Why Robotics Is The Next Trillion Dollar Industry - Andrew Kang
Summary
Andrew Kang’s core bet is that intelligence—not mechanical design—was the missing layer that kept general-purpose robots in science fiction, and ChatGPT made the coming physical-intelligence inflection visible. After discovering Figure AI in late 2023/2024, he pressure-tested the thesis against skeptical venture investors, competitors, references, and co-investors, eventually concluding that “the rest of the venture capital landscape was missing something.” The opportunity is enormous, but concentrated: as with AI, many robotics startups may fail while the top companies compound 30x or more.
Kang argues that conventional valuation models break when technological progress turns reflexive and growth reaches 100%, 200%, or 300% rather than a steady 8% or 20%. Automating research could increase experimentation 2x, 5x, or 10x and cut 90-day iteration cycles to 45 days, with each improvement accelerating the next. His portfolio implication is blunt: acknowledge risk management and possible 30%-50% drawdowns, but stop allowing short-term macro anxiety to eject you from exceptional companies whose opportunity may be almost infinite.
Robotics is already at what Kang calls the “ChatGPT-3 level,” but real deployment requires moving from being right roughly half the time to near-zero failures. He thinks “GPT-5-level” physical intelligence could arrive in roughly a year because robotics can reuse advances in pre-training, annotation, RLHF, mid-training, and automated experimentation rather than rediscovering them. Home-capable robots might follow in about two years, broader everyday visibility in three to four years, and—if production grows roughly 10x annually—hundreds of millions or billions of units in the early 2030s.
The economic prize spans both premium consumer hardware and a wholesale restructuring of labor-intensive industries. Kang expects at least one robot per household, potentially one per person, with brand, taste, and status supporting Apple- or Tesla-like margins; meanwhile, Walmart’s roughly $700 billion of revenue, $20 billion of net income, $30 billion of operating income, and approximately $100 billion of physical-labor expense illustrate how automation could move margins from 3% toward 10%-15%. He does not celebrate displacement: formerly “anti-socialism,” he now sees a credible case for UBI and a safety net if automation breaks the traditional labor system.
Humanoid robotics should support multiple major winners rather than collapse into one dominant supplier. Kang says large technology companies were “asleep at the wheel”: Google sold Boston Dynamics and scrapped Everyday Robots, while Google and Nvidia now emphasize intelligence and operating systems rather than wholly owned humanoid hardware. Tesla Optimus should become substantial, but Tesla is not clearly number one today; Kang instead expects an automobile-like market with perhaps 15-20 large players and greater value concentration among the leaders.
The most durable near-term moats may sit in hardware, manufacturing, deployment, and supply chains—not standalone physical-AI models. Kang hedges that he “may be wrong,” but points to Chinese open-source labs, Nvidia’s research, DreamZero reaching the frontier when released, and a shrinking open-versus-closed model gap as evidence that intelligence will commoditize once it is “good enough.” Actuators already represent roughly 30%-50% of a robot’s bill of materials, onboard GPUs are mandatory, and deployed fleets produce proprietary real-world data that hardware companies can use to improve their models.
Robo Strategy is designed to turn scarce private-market access into a permanent-capital compounding vehicle rather than a leveraged robotics proxy. Kang estimates new-generation private robotics companies are collectively worth only $100-$200 billion—versus roughly $50 billion for all Pokémon cards—and believes the industry could grow 100x or 1,000x. Borrowing MicroStrategy’s “accretive issuance” loop while avoiding unnecessary leverage, he wants to build a SoftBank-scale public venture platform whose access advantage is multiplied 100 times because investors cannot simply buy leading private robotics companies through an ETF or exchange.
Deep dive
1. Figure AI turned a science-fiction category into an investable inflection
Kang first became interested in robotics in late 2023 or 2024, when a Silicon Valley friend showed him Figure AI. Humanoids initially seemed “pretty out there,” but researching the company and its founder convinced him both that the team could execute and that robots able to learn, perceive, and act like humans were finally becoming feasible.
