The AdValorem AI & Robotics Research Thesis: Why Humanoid Supply Chains, Physical-World AI, and Embodied Intelligence Are the 2026–2030 Window
Why we cover this vertical: AI & Robotics is not a single theme — it is a convergence of three distinct but reinforcing transitions happening simultaneously: the industrialization of humanoid hardware, the maturation of physical-world AI (sometimes called “world models” or embodied intelligence), and the structural shift in how private capital enters early-stage deep-tech before liquidity events arrive. AdValorem tracks this vertical because we believe the 2026–2030 window is the period during which foundational positions are established — not the period when broad recognition peaks. Our research goal is to help our community understand what is actually happening inside the supply chain, the model stack, and the market structure, so they can reason about it more rigorously than the headline narrative allows.
1) The humanoid supply chain build-out: who is scaling, and what it costs
The most important development in robotics in 2025–2026 is not a single robot demo — it is the beginning of a genuine manufacturing scale-up, accompanied by the first serious pressure on bill of materials (BoM) costs. Several companies now have meaningful capital behind them:
Figure AI closed a Series C exceeding $1 billion at a post-money valuation of $39 billion in September 2025 — approximately 15x its Series B valuation of $2.6 billion from just 18 months earlier. Investors include NVIDIA, Brookfield Asset Management, Macquarie Capital, LG Technology Ventures, Salesforce, and Qualcomm Ventures.
Apptronik — a University of Texas spinout building the Apollo humanoid, backed by Google DeepMind and Mercedes-Benz — expanded its Series A to $935 million at a post-money valuation of approximately $5.3 billion as of February 2026, following strong investor demand that caused the round to reopen multiple times.
1X Technologies (backed by OpenAI) has announced a new California factory with stated capacity for up to 10,000 of its Neo household humanoid annually, a meaningful step toward production at scale.
Sanctuary AI has continued iterating on its Phoenix platform, focused on cognitive AI architectures that can be applied across multiple hardware form factors.
Tesla Optimus is the most-watched and most-debated program. Tesla has converted its Fremont production line (formerly Model S/X) to begin Optimus production, with a second factory planned at Giga Texas for 2027. Elon Musk has confirmed initial output will be “quite slow” given Optimus has over 10,000 unique parts — and acknowledged that the program has not yet reached the stage of robots doing sustained, commercially useful work. The production timeline has been revised repeatedly; analyst consensus leans toward meaningful commercial volume arriving 2027–2028.
The BoM story is where the structural investment thesis lives. A 2026 McKinsey analysis of humanoid robot cost structure found that a typical unit runs $30,000–$150,000 to assemble at pilot scale, with actuators (specifically reduction gearboxes) representing the single largest cost line — often approaching half the total mechanical bill. Vision systems ride a smartphone-driven commodity curve; AI compute runs on TSMC and Samsung foundry silicon. The implication is that humanoid cost reductions over the 2026–2030 horizon will be driven primarily by actuator manufacturing scale and supply chain consolidation around a small number of component categories — not by AI software improvements alone. Bank of America projects that a Chinese-manufactured humanoid’s BoM will fall from approximately $35,000 in 2024 toward $17,000 by 2030, while Western-produced units currently sit at $90,000–$100,000 at pilot stages.
This is the picks-and-shovels framing we use in our research: rather than betting which OEM wins the humanoid race, the more durable question is which component and subsystem suppliers capture margin on every unit shipped — regardless of brand.
2) Physical AI and world models: the software stack that makes embodiment viable
Hardware scale alone does not create useful robots. The technical bottleneck — the reason humanoid robotics failed to commercialize meaningfully for decades — was always the software: robots could not generalize across unstructured environments, could not interpret ambiguous instructions, and could not learn from small amounts of demonstration data. That constraint is now being attacked from multiple directions simultaneously:
Google DeepMind’s Gemini Robotics platform, including the Gemini Robotics-ER (Embodied Reasoning) 1.6 model released in April 2026, demonstrates a new capability class: the model reasons about physical common sense — understanding that a glass bottle is fragile while a plastic one is not, that a table bolted to the floor cannot be lifted, that task completion should be verified by watching the delta between frames rather than only the final state. The model is available via the Gemini API and Google AI Studio. The DeepMind RT-X line of research has underpinned much of this generalization work.
NVIDIA’s Isaac GR00T platform is the clearest example of a platform-layer bet on embodied AI. NVIDIA announced GR00T N1.7 in early access at GTC in March 2026 with commercial licensing for production-scale deployments, and has partnered with major robot OEMs including Apptronik, LG Electronics, NEURA Robotics, and others to run GR00T models as the “brain” layer across diverse hardware. The strategy mirrors what NVIDIA did in data center AI: own the compute substrate and the model training infrastructure, then capture margin as the ecosystem scales above it.
World models as a concept — the idea that a robot’s AI should maintain an internal model of the physical world that it can use for planning, not just pattern-matching on sensor inputs — is increasingly the lens through which the leading labs are describing their robotics AI work. The practical expression is robots that can handle novel scenarios without being explicitly programmed for each one, using simulation-to-real transfer pipelines and large-scale pre-training on synthetic and real-world data.
The convergence of large language model reasoning, vision-language-action (VLA) architectures, and high-fidelity physics simulation is moving these capabilities from research benchmarks into early industrial deployment in 2026. The window is not open indefinitely — once a platform layer consolidates (as happened in mobile with iOS/Android, and in cloud with AWS/Azure/GCP), the architecture tends to lock in.
