Is Robotics The Next Megabubble?

Summary

• Andrew Kang’s core call: robotics today is crypto in 2014–2015 — total private robotics market cap was ~$20B when he entered in late 2023/early 2024 and is now $100–150B “maybe pushing $200 billion,” versus XRP ($100B) and Pokémon trading cards ($50B). Against a destination he puts at tens of trillions, “the industry as a whole still has 100x, potentially 1,000x upside remaining,” and the edge comes precisely because “there’s no framework for evaluating it” yet.
• The timeline call: “we’re pretty much at the GPT-3 level of robotics” — and the jump to the GPT-5 of robotics takes “something like a year,” not the two years LLMs needed, because the research learnings (RLHF, mid-training, data mix, annotation infra) transfer directly. “Robotics AGI is going to happen a lot sooner than people think.”
• His new vehicle, Robo Strategy, is the MicroStrategy playbook applied to private robotics — a NASDAQ-listed closed-end fund rolling up his private book, keeping accretive share issuance as a “second engine” for NAV per share while largely leaving out leverage (no preferred shares). The demand driver is access: public markets offer basically Tesla and little else while SpaceX/Anthropic/OpenAI stay private past $1T, and the access problem for private venture “is multiplied a thousand X” versus Bitcoin.
• The ambition is explicitly SoftBank scale: MicroStrategy peaked ~$100B; a $10T robotics market is ~7x Bitcoin’s current cap, so Kang is “reaching for something… in the hundreds of billions” — tens of billions deployed into “the Anthropic of robotics,” targeting 100% year-over-year on billion-dollar entries versus the traditional 20–25% growth-stage bar.
• The capital rotation is already on: robotics VC interest “started to inflect the last 3–4 months” — the reason he gave was software’s repricing, which gave VCs “an existential moment” about SaaS durability, and physical AI became the apparent next stage. Term sheets are printing at multiples of valuations from just months ago.
• The moat answer for robotics companies: “You can’t just hook up Claude into a robot and have it work” — LLMs know the world via text, while robot foundation models must learn force, physics, and the priors/affordances humans get free (one hand or two for a bottle, don’t tilt the water glass).
• Risks, as hedged by Kang himself: the NAV premium cuts both ways — compression from 4x to 2x “is a 50% loss in market price” — and venture robotics will carry a “very, very high failure rate” even with big winners. The host’s sharpest point stands too: a public vehicle at 2x NAV on illiquid privates may itself be pulling forward the underlying marks — Kang conceded listed versions “could trade lower.”

Deep dive

1. From DeFi to humanoids: buying the inflection, not the framework

• Kang ran Mechanism Capital from ~2018 — early DeFi, no outside capital ever — and stayed inside crypto until late 2023/early 2024, when a friend showed him Figure AI and founder Brett’s humanoid pitch. His unlock: humanoids were a 100-year sci-fi dream blocked by one thing — “the barrier was always intelligence,” not hardware — and ChatGPT in 2022 made digital AGI extrapolable to physical AGI.
• Why jump then: “The best time to invest, in my opinion, is at these inflection points” — the market doesn’t reprice a change in the pace of development. “It kind of reminded me of crypto in 2014, 2015”: complex, interdisciplinary, no established valuation framework — so whoever builds one first has “tremendous edge over all the other investors in the industry.”
• His method since: “pure obsession… no days off” except surfing, and — only half-joking — “I just talk to Claude. That’s my primary point of contact with another being.” Papers stopped being daunting once he could interrogate every ablation and architecture trade-off conversationally; what obsession can’t replace is decades, hence Scott Walter (40 years in robotics, two companies sold, the last to Kuka) on the team.
• The epistemics: a “very truth-seeking” culture, high conviction only after diligence — and if you validate five separate ways you could be wrong and you’re not, “maybe the market’s wrong. That’s what happened with Figure AI.”

2. Robotics market size and upside

• Private robotics market cap when he entered: ~$20B. Today: ~$100–150B, “maybe pushing $200 billion” — versus XRP near $100B and Pokémon trading cards around $50B. Against tens of trillions, “100x, potentially 1,000x upside remaining.”
• Why robotics over space: SpaceX was already proven and priced, and “most people are not going to ride a rocket ship… robots are going to permeate everyday life” much sooner.
• The 2035 picture: a $20k economy / $50k premium humanoid versus human personal assistants at $40–50k a year (executive assistants up to $200–400k all-in), robots caring for your parents, working factories, hotels, restaurants — and robots in space, in environments humans can’t operate.

3. The under-hyped giant: intelligent arms, not just humanoids

• Industrial arms have always existed — what’s new is intelligence that removes “weeks of programming and tens of thousands, maybe hundreds of thousands of dollars of engineering cost just to set things up.” Kang’s symmetry: “if there’s going to be billions of humanoids, then there’s probably also going to be billions of mechanical arms” — coffee, cooking, warehouse unpacking.
• Standard Bots, a key portfolio holding, builds cobots and industrial arms and is “the only one that really exists at scale” in America — vertically integrated, not reliant on China in the supply chain. If the US re-industrializes, “they’re going to power the reindustrialization of America”; Kang frames the company as strategically important to the American government itself.