His technical framing was decisive: mechanical design is extremely difficult but ultimately solvable; the deeper bottleneck was intelligence. Once ChatGPT demonstrated a visible trajectory toward much stronger machine intelligence, Kang believed the same progress would extend into physical intelligence and unlock something as transformative as “the smartphone or even the internet or AI.”
Venture friends repeatedly advised him not to invest because robotics had not historically produced venture-scale outcomes and remained expensive to develop. Easy-looking SPV access initially made Kang suspect he was “getting ripped off,” so he studied competitors, industry dynamics, round participants, and team references before deciding that he was not missing something—the venture landscape was.
The category’s multidisciplinary difficulty became part of the attraction. Robotics demands expertise across AI, robot learning, mechanical design, hardware engineering, and manufacturing; “when there’s a huge gap in understanding,” Kang argued, “there’s going to be a lot of future alpha.”
2. Exponential development rewards holding more than macro timing
Kang’s “Exponential Horizon” framework begins with the claim that “technology development is not linear.” If AI automates AI research, experimentation could accelerate 2x, 5x, or 10x, operate for more hours, and compress a 90-day iteration cycle to 45 days; better AI then accelerates the next cycle again.
Traditional DCF assumptions—perhaps 8% or 20% annual growth along a stable line—struggle with companies expanding 100%, 200%, or 300%. Even assuming 100% growth for several years can produce an output “10x different,” helping explain why certain technology valuations appear parabolic before conventional models catch up.
His provocative response to macro anxiety was “Who cares?”—immediately qualified by acknowledging that risk management matters. A public stock may fall 30%-50% or go unrewarded for a year, but if the underlying company is exceptional and the setback temporary, its growth can rapidly overwhelm the correction.
Kang cited Peter Thiel and Chamath Palihapitiya selling their Facebook positions at the IPO: despite many later investments, they might have outperformed by simply holding. He includes himself in the critique—his largest wins came from holding until the thesis played out, while longer-term positions generally beat his shorter-term trading.
3. Talent migration and live demonstrations are changing what people believe
Kang sees elite talent moving from traditional finance, crypto, and other formerly fashionable fields into “the world of atoms.” His sharpest example was a founder in his twenties running a nine-figure proprietary fund, mostly with his own capital, who applied to intern at a physical-AI lab despite potentially being able to retire.
Better supply chains, development kits, tooling, education, and assistants such as Claude are lowering the cost of entering robotics. Newcomers still need time to acquire domain expertise, but Kang expects the growing talent pool and supporting infrastructure to accelerate the entire field.
Thread Guy compared robotics skepticism with constantly shifting AI goalposts: driverless Waymos already look ordinary in Los Angeles, while software engineers moved from denying AI could code to reporting that it generates 90% of their work. His equivalent robotics moment was Figure’s 20-plus-day livestream, whose consistency was more persuasive than a polished demonstration.
Kang called the livestream “one of those moments where there’s like a shift in the global consciousness.” It did not imply a million home robots tomorrow, but critics who once doubted Figure was real began asking how to invest; a live feed made it harder to allege editing or tricks.
4. Reliability—not raw capability—is the next intelligence threshold
“We’re at the ChatGPT-3 level of robotics already,” Kang argued: foundation models can perform a meaningful range of tasks, but may still fail roughly half the time. That can demonstrate capability, yet usefulness demands a much higher bar—“zero failures,” not a coffee machine that breaks or produces the wrong drink every other day.
He maps reliable workflow integration to a “GPT-5 level” of intelligence, where selected tasks no longer require a human checking every result. Although the move from 3 to 5 took roughly two years in language models, he thinks robotics might traverse its analogous gap in about a year because technology now accelerates future technology.
Physical research remains painfully labor-intensive. A robot may place a bottle in a basket 100 or 1,000 times, but a person must reset the scene, observe each trial, record whether it succeeded, and diagnose what went wrong; removing that human loop sharply increases experimental throughput.
Robotics can also inherit techniques already developed for language models: improved pre-training mixes, data annotation, RLHF, mid-training, and better ways of eliciting reasoning. Open research and talent movement let physical-AI teams reuse that work rather than repeat the whole experimental journey.
