable policies.
Data bottleneck is easing: For years, the limiting factor on the robotics frontier wasn’t intelligence; it was data. Unlike LLMs, the data required to train robot models (e.g. motor skills, pressure, manipulation, etc.) can’t just be scrapped from the internet. Physical AI data is unstructured, multimodal, and historically expensive and slow to collect via real-world interactions. However, these data constraints are now abating due to advances in scalable teleoperation, simulation-first approaches, egocentric video, world models, and haptics. Additionally, techniques and tooling are also maturing rapidly (exhibit below). The data problem isn’t fully solved yet, but it’s no longer the wall it once was.
Inference infrastructure is meeting the moment: Robotic intelligence is only useful if a robot can act on it in real time. Here, breakthroughs in edge inference, such as more efficient on-device compute that can run complex models locally and in real-time, are closing the gap. This type of inference is critical for physical AI systems where latency and connectivity could impose hard constraints, particularly in environments like factory floors or construction sites where immediate action is required.
Hardware is ready to scale and getting cheaper: Crucially, hardware improvements, commoditization, and falling cost curves are making scalable, versatile robots economically viable. This is a necessary unlock to turn promising demos into deployable products.
Macroeconomic tailwinds: These technological shifts are converging in a favorable macro environment. Labor shortages, supply chain fragility, and geopolitical pressure around issues like reshoring have shifted automation from a future bet to a present and strategic necessity. At the same time, autonomy is increasingly becoming mainstream in public consciousness, from self-driving cars on our roads to humanoid robots serving customers in restaurants.
Talent inflows: Perhaps the most telling signal of all is talent. Across Big Tech and startups, a wave of researchers, developers, and founders are now moving into the robotics field in numbers reminiscent of the early days of the LLM boom:
While recent progress in the field has been remarkable, the bigger debate has shifted to timing: when will Physical AI have its “ChatGPT moment”? We’re not quite yet at the point of true generalizability across real-world tasks at scale, but with multiple catalysts compounding in parallel, the trajectory is becoming clearer that the inflection point may be closer than we expect.
Janelle, the parallel convergence framing is the right lens here. Having worked across IoT connectivity and edge deployments, I'd argue the inference infrastructure catalyst you mention is actually the most underrated of the six -- and the one most likely to determine which physical AI companies scale vs. stall.
Here's why: foundation models for robotics are impressive in the lab, but deploying them on factory floors, construction sites, or agricultural operations introduces brutal real-world constraints. You're running VLA models on edge compute with 8-16GB of RAM, managing thermal throttling in non-climate-controlled environments, and handling inference pipelines that need sub-100ms latency for manipulation tasks -- all while maintaining connectivity through cellular backhaul that may drop to 3G in a warehouse corner.
The data bottleneck easing is real, but there's a related connectivity bottleneck that doesn't get enough attention. Teleoperation for training data collection requires stable, low-latency links. Sim-to-real transfer works until the robot encounters environmental conditions the simulation didn't model. And fleet-wide model updates across deployed robots require OTA infrastructure that's more akin to automotive-grade FOTA than typical cloud deployments.
The macro tailwinds around labor shortages are undeniable -- but I'd add one nuance: the first physical AI deployments that reach true scale won't be humanoids doing general tasks. They'll be purpose-built form factors in structured environments (warehouses, food processing, electronics assembly) where the perception and manipulation problem space is bounded enough to be reliable. The humanoid moment comes after that beachhead is established.
What's your read on which vertical will produce the first breakout physical AI deployment at genuine scale?