Robots will not have a "ChatGPT moment", but the boom has already begun.

Author: RoboStrategy Research Team

Translated by: TechFlow

TechFlow Summary: Within just two months of its launch, ChatGPT accumulated over 100 million users, igniting a frenzy of imagination around artificial intelligence across humanity. This also placed a longstanding question on the shoulders of the embodied AI and robotics sector: when will the ChatGPT moment for robots finally arrive? In this in-depth research report, the RoboStrategy team offers a starkly opposite, sober judgment: robots will never experience such a so-called ChatGPT moment.

The purely digital outputs of the software world naturally allow for trial-and-error and post-hoc manual fine-tuning, while cheap cloud computing infrastructure drives their marginal distribution costs close to zero; conversely, interactions in the real physical world are constrained by heavy capital hardware burdens, irreversible costs of failure, and highly strict physical safety boundaries.

Every hour a robot sits idle taxes its asset returns, and it must scale a reliability cliff of 99.9% or higher just to earn a ticket to commercial deployment. But this does not mean embodied AI is falling behind. On the contrary, from Waymo's autonomous vehicle fleets hitting the roads at scale to the vertical penetration of embodied models in welding, food service, and logistics, Physical AI is leveraging multimodal foundation models and Scaling Laws to follow an adoption curve similar to autonomous driving technology: gradual progression followed by sudden surges. By trading reliability for distribution, the true industrial boom has quietly already ignited in the deep end of the pool.

Look back to November 30, 2022. OpenAI launched ChatGPT.

Five days later: Users reached 1 million.

Two months later: Users surpassed 100 million.

That was a starting pistol echoing worldwide. The unprecedented pace of user adoption was only part of that "moment." What truly mattered was that the collective global consciousness almost instantly synchronized a paradigm shift in thinking: AI was no longer just a scientific fantasy confined to laboratories. It was undeniably real and poised to ignite the next technological revolution. C-suite executives across every industry made AI planning their top priority. Investors rushed to back potential winners or urgently exited companies ripe for disruption. Entrepreneurs flocked to launch new ventures in semiconductors, biotech, and robotics. AI was destined to sweep every field. Ordinary people around the world began incorporating AI chatbots into their daily lives.

Since then, everyone in the robotics field has been asked one consistent question: When will our own robotics ChatGPT moment finally arrive?

The robotics field will never see any so-called ChatGPT moment.

Figure 1: Jensen Huang stands on stage alongside leading humanoid robots, discussing whether the robotics "ChatGPT moment" has already arrived.

Intelligent robots have already arrived among us. They will be adopted at scale like any transformative technology before them, but this will certainly not happen in a matter of weeks; it will take years.

The internet, smartphones, and more recently, autonomous vehicles, all prove this rule. Anyone still desperately waiting for a robotics ChatGPT moment overlooks a simple fact: the starting pistol fired years ago, and the first wave of real-world large-scale deployment has already begun this year. From UPS to BMW to Samsung, major corporations are pouring relentless resources into automating operations with robots. Top talent is flocking to the field, building an entirely new enterprise ecosystem for Physical AI, with capital following close behind—albeit still a fraction of the massive funds that have flowed into pure software AI over the past few years.

So why is it impossible for us to witness a growth miracle on robots matching ChatGPT's explosive rise from zero to 100 million users in just two months? Beyond the physical hardware manufacturing constraints of producing 100 million robots, the capability requirements for deploying robots differ fundamentally from those of general AI models. We often forget how flawed ChatGPT was upon its initial release. It frequently hallucinated, provided answers riddled with errors, and couldn't even solve problems within a middle schooler's grasp. Yet in certain large-scale use cases, these flaws didn't matter much. In scenarios like copywriting, text summarization, and basic research, it still demonstrated practical utility. It didn't replace entire workflows, but it saved ordinary people a bit of time in their daily tasks. Essentially, it was a decent, albeit imperfect, productivity tool. While few would pay a premium for it, many were willing to pay a tiny fee or use it free simply to boost efficiency. More importantly, because it runs in the cloud, its delivery cost is incredibly low. Nobody needs to purchase and host their own GPU hardware; everyone shares the same centralized cloud infrastructure. When the marginal delivery cost of a service is negligible, the threshold to make it "worth trying" is naturally extremely low.

