Per-Unit Pricing: Examples & Companies

What is it

Per-Unit Pricing is a billing unit used by robotics, hardware AI, and some SaaS companies where the metered object is a physical or abstract “unit” — a robot deployed, a device sold, or a defined deliverable.

The word “unit” is doing a lot of work here, and that ambiguity is the whole point. In physical AI, a unit is a tangible object: a humanoid robot frame, an autonomous drone, a licensed model deployment running on a device. In software, a unit is an invented accounting abstraction — Labelbox prices data work per Labelbox Unit (LBU) and Databricks’ Mosaic AI prices every AI workload per DBU (Databricks Unit) — a normalized token that lets one meter span workloads with no shared natural dimension. What unites both families is that the definition lives in the contract or the rate card, not in any universal standard the buyer can independently verify.

The physical family dominates embodied AI: humanoid and robotics vendors — 1X Technologies, Figure, Apptronik, Sanctuary AI, Covariant — all sell or lease per unit, almost always behind a sales quote. The abstract family solves the opposite problem: when a platform sells several products with wildly different cost structures, one invented unit collapses three rate cards into a single number. Either way, per-unit pricing trades a clean sticker price for flexibility the vendor keeps for itself — and the two families fail buyers in mirror-image ways.

How it works

Per-unit pricing splits into two mechanically distinct families that happen to share the word “unit.” Getting the distinction right is the key to reading any of these bills.

The physical unit is the simplest: it’s a hardware transaction wearing a subscription costume. 1X Technologies sells NEO for a one-time $20,000 (Early Access, priority 2026 delivery, $200 refundable deposit) or a $499/month RaaS lease on a six-month minimum. The monthly figure is a flat fee, not a metered per-task charge — the “meter” is simply how many robots you have. Channel data on Skydio drones shows the same shape: analysts cite X10 kits around $15,000–$20,000 and X10D units at $30,000–$50,000, though Skydio itself publishes no first-party rate.

The abstract software unit is where per-unit pricing behaves like genuine consumption billing. Consider Labelbox’s LBU, the most transparent example in the corpus:

At a flat $0.10/LBU, the same dollar buys 60 curated rows or a single labeled row — a 60x asymmetry hidden inside one meter. Databricks’ Mosaic AI works the same way: a token of inference, an hour of GPU serving, and a fine-tuning run all roll up into DBUs (AI workloads list at $0.07/DBU, model training references $0.65/DBU), metered per second, with the crucial caveat that the DBU is the software charge only — you pay your cloud provider separately for the underlying compute.

Unit math (physical): Total = (units deployed × per-unit price). A 10-robot NEO fleet on subscription = 10 × $499 = $4,990/mo.

Unit math (abstract): Total = (rows × per-row LBU rate × $/LBU). Labeling 50,000 rows in Labelbox Annotate = 50,000 × 1 LBU × $0.10 = $5,000.

Companies using this

Fourteen companies in the corpus meter on units, and they split cleanly into two camps: physical AI (humanoids, drones, autonomous aircraft) where the unit is a robot, and ML/observability platforms where the unit is a software abstraction like an LBU or DBU. 1X Technologies is the canonical public-price example on the physical side; Labelbox is the clearest abstract-unit example on the software side.

Patterns observed

A $20,000 humanoid and a $0.07 software unit share no obvious category — yet a few patterns hold across the whole set.

• The unit is deliberately vague, and that’s a feature. Every company here uses “unit” precisely because it doesn’t commit them to a natural dimension. One LBU meter spans curation, labeling, and evaluation; DBUs fold inference, serving, and training into one number; on the physical side, Figure and Apptronik leave the definition — outright sale vs. RaaS lease, what’s bundled — entirely to the contract. That vagueness preserves pricing flexibility a per-token or per-seat meter would surrender.

• Physical-unit vendors overwhelmingly hide the price. Only 1X Technologies publishes a real, self-serve number; Figure, Sanctuary AI, Covariant, Apptronik, Shield AI, and Physical Intelligence all quote via sales. They almost all carry sales-led as their motion because per-unit production cost is still falling and each deployment needs custom integration.

• Abstract-unit vendors publish the rate but hide the conversion. The headline number is clean; the real cost hides in the per-operation rate. Beyond Labelbox and Databricks, Langfuse meters “units” (traces + observations + scores) at $8 per 100k overage and LangChain prices LangGraph compute in LCUs at $1.50/LCU — same shape, different token.

• The published unit is often the smaller half of the bill. Software-unit vendors rarely bill units alone: Databricks Mosaic AI makes you pay your cloud provider for compute on top of the DBU, and Labelbox layers sales-quoted human labeling (Alignerr) over the per-LBU charge — services that typically dwarf the software meter.

• Robotics-as-a-Service reframes the unit as a subscription. 1X Technologies and Apptronik (Apollo) offer a per-unit lease rather than an outright sale — the unit stays the robot, but the billing shape shifts to a recurring fee that anchors against a human wage rather than a hardware sticker.

Counterexamples & variants

The cleanest counterexample to “per-unit means opaque enterprise pricing” is 1X Technologies, which attaches a real self-serve figure to a consumer humanoid while the rest of the sector quotes bespoke. The catch is what the price buys: at launch NEO leans on remote human teleoperators (“Expert Mode”) for many chores, so the sticker buys an evolving, partly-teleoperated product whose autonomy is expected to grow. The unit is public, but the value behind it is still moving — the mirror image of the abstract family, where the rate is published but the per-operation conversion is booby-trapped, so the price is visible while the bill isn’t.

The “unit” label sometimes papers over a different model entirely. Snorkel AI and Essential AI carry units in their billing dimensions but sell annual enterprise subscriptions quoted via sales — the “unit” is closer to a licensed platform or a contracted deliverable than a metered token. Physical Intelligence is the extreme edge: a pre-commercial lab that sells nothing yet and releases its π0 weights free under Apache-2.0, so its “unit” is entirely prospective. When you see units on a company really running commitment or subscription pricing, treat the label as a placeholder for whatever the contract defines — not as evidence of usage-based billing.

What this means for buyers vs vendors

For buyers

First figure out which kind of unit you’re buying. For a physical unit, treat the quote like a hardware purchase or lease: get total cost of ownership — integration, teleoperation, maintenance, any RaaS minimum term — and don’t assume the sticker is the whole bill. For an abstract unit, the published rate is nearly useless alone; demand the per-operation conversion table before you sign, and always model the separate cloud-compute line where one exists. Our guide to choosing the right usage metric walks through pressure-testing whether a vendor’s unit maps to the value you receive, and the pricing calculator hub helps you model these scenarios before you commit.

For vendors

Per-unit pricing fits when your product has no clean natural meter — a physical object that can’t bill “per token,” or several workloads that don’t share a dimension. The abstract unit buys two things: you can change one product’s underlying economics without renegotiating contracts, and you present one bill instead of three rate cards. The cost is buyer trust — an opaque unit invites suspicion, so publishing your conversion rates builds more credibility than hiding them. In physical AI, decide deliberately: follow 1X Technologies into public pricing (which wins PLG-style demand and press) or stay sales-led like Figure and Shield AI (which protects margin while unit cost is still falling). For the mechanics of standing up a unit meter, our guide to tracking and metering usage events covers the instrumentation.

Explore this theme in the knowledge graph →