4. We’re at GPT-3 robotics — and GPT-5 arrives in a year, not two

• The calibration: “we’re pretty much at the GPT-3 level of robotics” — it just doesn’t look impressive, because a robot that picks up a cup only half the time reads as failure rather than progress. What frontier models now show: generalization to homes and factories never seen in training, in-context learning migrating into physical AI, pick-and-place expanding into long-horizon tasks (“package it, then take that box and put it over there”), and internal memory — Rohde AI (as heard) demoed a robot winning the three-cup shell game. These capabilities are “emergent” as data and compute scale.
• The compression argument: GPT-3 to GPT-5 took two years of research learnings — RLHF, mid-training structure, pre-training data mix, annotation infrastructure — and those don’t need re-discovering for physical AI. “I think it’s going to be something like a year… robotics AGI is going to happen a lot sooner than people think.”

5. No, you can’t just plug Claude into a robot

• The host’s moat question — if the frontier labs hit AGI, can’t you hook their models to any old hardware? Kang: “You can’t just hook up Claude into a robot and have it work.” LLMs like Opus understand the world via text; his analogy, as told: imagine a blind person with no limbs suddenly given sight and limbs and told to interact with the world — “maybe they could figure it out after a lot of trial and error, but it’s not going to work out of the box.”
• The missing layer is what researchers call priors or affordances: how much pressure keeps a cup held without dropping it, one hand or two for a bottle of a given size, don’t tilt the glass or the water falls out. Humans get these free; robot models must learn them — and learn to learn, since robots will face situations “we haven’t even thought of.”

6. Software’s existential moment is funding the rotation

• Interest “started to inflect the last 3–4 months,” and the reason he gave was software’s repricing: VCs realized some software businesses “were not as durable as we thought they were,” had “an existential moment” about where to deploy if not SaaS, and physical AI became the apparent next stage. Previous concerns included CAPEX, China competition, and US manufacturing; the trend wasn’t established yet.
• The result: basically every top-tier and Tier 2 VC now looking, some term sheets at valuations multiple times higher than these companies were raising at just a few months ago — and Kang thinks today’s AI-company interest “is the equivalent of what it will be in the next 1 or 2 years” for robotics.

7. Robo Strategy: the MicroStrategy playbook, limited leverage

• The origin: after posting robotics deals on X, Kang got “over a thousand DMs” asking for exposure — and public markets offer essentially Tesla ($2T market cap) and one or two others, while “SpaceX, Anthropic, OpenAI are going to a trillion dollars plus before they’re going public.” His candid framing of motive: “my primary goal is winning. It’s not like I’m an altruistic person” — access for everyone is the “really nice factor,” but it’s also the product.
• The mechanics, as he laid them out: MicroStrategy was really the private-equity roll-up model (TransDigm, Constellation Software) applied to Bitcoin — public comps trade at 15–40x where small privates trade 3–8x, so “I buy this private company, maybe I spend 100 million. Once it’s on my books, it re-rates immediately to 400 million.” And MSTR’s real demand driver was access: the premium survived the ETF because the restricted capital base — RIAs, endowments, pensions — “turns out is really, really large.”
• His claim: that access problem “is multiplied a thousand X for private robotics investments,” and the TAM is bigger — Bitcoin sits below $1.5T while robotics goes an order of magnitude past it. What he keeps from MicroStrategy: accretive issuance at a premium as a “second engine” to grow NAV per share even in a flat-mark bear market. What it limits: leverage — no preferred shares, and any borrowing “very, very selective” and short-lived.
• The hedge, stated plainly: the premium adds real risk — compression from 4x to 2x “is a 50% loss in market price” — though NAV (marked to last-round valuations) also misses imminent rounds and growth that public markets price real-time.

8. SoftBank scale — and the host’s sharpest pushback

• The target: MicroStrategy peaked around $100B; “I’m reaching for something that’s more like SoftBank scale… in the hundreds of billions.” His napkin math: $10T robotics is ~7x Bitcoin’s market cap, so 7x MSTR’s peak before even counting the worse access. Asked whether that requires expanding beyond robotics: a flat “No.”
• The power-law logic underneath: without scale a robot costs “$500,000 to a million” to produce; with scale, 90% cheaper — so scale compounds and “there’s going to be a multi-trillion-dollar winner.” He wants “tens of billions of dollars to work in the Anthropic of robotics,” targeting 100% year-over-year on billion-dollar entries versus the 20–25% traditional growth-stage bar — while conceding venture robotics will have “a very, very high failure rate” alongside the big winners.
• The host’s pushback — worth keeping: a public vehicle trading at 2x NAV on assets that don’t trade 24/7 may itself be setting the price of the underlying, pulling forward the next round’s markup. Kang conceded the point: if these companies went public, “they could trade lower, they could trade the same, they could trade higher.”

9. The book, rapid fire

• Figure is “the Apple of this generation” — taste and design nobody else matches, certainly not the “pretty industrial, cookie-cutter” Chinese entrants; a home robot will carry iPhone-style status. Apptronik is “the next best, maybe a year or two behind” — an early humanoid innovator that built prototypes for today’s biggest names, with manufacturing depth (“Stable,” as heard — one of the few large-scale US hardware manufacturers) plus a Google DeepMind partnership; per Elon, manufacturing was harder than design at Tesla.
• Dyna Robotics: an elite robot-learning group, strongest in post-training, among the first to hit 99.9% task success — the bar that matters, because “you can’t have a human intervene every hour… the goal is that humans don’t have to intervene ever.” They’re building their own hardware platform too.
• Dexmo: the “Unitree of America” — first US company to commercially sell humanoids (a wheeled version last year to “basically all of the major US research labs,” legs “later this summer”). Kang’s platform thesis: robots become the next iPhone, with a developer skill economy on top — “somebody create a skill to have my robot cook as well as Gordon Ramsay.”