Robotics: The Vertical-vs-Horizontal Split Is the Trade

Robotics has finally split into two coherent bets that disagree about what comes first.

The horizontal bet, embodied by SoftBank, Figure, 1X, Apptronik, and Tesla, says: build a general-purpose humanoid platform, plug a foundation model into it, and let one form factor swallow most of the world's physical labor. The vertical bet, embodied by Forest3, Carbon Robotics, Built Robotics, Gecko, Monarch, CMR Surgical, and the long tail of task-specific robots, says: pick one job, beat the human at it, and ship revenue today. Capital is flowing into both sides -- but the path to ROI looks very different.

This is the next 36 months of robotics. Both bets are correct in different timeframes. The investable question is when they converge, not whether.

The SoftBank bet, decoded

The clearest version of the horizontal thesis is SoftBank's, and 2025 was the year they put their balance sheet behind it. Three moves matter:

1. Acquisition of ABB's robotics business for $5.375B (announced Oct 8, 2025) -- a profitable industrial-robotics business with a global installed base. That gives them the actuator IP, the manufacturing scale, the safety certifications, and the enterprise customer list.
2. Partnership with Yaskawa Electric on "Physical AI" via AI-RAN (Dec 2025) -- telco-network-native robotics infrastructure, with Yaskawa providing the arms.
3. Masayoshi Son's "ASI Platform Provider" framing in the 2025 Group Report -- humanoid and physical AI explicitly positioned as the embodiment layer of a broader artificial superintelligence strategy, alongside Arm and the Vision Fund book.

The thesis: foundation models for embodiment are coming, the form factor that scales them is humanoid, and the moat is capex (factory floors), data (real-world deployment hours), and distribution (industrial customers via ABB, telco via Yaskawa/AI-RAN). Whoever owns the most actuators in the field by 2030 owns the data flywheel.

Figure, Apptronik, 1X, and Tesla Optimus are running variants of the same playbook with different distribution wedges:

• Figure -- $1B+ Series C at a $39B post-money valuation (Sept 2025), Parkway Venture Capital leading. BMW pilot, Helix in-house foundation model. Industrial wedge.
• Apptronik -- $350M Series A (Feb 2025, B Capital + Capital Factory), then $935M total at a $5B+ valuation (Feb 2026). Google DeepMind partnership, UT Austin spinout. Industrial wedge with logistics deployments.
• 1X (Norway, OpenAI-backed) -- raising $1B at a $10B valuation (Sept 2025) to ship NEO into homes. Consumer wedge.
• Unitree (China) -- Series C, ~$1.7B valuation (June 2025), quadrupeds and humanoids at aggressive price points. Volume wedge.
• Tesla Optimus -- in-house, vertically integrated, captive Tesla factories as customer #1.

The horizontal thesis is well-funded, intellectually coherent, and aimed at a real prize. The risk: the gap between "impressive demo" and "cheaper than a human worker doing one specific job" is still very wide, and most of the unit-economics math assumes that gap closes in 24-36 months. It might. It might not.

Why vertical wins the next 36 months

The vertical bet doesn't need foundation-model AGI. It needs a robot that does one job at lower cost than a human, today, with the data and physical constraints of that single domain. Three reasons it's running ahead on ROI:

1. Narrower data requirements. A LaserWeeder doesn't need to fold laundry. It needs to identify 200 weed species under varied light conditions. That dataset is collectable in months, not decades.
2. Customer pull is real. Farms, construction sites, hospitals, ports, and factories have labor shortages and budgets right now. They are not waiting for general-purpose robots; they are buying the thing that solves their specific bottleneck this quarter.
3. Unit economics work at lower volume. A $1.5M agricultural laser-weeder doesn't need 100,000 units sold to be profitable. It needs hundreds.

The vertical bet is winning the revenue race. Whether it wins the long-term value race depends on what happens when a horizontal platform crosses the cost-per-task curve.

Vertical map by use case

Forestry. Forest3 -- small semi-autonomous equipment for fire prevention and precision harvesting. The category is also producing academic prototypes (gibbon-inspired and primate/inchworm tree-climbers from Taiwan and elsewhere) that point at a serious commercial wedge: post-fire inventory, selective thinning, and biomass monitoring at a unit cost no human crew can match. Climate insurance, utility right-of-way maintenance, and carbon credit verification are the buyers.

Agriculture. Carbon Robotics -- LaserWeeder G2 (Physical AI for ag), $70M Series D (Oct 2024), additional $20M Oct 2025 for a new "AI robot" beyond weeding, plus Carbon ATK tractor autonomy. Monarch Tractor -- $133M Series C (Jul 2024), fully electric driver-optional MK-V, plus MonarchOne platform (Aug 2025) extending autonomy beyond ag. Bear Flag Robotics (now John Deere). Burro (autonomous payload carriers, specialty crops).

Construction. Built Robotics -- RPD 35 piling robot for utility-scale solar (224-pile capacity, 34,000 lb payload), deployed in US and Australia. Canvas (robotic drywall). Dusty Robotics (layout printing). Toggle (rebar). These all share the pattern: take one extremely repetitive trade, fully automate it, sell to general contractors as a productivity SKU.

Inspection. Gecko Robotics -- Cantilever AI + robot fleet, US Navy customer, L3Harris partnership (Apr 2025) for extended-reality 3D maintenance models, StratoSight platform launched 2025. Skydio (autonomous drones for utility / infrastructure inspection). Saildrone (autonomous maritime surveillance). ANYbotics (legged inspection for energy / mining). The thesis: critical infrastructure has a labor shortage and a generational replacement cycle hitting at once.

