# A Robot Cost Curve Needs Somebody to Buy the First Thousand

> A robot cost curve only bends when someone buys the first thousand units. Digit 5 BOM targets, the AI stack, and who funds the autonomy data flywheel.

- Canonical URL: https://webofmike.com/a-robot-cost-curve-needs-somebody-to-buy-the-first-thousand/
- Author: Mike Moore (https://webofmike.com/about/)
- Published: 2026-11-20
- Last modified: 2026-11-08
- Tags: Robotics, Metrics, Generative AI, Thoughts
- Cite as: Mike Moore, "A Robot Cost Curve Needs Somebody to Buy the First Thousand", Web of Mike (webofmike.com), 2026-11-20. https://webofmike.com/a-robot-cost-curve-needs-somebody-to-buy-the-first-thousand/


<!-- ART: a printed BOM with one line, "left foot, spare", cost column blank. Never explained. -->

A robot cost curve is not a materials story. It is a who-buys-the-first-thousand story, because the bill of materials only falls when someone funds the volume, and the autonomy stack only gets cheaper when those units produce data you can train on. Agility's own filings put Digit 5 parts at about $150,000 at launch and a target of about $30,000 at 10,000 units a year. That drop does not happen on a slide.

This is [part 7](/capable-isnt-cheap-robots/) of the robot economics series.

## Why does a robot cost curve need someone to buy the first thousand?

Andra Keay's [November 7 Robots & Startups newsletter](https://robotsandstartups.substack.com/p/the-impact-on-robotics-of-geopolitics) put the headline numbers in one paragraph: each Digit 5 starts at about $150,000 in parts, with Agility expecting $75,000 at 1,000 robots a year and $30,000 at 10,000, plus an order on the books worth more than $300 million from one unnamed customer covering 1,000 robots over three years.

I went and checked those against Agility's public materials, not the newsletter.

The parts curve is Agility's own. In a [Form 425 transcript](https://www.sec.gov/Archives/edgar/data/2074973/000121390026113339/ea0309918-425_churchill11.htm) filed October 28, 2026, CFO Michael Beer said Digit 5 has "a bill of materials cost of the system of roughly $150,000," coming down to "around 75,000" at about 1,000 units a year, and "about $30,000 per unit" at 10,000. [Ars Technica](https://arstechnica.com/ai/2026/11/agility-robotics-nasdaq-debut-invites-scrutiny-of-humanoid-robot-market/) read the S-4 the same way. The October 6 investor-day deck shows Digit 4 around $125,000 and an illustrative Digit 5 path, with a footnote that the figures are rounded estimates.

So: **verified as Agility's disclosed estimates and targets.** Not verified as what a Digit 5 actually costs to build today, because Digit 5 is not in general production. Early access is still aimed at the first half of 2027.

The 1,000-robot order is also Agility's disclosure. The investor-day note on the $300 million-plus figure says it is potential multi-year value as of May 2026, subject to contractual milestones, for 1,000 Digit 5 robots on a three-year robots-as-a-service contract, with warrants that vest as robots are deployed. It is "not a measure of current period revenue." The purchaser is not named. Ars makes the same point: the contract does not mean Agility is guaranteed $300 million.

So: **verified as Agility's reported commitment, milestones and unnamed buyer included.** Not verified as cash. I am not going to guess the customer.

That is still the right shape of the problem. Somebody has to take the expensive units so the next units can be cheap. The difference with a humanoid is that those units are also the training set for the software that makes the cheap ones useful.

## What is in a robot bill of materials versus the AI stack?

A BOM is the parts: actuators, structure, batteries, cameras, compute, harnesses, the left foot. Beer said volume should yield about 20% of the planned savings, with the rest from suppliers and redesign.

I don't have Agility's spreadsheet. I can still split the cost the way a test engineer splits a station: what you buy once, and what you keep paying because the station thinks.

```text
Digit 5, one unit, launch-era sketch
(Agility's numbers where cited. The rest is a split, not a leak.)

Hardware BOM (Agility, launch estimate)          $150,000
  actuators, structure, battery, sensors, compute
  left foot, spare                                      ____
Vendor deploy (Agility, illustrative)              $15,000
Vendor software + maintenance / year (Agility)     $15,000

What the customer is shown (illustrative S-4 model)
  RaaS: ~$25,000 deploy + $8,500 / month
  or buy: ~$200,000 hardware + ~$20,000 deploy
          + ~$36,000 / year software and maintenance

AI / autonomy stack, ongoing, not in the BOM
  on-robot perception and control models
  skill training (sim, teleop, site data)
  fleet software (Arc-class: dispatch, KPIs, updates)
  evals and safety re-cert on every model drop
  human interventions while the skills are thin
```

The blank spare-foot line is the tell. If you cannot price the spare, you don't have a cost curve. You have a prototype with a hole in the spreadsheet. The same hole shows up in the AI column: the extra policy, the second eval suite, the night-shift fine-tune. If that cell is blank, the $30,000 target is a parts story that forgot the brain.

Look at the RaaS row. Even in Agility's own model, the customer keeps paying a monthly software-and-maintenance fee after the robot is on the floor. Beer also said Agility planned about $110 million of R&D in the then-current year. Hardware BOM can fall to $30,000 and the subscription can still be the business. That is the AI stack refusing to sit inside a parts spreadsheet.