5. Intelligence may arrive before factories can satisfy demand
Kang expects hardware, not intelligence, to become the binding constraint. He would not promise a home-ready robot in six to twelve months, but suggested “maybe we’ll have that in two years”; once capability crosses that threshold, demand could jump far faster than factories can produce physical units.
More visible everyday deployment could follow in three to four years, with buyers willing to pay more gaining earlier access. Kang floated roughly 10x annual unit growth, enough to reach hundreds of millions or billions of robots relatively quickly—potentially in the early 2030s.
His optimistic household robot is proactive rather than merely commanded: it shops, handles chores, collects children from school, supports aging parents, and asks, “How can I make Thread Guy’s life better?” Robots could also construct businesses or buildings, then prepare infrastructure on Mars or other inhospitable environments before humans arrive.
Kang expects “at least a robot per household,” and perhaps one per person if abundance makes the economics work. Buyers will not necessarily choose the cheapest machine; like phones or cars, robots displayed inside the home may reward companies with brand, taste, and style.
6. Automation attacks labor costs—and forces a political reckoning
Kang used Walmart to show why cost reduction may matter more than top-line expansion: approximately $700 billion of annual revenue produces around $20 billion of net income and $30 billion of operating income, while physical labor costs roughly $100 billion. Moving margins from 3% toward 10%-15% could multiply enterprise value.
UPS and FedEx similarly generate close to $100 billion of revenue yet trade below that level because costs are high and margins thin. Physical labor appears throughout the economy—and is itself slowing AI infrastructure, as data-center construction lacks enough plumbers, electricians, and other skilled workers.
Kang explicitly rejected reading this thesis as advocacy for displacement: “You can’t stop it,” so policy must address the consequences realistically. Having previously been anti-socialism, he now thinks UBI and a safety net may become necessary when automation leaves “not much of an alternative” and breaks the traditional system.
Thread Guy raised the possibility of a modern Luddite reaction similar to protests against data centers. Kang’s answer was that households will actively want robots once they eliminate dishes, laundry, trash, and cleaning: the time returned to people should make adoption broadly appealing.
7. Big Tech’s delay leaves room for an automobile-like field of winners
Kang’s early excitement around Figure and Apptronik partly came from believing that “big tech companies were asleep at the wheel.” Google once owned Boston Dynamics and later scrapped its Everyday Robots initiative.
Google and Nvidia are now emphasizing robot intelligence and operating systems rather than commercializing complete humanoid hardware. Kang imagines Google partnering with manufacturers such as Apptronik in a Pixel-like arrangement: another company builds much of the machine, while Google supplies software and possibly the brand.
Microsoft, Meta, Apple, and others are becoming more interested, but Kang argues that building excellent robots is not naturally in their DNA; Apple’s abandoned car effort is his cautionary example. Partnerships, distribution agreements, or acquisitions may therefore be more realistic than internally creating everything.
Tesla is the major exception with its own humanoid program. Kang expects Optimus to become “a big success,” but does not regard it as clearly number one today, partly because Elon Musk’s focus spans SpaceX, Neuralink, xAI, and Tesla; he expects perhaps 15-20 large robotics players rather than winner-take-all.
8. Actuators, compute, and general-purpose design define the hardware stack
Actuators—the motors powering robot joints—represent roughly 30%-50% of a humanoid’s bill of materials. Existing markets never required billions of actuators in humanoid-specific shapes and sizes, so China is investing heavily in capacity and making enough progress that even companies such as Tesla reportedly source components there.
Robots also require onboard GPUs: a machine cannot lose Wi-Fi or 5G and simply stop functioning. The existing compute shortage could therefore intensify as intelligence migrates from data centers into physical machines.
Thread Guy’s challenge—why retain a human form if the world could be redesigned for robots?—drew a first-principles rebuttal. General-purpose mobility needs legs for varied terrain and stairs, arms and hands for manipulation, and elevated eyes for visibility; cameras on robot wrists can add a nonhuman advantage without abandoning the basic form.
