Humanoid Deployment Meets the Data Factory: How Real Factory Rollouts Are Rewriting the AI-Robotics Capital Stack

The last week of June marked a quiet regime change in humanoid robotics. Three announcements — Apptronik's Robot Park + Apollo 2, BMW's Figure 03 deployment, and CATL's heavy-duty humanoid going live inside a CATL factory — moved the sector past "impressive demo videos" and into "real factory floor, real production KPIs." For allocators tracking AI-robotics as a research vertical, the important story isn't which humanoid company will "win." It's how the capital stack is bifurcating between the visible robot layer and the invisible data-factory layer that trains them.

What actually happened

On June 30, Apptronik opened Robot Park — a purpose-built training and demonstration facility for its Apollo humanoid — and unveiled Apollo 2, the next-generation platform. Robot Park's stated purpose is more revealing than the robot itself: it's a data-generation environment designed to log tens of thousands of hours of physical manipulation data across warehouse, assembly, and logistics tasks. Apptronik is treating the training environment as the differentiated asset.

On July 1, video surfaced of BMW deploying Figure 03 humanoids at its Spartanburg, South Carolina plant. Unlike prior humanoid pilots (which mostly involved carrying totes between marked stations), the Figure 03 units are performing genuine assembly tasks alongside human workers — inserting sheet-metal components, torquing fasteners, and moving parts through actual production cells. BMW has been Figure's most public partner since 2024; this is the first footage that looks like production rather than choreography.

And on the same day, CATL announced the first heavy-duty humanoid powered by CATL batteries went live inside a CATL factory. The robot is built by Galbot, a Chinese humanoid startup, and is designed for high-payload tasks — battery-module handling, chassis positioning — that are physically brutal for human workers. CATL disclosing this deployment publicly matters because CATL is both the largest EV battery manufacturer in the world and one of the few counterparties with the internal factory footprint to prove humanoid economics at scale.

Why the timing matters

The 2024–2025 humanoid narrative was demo-driven: Tesla Optimus stage walks, Figure GPT integrations, Agility Digit tote-carrying. Impressive but bounded. What changed in Q2 2026 is that the same customers who buy industrial automation now have humanoid line items in real capex plans. BMW's Spartanburg deployment isn't a pilot — Figure has been embedded on the plant floor for over a year, and the July footage is what productionization looks like at the boring, incremental end. CATL's Galbot deployment is similar: not a partnership press release, an actual load-bearing production unit inside the world's largest battery plant.

This shift changes what "AI robotics" means as an investment vertical. Two years ago the debate was hardware — actuators, batteries, dexterity. Today the hardware bottleneck is mostly solved. The new bottleneck is training data: how many hours of high-quality, task-diverse manipulation data can a company log per week, and how efficiently can they convert that data into policies that generalize across tasks and platforms?

The data factory as the real asset

Robot Park is the clearest expression of this thesis. Apptronik isn't primarily marketing a new humanoid — it's marketing a training environment. The reasoning is straightforward: once hardware converges on a rough set of viable form factors, the winning position is whoever can generate the most useful training data fastest. Every hour a humanoid works in a real factory generates data. Every hour a competitor's humanoid does not, is a compounding disadvantage.

Figure's version of this is the BMW relationship itself — over a year of continuous manipulation logs from a real assembly line. Galbot's version is the CATL deployment. What all three share is a customer partnership that doubles as a data-generation contract. The customer gets useful labor; the humanoid company gets proprietary training data at industrial scale.

For allocators, this reframes valuation. Humanoid companies without durable factory deployments generating proprietary data are hardware companies competing on unit economics. Humanoid companies with durable deployments are data companies with a hardware distribution channel. Those are very different businesses, and public secondary markets are only beginning to price the distinction.

What we're watching in the AdValorem research desk

• Deployment velocity as the primary KPI. Not "robots produced" but "robot-hours logged on real production floors per month." Companies that don't disclose this will increasingly be assumed to have low numbers.
• Battery + power-electronics counterparties. CATL's Galbot deployment signals that battery leadership can extend into humanoid platform leadership. The overlap between EV cell manufacturers, energy-storage vendors, and humanoid power systems is a research topic we're actively expanding coverage on.
• Training-data licensing structures. A humanoid company operating inside BMW or CATL is generating data that both parties have arguable claims to. The contract structures here — who owns the logs, who can train on them, whether policies trained inside one customer can be sold to competitors — will decide who captures the economic surplus.
• Secondary tape on Figure, Apptronik, and Physical Intelligence. Pre-IPO secondary trading on Hiive, EquityZen, and Forge has been active for all three names. Fair-value math needs to increasingly account for deployment data velocity, not just headline round valuations.

The takeaway for our research community

The interesting question in humanoid robotics is no longer "will they work?" — the June deployments answer that. The question is "who has the data-generation flywheel, and how defensible is it?" That's the frame we're using across the AdValorem research desk as we cover Apptronik, Figure, Physical Intelligence, Galbot, and the broader AI-robotics stack. If you're an operator or founder in AI, robotics, or industrial automation and want to compare notes with the community following this closely, our Discord and weekly research are the place — links below.