I already made this point the other way in [capable isn't cheap](/capable-isnt-cheap-robots/): Anthropic's cost column includes software and supervision, not just the arm. $8,500 times 12 is $102,000. That is Agility's illustrative fee, and the S-4 says actual terms are subject to negotiation. **Software does not follow the same slope as steel.**

Duty cycle still multiplies both columns. [The 20-hour robot](/the-20-hour-robot/) is why Agility led with a 10:1 charge ratio. More hours amortize the BOM. They also generate more traces for the model. A robot that sits on a charger is expensive twice: once in unused iron, once in unused data.

## Who pays for the autonomy learning curve?

Three parties can fund the first thousand. They are not equally motivated.

**The launch customer.** They get capacity. They also get the bugs, the interventions, and the argument about whose lawyers own the site video. If the contract is robots-as-a-service, they are paying for uptime, not for Agility's next skill. They will tolerate data leaving the building only as far as the contract forces them to.

**The vendor.** Under RaaS, Agility's deck says the deployed robots sit outside a small capex line, with the company wanting cheap debt to fund the fleet once public. That means Agility holds the learning-curve cost on its balance sheet and makes it back on the subscription. Fine, if the subscription holds and the milestones on that 1,000-unit commitment actually release the batches. The filing is explicit that they might not.

**Investors.** Capital buys time before the first thousand exist. It is not a substitute for a factory that has built them.

The AI question is which of those three is paying for labels and evals, not just actuators. A cheaper drive does not label a missed sit. A thousand robots that stream traces into a training pipeline do, if you are allowed to use the traces.

Agility's investor-day slides say the quiet part: deployments power a data flywheel. Digit in the field, operational data, better embodied AI, more workflows, more deployments. That is the same loop I wrote about in [two percent a year](/robots-two-percent-a-year/), where Agility's CTO argued that getting robots out to practice is the main lever, and roboticist Chris Paxton argued that most deployment data "is almost certainly less valuable than you think it is."

I still land where I landed then. The hardware flywheel is the solid one: volume, suppliers, redesign. The data flywheel is real and messy. Site data is biased toward the jobs, lighting, and people you already have. It is not a held-out safety suite. A skill that only works in one building can still sell to the next building like it, which is how you get from 1,000 to 10,000.

Hands still set the ceiling. [The hard part is the hands](/robots-hard-part-is-the-hands/). A cheaper BOM with the same gripper does not pick the tangled wire. Buy the first thousand for jobs the hand can already do.

## What data do the first thousand robots actually produce?

Not "the internet, but for warehouses." Specific traces:

- Joint torques, contact, slips, sits, e-stops
- Camera and depth frames around every intervention
- Task success and the reason for a human takeover
- Battery, thermal, and charge-cycle logs
- Model hash and skill version on every hour of work

That list is an eval corpus if you keep it, version it, and refuse to train on unlabeled failures. It is a liability if it streams off-site with no inventory. The cost-curve point is narrower: **the first thousand are expensive because they are the calibration run for the model, not only because the actuators are still full price.**

If you are the buyer, ask who pays for that calibration:

```text
First-thousand cost questions
[ ] Is the $150,000 a BOM, a price, or a slide?
[ ] What is the spare-foot line, and every other spare, in dollars?
[ ] How much of the year-one invoice is software, not iron?
[ ] Who owns site video and intervention traces?
[ ] Are those traces in the training set for the next skill I didn't buy?
[ ] What eval re-runs when the perception hash changes?
[ ] How many human hours per robot-hour in month 1 vs month 12?
[ ] What happens to my rate if the 1,000-unit milestones slip?

If the spare foot is still blank, the rest of the sheet is a wish.
```

Hardware is still the teeth. I wrote that in [hardware is AI's teeth](/hardware-is-ais-teeth/). Teeth without a buyer are a lab. Aim the first thousand at boring jobs where a cheaper robot and a better detector compound, not at [the last 20 percent](/the-last-20-percent-of-work/) of hands-on work a BOM target cannot touch.

Someone has to write a check that makes the $150,000 unit exist so the $30,000 unit can. Until that check clears, the cost curve is a drawing.

## Related

- Andra Keay, [The impact on robotics of geopolitics](https://robotsandstartups.substack.com/p/the-impact-on-robotics-of-geopolitics), Robots & Startups, Nov 7, 2026
- Michael Beer, [Churchill Capital Corp XI Form 425](https://www.sec.gov/Archives/edgar/data/2074973/000121390026113339/ea0309918-425_churchill11.htm), Oct 28, 2026
- Ars Technica, [Agility's Nasdaq debut](https://arstechnica.com/ai/2026/11/agility-robotics-nasdaq-debut-invites-scrutiny-of-humanoid-robot-market/)
- Agility, [Analyst and Investor Day exhibit](https://www.sec.gov/Archives/edgar/data/2074973/000121390026107259/ea030784201ex99-1.htm), Oct 6, 2026
- Series: [part 1](/capable-isnt-cheap-robots/), [part 2](/robots-hard-part-is-the-hands/), [part 3](/robots-two-percent-a-year/), [part 4](/the-20-hour-robot/), [part 5](/hardware-is-ais-teeth/), [part 6](/the-last-20-percent-of-work/)