3) The market opportunity: what institutional forecasts actually say
Several major institutional research franchises have published substantial humanoid robotics forecasts in 2025–2026, and the numbers bear examination:
Goldman Sachs revised its humanoid TAM estimate upward more than sixfold in 2024, reaching $38 billion by 2035 (up from an earlier $6 billion projection), with a base case of more than 250,000 humanoid robot shipments in 2030 — nearly all for industrial use. Goldman also projects 100,000 humanoid robots in U.S. deployment by early in the next decade, driven by labor substitution in manufacturing and logistics.
Bank of America published a comprehensive March 2026 research piece projecting the global humanoid robot population could reach 3 billion units by 2060 — surpassing cars on a per-capita basis — with annual shipments growing from approximately 90,000 units in 2028 to 12 million by 2035, implying an 86% CAGR that would outpace the early EV adoption curve. BofA has tracked funding in the sector growing from roughly $0.7 billion in 2018 to $4.3 billion in 2024.
Morgan Stanley Investment Management published its “Embodied AI and the Rise of Humanoid Robots” research note in January 2026, identifying the intersection of vision-language models, labor economics (aging workforces, persistent labor shortages, rising wages), and simulation advances as the structural driver behind early industrial deployments — while noting that broad adoption remains years away.
McKinsey projects general-purpose robotics reaching a $370 billion market by 2040, with humanoid supply chain opportunities representing a distinct “billion-dollar win” category for component suppliers who can solve the actuator cost problem at scale.
The IFR World Robotics 2025 Report documented 542,000 industrial robot installations in 2024 — the fourth consecutive year above 500,000 units — with global installations forecast to reach 700,000 annually by 2028. Humanoid robots are tracked separately in IFR’s position paper series as an emerging sub-category of service robotics.
These projections vary substantially in methodology and time horizon. Our role in the research is not to adjudicate between them but to help our community understand the assumptions that drive each model and where execution risk actually concentrates — which, in most cases, is at the supply chain and deployment layer, not the demand layer.
4) Why pre-IPO and secondary markets are the relevant access vehicle
The companies doing the most consequential work in this vertical — Figure AI, Apptronik, 1X, Sanctuary AI, and their direct peers — are private. They will likely remain private through the core build-out phase of 2026–2029. This creates a structural feature of the market: the period during which foundational technology choices are made, unit economics are established, and OEM partnerships are locked in is largely invisible to public market participants.
The primary vehicles for non-institutional participants to access this exposure are pre-IPO secondary markets (where existing shareholders in private companies transfer shares to new buyers) and structured research programs that track the ecosystem. Secondary platforms like Forge Global report active price discovery on Figure AI shares — with a Forge Price of $174.00 as of late May 2026 implying a valuation of approximately $34 billion — creating a secondary market reference point that did not exist even two years ago. Sanctuary AI trades on secondary platforms including EquityZen.
The analytical challenge in secondaries is not finding the shares — it is understanding what you are actually buying. Pre-IPO secondary positions in private deep-tech companies carry specific structural considerations: information asymmetry relative to insiders, transfer restrictions, no guaranteed liquidity event, and sensitivity to how corporate actions (new primary rounds, preferred stock terms, company-side rights of first refusal) affect the economics of secondary positions. This is precisely why AdValorem’s research covers the structural mechanics of these instruments alongside the thematic market analysis — both sides of the picture are required for rigorous reasoning.
5) What we publish on this thesis
AdValorem’s AI & Robotics vertical is one of our most active research tracks. We publish weekly research notes covering specific companies, supply chain developments, model releases, and funding events. Our podcast covers both the technical layer (how embodied AI actually works) and the market structure layer (how capital flows into and through this ecosystem). Our Discord community is where members share deal flow intelligence, exchange views on specific companies, and ask questions of the research team in real time.
Recent research outputs have included our deep-dive on the Figure AI and Apptronik funding stories as a lens into humanoid capital concentration dynamics, actuator supply chain analysis framed around the McKinsey BoM data, and coverage of NVIDIA’s GR00T N1.7 release and what the commercial licensing model implies for the platform layer. We treat this vertical the same way a sector-specialist research team would: with consistent coverage cadence, primary source analysis, and explicit acknowledgment of where our confidence is high versus where it is speculative.
Bottom line
The 2026–2030 window in AI & Robotics is defined by the intersection of three forces arriving simultaneously: manufacturing scale-up (the supply chain is getting built now), physical AI maturation (world models and embodied reasoning are crossing commercial viability thresholds), and capital market structure (the pre-IPO secondary ecosystem is developing just as these companies approach their most consequential growth phases). None of these forces is guaranteed to produce a specific outcome — the execution risks are real, the timelines are uncertain, and the competitive landscape includes both Western and Chinese participants at scale. But the combination creates the kind of compressed research opportunity that AdValorem is built to cover: a window in which deep understanding of the technical and structural details matters more than general familiarity with the theme.
If this thesis resonates with your portfolio focus, we’d love to discuss the research further.
Get Weekly Research
Analysis, education, and market intelligence — delivered to your inbox.
Subscribe
Join 586+ members for weekly research. Unsubscribe anytime.
Sources
Bank of America Institute — “Physical AI, Part 2: Humanoid Robots” (March 2026)
Morgan Stanley Investment Management — “Embodied AI and the Rise of Humanoid Robots” (January 2026)
Google DeepMind — “Gemini Robotics ER 1.6: Enhanced Embodied Reasoning” (April 2026)
IFR International Federation of Robotics — World Robotics 2025 Report (September 2025)
If this thesis resonates with your portfolio focus, we’d love to discuss the research further.
Schedule time with the AdValorem research team to go deeper on AI & Robotics coverage.