However, robots face two core pain points that chatbots have never had to encounter.

First is the "idle-robot tax" (The idle-robot tax). Regardless of whether a robot is working hard or sitting idle, buyers must continuously pay for it, whether paying upfront as a capital expenditure (CapEx) or converting it into an operating expense (OpEx) through leasing or Robot-as-a-Service (RaaS). A robot used for merely one hour a week must generate staggering value within that single hour, because the cost to transfer it and retrain it for other tasks is prohibitively high. As Physical AI equips general-purpose robots with multi-tasking capabilities and human-like mobility, I expect this dynamic to change. But right now, every idle hour of a robot acts as a heavy financial tax.

Second is the "99.9% wall" (The 99.9% wall). There is a vast difference between knowledge work and physical work in how they define a "passing grade." A mediocre, rough draft can still save time because humans can curate the best parts and fix flaws before finalizing it. Copy, slide decks, and even legal contracts are all subject to endless editing and revision.

Physical work is far less forgiving. A poor weld could scrap an entire metal component; a wrong medication dosage could lead to irreversible, catastrophic consequences.

Figure 2: Classification matrix comparing error recoverability and quality tolerance between knowledge work and physical work. Physical work predominantly clusters in high-risk zones characterized by irrecoverable outcomes and narrow tolerance margins.

The nature of knowledge work allows for error correction before AI outputs become final deliverables. In physical work, although humans can also intervene, the cost required to recover from a physical mistake is often far heavier than editing a paragraph. Physical work typically tolerates errors very poorly, which means the threshold for reliability is exceptionally high. For robots performing physical labor to hold real value, they must consistently produce high-quality results, demanding not just a 99% success rate but often exceeding 99.9%. This erects a much higher barrier for large-scale deployment. Aside from occasional successful demo runs, robots must complete tasks at sufficiently low costs and with sufficiently high reliability to genuinely persuade clients to integrate them into actual operational production.

Figure 3: Comparison of success rates required to meet usability thresholds between knowledge work and physical work. Knowledge work typically creates value at rates between 50% and 90%, whereas physical operations often demand absolute reliability ranging from 99.9% to 99.9999%.

Between knowledge work and physical work, the quality thresholds that define a task as "practically useful" differ drastically.

In the years since ChatGPT's launch, the robotics field has achieved massive leapfrog progress, partly fueled by spillovers from external technologies. General AI models and video generation models have jointly accelerated robotics R&D and formed the foundation of robots' core brains.

Many cutting-edge robotics models today utilize a dual-system architecture design, inspired by psychological classifications of human thought processes. System 2 is the slow, deliberate thinker: a specially adapted general AI large model responsible for understanding environmental contexts, simulating plans, and executing logical reasoning. System 1 is the fast, instinctive executor: a motor execution model (derived from video prediction models in some architectures) tasked with translating the aforementioned plans into high-frequency motor control commands in real-time.

This means that every technical breakthrough achieved by OpenAI, Anthropic, Google DeepMind, NVIDIA, Alibaba, and top university labs is indirectly or directly feeding back into the advancement of physical intelligence. In the table below, the core capabilities of cutting-edge robotics models are comprehensively compared against ChatGPT at its initial launch and its current state.

Figure 4: Comprehensive evaluation comparison of core capabilities in learning memory, autonomous execution, and reliability between frontier robotics models (such as Dyna 2.1, Rhoda DVA) and early-stage ChatGPT (GPT-3.5) alongside today's general-purpose models.

In my view, frontier robotics models have actually surpassed the GPT-3.5 model that powered ChatGPT's initial debut across many critical dimensions. Embodied models from companies like Dyna Robotics and Rhoda AI can now handle long-process workflows, conduct multi-step reasoning, maintain longer-context states, and keenly detect their own errors to autonomously retry and correct course.

Meanwhile, we have already witnessed the first wave of Physical AI companies stepping out of laboratories and prototype testing phases to boldly advance toward large-scale deployment in the real world. In fact, Waymo has been walking this path in the autonomous driving sector for years.

In recent years, the internet has seen a flood of stunning videos showcasing general-purpose robots flawlessly completing tasks. But the core issue remains: observers have no way of knowing if these polished short clips were meticulously edited from dozens or even hundreds of failed attempts (and reality is often exactly that). Figure AI's unbroken 200+ hours of live web streaming its humanoid robot sorting packages ranks as a "mini ChatGPT moment" in the robotics field: it is the first time in history that real, indisputable evidence exists proving general mechanical systems can stably execute complex physical labor just like humans.