Warehouse / Logistics. Symbotic ($60B+ public, Walmart-anchored). Locus Robotics. Pickle Robot (truck unload). Dexterity (multi-arm logistics). Covariant (now Amazon, foundation-model approach narrowed to vertical). This is the most mature vertical category and the one closest to "solved" for grasp + move tasks.

Surgical. Intuitive is the incumbent. CMR Surgical raised $200M+ (April 2025, Trinity Capital + existing) to push Versius into the US. Distalmotion closed a $150M Series G (Nov 2025) for DEXTER US commercialization. Moon Surgical (collaborative laparoscopy). Vicarious Surgical (single-port platform). Surgical robotics has long sales cycles, but unit economics per-system are extraordinary when adoption clicks.

Kitchen. Chef Robotics (Series A 2024, autonomous food assembly). Hyphen (digital makelines for restaurants). Mukunda Foods (vending-grade). Vertical SaaS for QSR with a hardware attach.

Elder care. Ties back to last week's Big Button conversation. Labrador Systems (in-home companion + mobility). Aeolus Robotics. Sensi.AI ($45M Series C Oct 2025). The aging population is one of the deepest buyer pools in the vertical map, and is structurally underserved.

Industrial / cell-line. Atomic Industries ($25M Series A, Detroit, AI injection mold automation). Inbolt (vision-enabled adaptive robot programming, opened US HQ in Detroit). HEBI Robotics (Pittsburgh, modular actuators for NASA / Army). Whelix / Mito Robotics (cell culture automation). roboLoop (e-waste dismantling). These are the picks and shovels for everyone else's robotics deployment.

The vertical map is no longer a list of demos. It's an annual-revenue chart.

The horizontal contenders, beyond SoftBank/Figure/1X

A few more worth tracking that don't fit cleanly in either camp:

• Agility Robotics -- Digit humanoid, GXO Logistics commercial deployment, the most public proof point that bipedal humanoids can do real warehouse work. Quietly the most "operationally deployed" of the humanoid set.
• Sanctuary AI (Canada) -- Phoenix, dexterous manipulation focus, slower public arc but interesting IP.
• Mentee Robotics (Israel) -- humanoid, deep-learning-first design.
• EngineAI and Fourier Intelligence (China) -- aggressive price/performance entries in the Chinese humanoid wave, expect them to compress US incumbent margins.
• Boston Dynamics (Hyundai) -- Atlas electric, Spot ongoing in industrial inspection. Strategic ownership, less venture-style.

The Chinese humanoid wave (Unitree, EngineAI, Fourier, plus Xpeng, Xiaomi, BYD entries) deserves its own scout pass. Their cost curve is structurally different from US humanoid economics, and the implications for the global vertical market -- especially in industries with low margins -- are going to be aggressive.

The convergence trade

The vertical and horizontal bets converge along three vectors, and that convergence is where most of the asymmetric scout returns are likely to live.

1. Vertical specialists become foundation-model fine-tunes. A horizontal humanoid platform with a good embodied-AI model can be specialized to laser-weeding, drywall, or surgical assist by retraining the policy layer. The vertical companies that survive are the ones whose moat is not the form factor but the customer relationship, the certification pathway, or the proprietary task data. Carbon Robotics is moat-by-data. CMR is moat-by-FDA. Symbotic is moat-by-Walmart.
2. Horizontal platforms only matter once three curves cross. Foundation model capability for embodied action, actuator and component cost, and deployment economics (training, integration, support) all have to clear simultaneously. The first time that happens in a single industry will be the moment the horizontal bet starts compounding faster than the vertical bet -- and the most likely first industry is general warehouse labor, followed by light manufacturing.
3. The middle gets acquired. Pure-play single-task companies with one customer get rolled up. Pure-play "platform" companies without revenue burn through their humanoid rounds. The companies that survive are either (a) deep vertical with regulatory or data moats, or (b) horizontal with a real customer and a clear path to per-unit economics. Everything in between is a 2028 acquisition by SoftBank, Amazon, or a Chinese strategic.

Investable layers

For scouts, four layers to map regardless of which side of the bet wins:

1. Actuators, sensors, batteries, hands. The picks-and-shovels of physical AI. Margin compression here helps everyone -- Apptronik, Figure, Symbotic, Built -- but the suppliers themselves can be 10x outcomes. HEBI, MIM-Robotics, dexterous-hand startups.
2. Vertical task winners with proprietary data. Carbon Robotics, Built, Gecko, CMR, Forest3, Sensi.AI. The "boring" companies that look like SaaS revenue plus hardware attach. These are scout layups because the buyer profile and ROI are knowable today.
3. Simulation, data, and embodied foundation models. The horizontal thesis only works if this layer matures. Companies training generalist robotics models, world models, and synthetic-data pipelines. This is where the asymmetric returns sit if the horizontal bet pays off.
4. The factory floor itself. Anduril Arsenal-1, Atomic Industries, Hadrian, and the next wave of robot-built factories. Robotics deploying robotics is a tighter loop than most of the market is pricing in.

What we're watching

• Whether SoftBank's ABB integration produces a credible humanoid play within 18 months, or whether it stays as an industrial cash cow
• Figure's Helix model performance on multi-step tasks at scale (not demo-tasks)
• Whether 1X's NEO ships into homes in 2026 and what happens in the data layer when it does
• The first vertical-task company to cross $100M ARR using foundation-model architecture (likely warehouse or ag)
• Chinese humanoid pricing -- and whether US export controls reshape the global market for components

For scouts: the easiest call is the vertical task winners -- they ship revenue and have clear customers. The hardest, highest-upside call is the embodiment foundation-model layer. The convergence is where the next decade of robotics outcomes get decided, and we should be talking to founders in both camps.