Industrial: the one you can underwrite today

01The subsector that already works

Everywhere else in Physical AI we are underwriting a future. In industrial robotics the future has partly arrived: robots are in production, generating revenue, and clearing payback. The master thesis called applied autonomy a conviction because you can underwrite it on unit economics rather than narrative, this chapter is where that claim is cashed. The reliability numbers are here too: narrow, structured tasks already run at production grade (Amazon's Vulcan picking above 99%; DYNA folding at 99.4%), which is exactly the regime a factory or fulfilment centre lives in.

Demand is structural: turnover in warehousing runs above 100% a year, the roles are dull and physically punishing, and the labour simply is not there to hire. That is why the market compounds at high-teens rates and why Symbotic can carry a $22.7B backlog on roughly $676M of quarterly revenue, years of visibility, and, as of this year, the segment's first GAAP profit.

02The framework, stress-tested

Industrial runs the master framework forward and surfaces a risk the other verticals do not: the recombination future is arriving here, but the biggest beneficiary of it may be the customer, not a startup.

The arm is a commodity; value moved to the model and the data

The traditional industrial-robot market is mature, concentrated in the "Big Four" (Fanuc, ABB, Yaskawa, and Midea-owned KUKA), and increasingly Chinese, China alone was 54% of 2024 installations. An entry-level six-axis arm is $17–20k and falling. This is the master's recombination world made concrete: the body is a commodity substrate, and the value migrates to the perception-and-policy model that lets a cheap arm do the variable, high-mix, brownfield tasks classical hard-automation never could, bin picking, mixed-SKU handling, defect detection. The winners are not the arm makers; they are whoever owns the model and, crucially, the proprietary task data that improves it.

The twist: the buyer is also the builder

Defense taught us that the buyer's preference shapes the market structure. Industrial teaches the darker version of the same lesson. Here the largest buyer, Amazon, is simultaneously the largest builder: it operates its own million-robot fleet, has built DeepFleet, a generative-AI foundation model that coordinates that fleet and cut travel time ~10%, unveiled its own multi-task Blue Jay workstation, and in 2024 reverse-acquihired Covariant, the best AI-picking team in the market, licensing its foundation-model tech and hiring its founders. The lesson is unambiguous: in industrial, a third-party vendor whose value is a general-purpose model sells into customers who have every incentive, and increasingly the capability, to internalise it. The warehouse-robotics SPAC wave already delivered the cautionary tale, Berkshire Grey taken private at a fraction of its peak.

03Humanoids on the line: optionality, not production

The most-hyped datapoint in industrial deserves a cold read. As of mid-2026 no humanoid is in true volume production on a paced line. The best hard evidence is Figure's BMW Spartanburg pilot: over roughly ten months, Figure 02 ran 10-hour shifts and moved 90,000+ components across ~1.2M steps, supporting production of 30,000+ vehicles, genuinely impressive, and still a fixed-term pilot, not permanent line integration. Agility's Digit at GXO is the closest to commercial, structured as a multi-year Robotics-as-a-Service agreement, but limited in scope. Apptronik's Apollo is in pilots at Mercedes and GXO.

Meanwhile capital has run ahead of production: Apptronik raised $520M at a ~$5B valuation, Agility is taking the SPAC route to Nasdaq, Figure carries a mega-valuation. This is the master thesis's humanoid caution in its native habitat, the flexibility-weighted case is strongest in mixed-task brownfield industrial settings, but throughput at human parity for general tasks is not yet demonstrated at scale. We treat industrial humanoids as a high-optionality, pre-production sleeve, sized accordingly, and we watch the throughput numbers, not the demo reels.

04Orien's verdict

Industrial signposts we track

• Does RaaS margin expand? If a leading RaaS operator discloses software-like gross margins as fleets scale, the "SaaS in the physical world" thesis is validated. If margins stay in the low 20s, it stays a capital-intensive business to underwrite conservatively.
• Hyperscaler insourcing pace. Watch Amazon's DeepFleet/Blue Jay expansion and any further reverse-acquihires. The more the hyperscalers build, the narrower the third-party opportunity and the more the mid-market discriminator matters.
• First humanoid in paced production. The moment any humanoid moves from fixed-term pilot to permanent, throughput-guaranteed line integration, the optionality sleeve re-rates. Track operating hours and interventions, not pilots announced.

05What breaks this call

Hyperscaler vertical integration. The central risk, and it is structural: the biggest buyers are also the biggest builders, and they can strand third-party vendors by insourcing the model layer or acquihiring the team. Our mitigation, serve the fragmented mid-market and own the data, is exactly that, a mitigation, not an immunity.

The margins never turn software-like. If RaaS and integration stay hardware-heavy at low-20s% gross margins, the segment is a good business but not a venture-return business, and entry multiples matter enormously. Underwrite to the economics that exist.

Commoditisation from below. Chinese arms and, increasingly, Chinese full systems are cheaper and improving; the security overlay protects Western high-value deployments but not price-sensitive, low-sensitivity ones. Margin compression is the base case for anything without a data or relationship moat.

Humanoid capital misallocation. The sleeve is real optionality, but the current funding pace is well ahead of demonstrated production throughput. Sizing discipline is the whole risk-management story here, this is where the master thesis's "heavy hardware is often a trap for an independent fund" warning bites hardest.

06Sources