Path Robotics has now widely deployed industrial welding robots to leading clients in energy, defense, shipbuilding, and other sectors, and earlier this year announced a massive contract worth up to $600 million (potentially reaching $900 million) with military conglomerate HII. Dyna has similarly rolled out automated robots across networks of large restaurant chains, hotels, and commercial laundries, stably running various complex procedures.

Figure 5: Compilation of recent reports from major global business media and institutions regarding the large-scale deployment of Figure AI, Standard Bots, Dyna Robotics, and Path Robotics in real factories and commercial networks.

How is all of this achieved? The answer lies in specialized models and targeted fine-tuning. The models powering Waymo's autonomous car navigation and Path's precision welding cannot handle bin-picking tasks. But these vertical niche markets are large enough to justify training dedicated models for mastering a single skill, ensuring they run robustly under various complex real-world conditions.

Embodied AI companies like Dyna and Rhoda typically start with general-purpose robot foundation models, collect real-world physical data tailored to specific workflows, and continuously fine-tune until the model's reliability on that task meets stringent industrial standards. They prioritize entering markets with the widest touchpoints and most prominent pain points through common procedures, gradually expanding the boundaries of physical data collection as subsequent funding rounds progress and capital accumulates. This rollout strategy mirrors exactly how general AI labs initially focused on conquering the software engineering coding domain.

In the past, achieving commercially deployable reliability for robots on single tasks often took months. However, as underlying foundation models grow increasingly smarter, some leading companies have found this cycle compressing dramatically to mere weeks. Earlier this year, several frontier teams successively released latest research findings indicating they have found conclusive evidence that robotics models likewise follow Scaling Laws. In other words, they assert that by increasing data scale and compute investment, embodied models will achieve qualitative leaps in generalization capabilities, while the fine-tuning data required for each new skill drops significantly. Every reduction in Post-training cost lowers the friction of technology landing, extending the list of tasks general-purpose robots can handle.

So what will the true form of mass adoption for robots look like?

Figure 6: Growth trajectory and forecast of autonomous vehicles on the road from 2021 to 2026. Gradually scaling from an early base of 500,000 units, projected to exceed 16 million by 2026, demonstrating a typical infiltration curve where embodied intelligence gradually accelerates.

You can view autonomous vehicles (considered a key sub-branch of Physical AI) as an excellent reference blueprint. Autonomous algorithms have undergone a qualitative leap in recent years, pushing the global total of autonomous vehicles past the 10 million mark. Meanwhile, autonomous cars face unique challenges many other industrial robots haven't encountered: their demands for reliability are two to three full orders of magnitude stricter, because a single error could cost a life; national safety regulations across countries and regions, as well as municipal operational permits, require painstaking, step-by-step conquest; furthermore, unmanned taxi fleets like Waymo and Cybercab must establish exclusive, heavy-asset ground physical operation networks in every city, covering dispatch centers, charging infrastructure, vehicle cleaning, maintenance, remote safety assistance, and round-the-clock roadside rescue teams.

Different from ChatGPT's pure consumer-side breakout and Waymo's ride-hailing penetration, I anticipate that before robots truly explode in the consumer market, they will first achieve substantively meaningful large-scale commercial deployment in business and industrial sectors. I firmly believe that someday in the future, home service robots will become as commonplace in households as cars are today; but in the foreseeable near term, training robots to achieve absolute reliability in higher-structured environments—such as industrial laundry, machine tool loading/unloading inspection, and warehouse material handling—remains the relatively controllable direction for engineering R&D. The gears of scalable expansion have already started turning, and the technological ceiling of physical intelligence is being reset anew every week.

Thus, the "ChatGPT moment" may have been a flawed analogy from the very beginning. In ChatGPT's world, channel distribution preceded system reliability; in the robotics field, reliability must be established before any scaled distribution, and the engineering pathways to achieve this goal have now been clearly validated. We stand at a historical starting point across industries resembling the autonomous vehicle adoption curve. History may not simply repeat itself, but it often rhymes.

Disclosure: RoboStrategy's portfolio may contain the companies discussed in this article.