Kang compared humanoids with GPUs and iPhones: general-purpose products gain enormous economies of scale and replace many specialized devices. A common robot platform could reduce something theoretically costing $1 million to $10,000, while outside developers train skills and personalities instead of designing new hardware from scratch.
9. Model intelligence may commoditize before hardware does
Proprietary data can still create value where expertise is scarce: people may pay for robotic skills trained from a Gordon Ramsay or Roger Federer rather than the third-best practitioner. Yet most customer service, data entry, basic research, factory, or restaurant tasks need “good enough,” not the equivalent of Einstein.
Venture firms recently began pursuing physical-AI model companies as potential analogues to OpenAI and Anthropic. Kang’s qualified dissent—“Maybe I’m wrong, by the way. I should hedge myself here”—is that Chinese open-source labs and Nvidia are shrinking the moat faster than investors expect.
He said the language-model gap between open source and the frontier narrowed from roughly two years to six months, while physical AI’s gap may be smaller still; Nvidia’s DreamZero was “at the frontier” when released. Once a model performs a job reliably, incremental intelligence may carry little economic value.
Hardware retains harder constraints: supplier relationships, actuator quality, five-year durability, shortage allocation, high-rate manufacturing talent, and deployed fleets that generate real-world data. Model-first companies are already attempting to build hardware as they recognize where those moats sit.
10. Simulation could become robotics’ Hyperbolic Time Chamber
Kang likened mature simulation to Dragon Ball Z’s “Hyperbolic Time Chamber”: a million virtual robots could train simultaneously, turning vast amounts of learning into a tiny interval of real time. Today, simulation is most dependable for locomotion because foot-ground contact and internal forces are comparatively tractable.
Manipulating the broader world is harder because water and other deformable objects bend or change shape. If their physics are modeled imperfectly, the robot learns the wrong behavior, preserving the need for expensive real-world training.
Kang expects that problem to be solved, though he moved his own estimate from perhaps five years to maybe three because “the world’s moving so fast.” At that point, physical deployments may lose some of their data advantage as simulation supplies enough realistic experience.
The endpoint extends beyond robots: atomic-level simulation could model the human body and transform drug development. Work that currently takes seven to ten years might, in his optimistic example, take a month or less.
11. Robo Strategy applies permanent-capital compounding to private robotics
Kang is not trying to write a robotics narrative. His method is to notice a new fact pattern, reconsider prior beliefs, and develop “an allergic reaction to something that doesn’t make sense,” then keep researching until competitors, technologies, and failure scenarios produce a defensible risk-reward estimate.
That process requires conviction without dogma, specialists with low ego, and disciplined attention across thousands of variables. For humanoids, a multi-trillion- or “deca-trillion-dollar” potential market can justify paying several billion dollars for an early leader—but only if leadership, team, technology, and manufacturing materially reduce the otherwise high startup risk.
Robo Strategy avoids MicroStrategy-style preferred-share leverage because Kang believes robotics already offers sufficient upside. He estimates the collective market capitalization of new-generation private robotics companies at only $100-$200 billion, versus roughly $50 billion for all Pokémon cards, leaving room in his view for 100x or even 1,000x industry growth.
What he does borrow is “accretive issuance”: sell equity when public markets value the portfolio above its acquisition cost, then buy more private assets below that implied valuation. He credits that loop—not simply leveraged Bitcoin exposure—for MicroStrategy’s NAV per share rising from roughly $2-$4 in 2020 to around $100-$120 despite an average Bitcoin purchase price near $75,000 and possibly more than $3 billion in net losses on Bitcoin.
Kang compared the mechanism with TransDigm and Constellation Software buying private companies at three-to-eight-times earnings while their own cash flows trade at 15-40 times. A $100 million acquisition can therefore add roughly $300 million of public enterprise value through rerating.
He expects leading technology companies to remain private longer, potentially beyond trillion-dollar valuations, because post-2008 reporting burdens and control risks discourage founders from listing while venture, sovereign-wealth, growth, and crossover funds can still supply capital. His proposed answer is a wave of public venture funds—and a SoftBank-scale Robo Strategy offering permanent capital, selection, and access unavailable through an ETF or exchange.