## Is Sanctuary AI Pivoting Away from Its Phoenix Humanoid?

**99.5% task success rate, 2.54-second cycle time, on a competitor's robotic arm.** That is the headline number from [Sanctuary AI](https://humanoidintel.ai/companies/sanctuary-ai)'s latest factory validation, and it tells you almost everything about where the Vancouver company is heading. On June 17, Sanctuary reported that its Carbon AI control system completed a wire-plugging task for a global Tier 1 automotive supplier — threading flexible wires into moving targets — at production-line speed, validated against the customer's live benchmarks. The critical detail: Phoenix, Sanctuary's own humanoid, was not involved. Carbon ran on a rival's robotic arm.

Under CEO Daniel Friedmann, who the company says moved from chairman to chief executive in June after nearly two decades running aerospace company MDA, Sanctuary is now explicitly positioning Carbon as a licensable software layer for existing industrial robots. With total funding reported by Bloomberg and CB Insights at approximately C$140 million to $148.6 million — a fraction of what [Figure AI](https://humanoidintel.ai/companies/figure-ai) closed in its Series C — the logic is straightforward: you do not compete on hardware capital intensity when you do not have hardware capital. You find the layer where a smaller company can still matter.

That layer, for Sanctuary, is Carbon.

---

## What Carbon Actually Did in That Factory

The wire-harness task is a meaningful benchmark, not a marketing one. Wire plugging is a canonical hard problem for [dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation): flexible cables resist predictable positioning, connector tolerances are tight, and the task demands real-time compliance to avoid damage. Doing it on a moving target at a 2.54-second cycle time — at 99.5% success — is contact-rich work at the edge of what most industrial automation handles reliably.

What makes Sanctuary's result analytically interesting is the platform it ran on. This was not an internal demo on Phoenix hardware tuned to Carbon's assumptions. Running Carbon on a third-party arm means the system had to generalize across a different kinematic chain, different actuator dynamics, and different sensor modalities than the ones Sanctuary's engineers built around. That is a meaningful, if still limited, proof of cross-platform generalization.

The missing details matter, though, and Sanctuary has not provided them: the automotive supplier is unnamed, Carbon's standalone licensing price is undisclosed, and there is no published evidence yet that the same performance holds across multiple customer sites, part variants, or robot platforms. A 99.5% result on one wire harness line is a serious proof point. It is not a product-market fit story.

---

## The Hardware Gap Sanctuary Is Working Around

Sanctuary's pivot makes more sense when you map it against the competitive landscape it cannot win on current terms.

[Figure AI](https://humanoidintel.ai/companies/figure-ai) closed more than $1 billion in Series C funding in September 2025 at a $39 billion post-money valuation, per the company's own announcement. It is scaling Helix, its embodied intelligence system, alongside BotQ, a manufacturing facility whose first line was designed for up to 12,000 humanoids per year, with stated supply chain ambitions reaching 100,000 robots over four years. [Tesla (Optimus Division)](https://humanoidintel.ai/companies/tesla-optimus) is integrating its humanoid program into a broader manufacturing machine. [Unitree Robotics](https://humanoidintel.ai/companies/unitree-robotics) is pressing from the cost floor, with its own store listing the G1 at $13,500 and the R1 from $4,900, and the company opened book-building for a Shanghai STAR Market IPO on August 5, targeting a valuation near 42 billion yuan.

Against that field, Sanctuary's reported funding — approximately C$140 million as of early July 2024 per Bloomberg, with CB Insights listing total funding at about $148.6 million — is real capital for a focused software bet. It is not sufficient to compete on humanoid hardware volume. Friedmann appears to have read the same spreadsheet the rest of the industry should be reading.

---

## Why the "Robot Brain" Model Has Historical Precedent — and a Specific Problem Here

The software-layer strategy is not new in robotics. Nvidia's Isaac GR00T N1.7, a [vision-language-action model](https://humanoidintel.ai/glossary/vision-language-action-model) for generalized humanoid skills, is distributed as open weights under Apache 2.0 through GitHub and Hugging Face — essentially the same playbook: let others build bodies, capture value in the intelligence layer. The smartphone era codified the broader principle: the operating system can compound value even when someone else manufactures the device.

Sanctuary's version of this model carries a wrinkle that Nvidia does not face. Nvidia does not sell a competing humanoid body. Sanctuary does. Phoenix has gone through at least eight generations since its public debut in May 2023, with a sixth-generation announcement that year, a seventh generation announced in April 2024, and a Generation 8 platform described as built for higher-quality data capture. If a factory procurement team can get near-production reliability from Carbon running on robot arms it already understands, maintains, and has safety-certified, the urgency of a Phoenix purchase drops materially.

Sanctuary's implicit answer seems to be that it can tolerate that cannibalization risk — because every Carbon deployment on third-party hardware generates task data that improves the one asset the company is clearly compounding: the AI system itself. Licensing Carbon on installed hardware should close faster than selling a full humanoid, and faster closure means faster data. In a field where training data quality increasingly determines model capability, that is not a trivial tradeoff.

---

## What This Means for the Full-Stack Humanoid Thesis

The conventional full-stack argument — pursued by Figure, Tesla, and [1X Technologies](https://humanoidintel.ai/companies/1x-technologies) — is that hardware and software must co-evolve under one roof to achieve the feedback loops necessary for real-world generalization. 1X said in July that its NEO hands are tightly integrated with in-house motors, electronics, [tendon-driven](https://humanoidintel.ai/glossary/tendon-driven) systems, and firmware. The argument is that you cannot optimize [whole-body control](https://humanoidintel.ai/glossary/whole-body-control) if you do not own the whole body.

Sanctuary's wire-plugging result is a direct challenge to that instinct — or at least a boundary condition on it. A contact-rich manipulation task, completed at line speed, on hardware Sanctuary did not design, at 99.5% reliability: that is exactly the kind of evidence the full-stack argument has to account for, not dismiss. It does not prove that Carbon generalizes broadly. It does prove that Carbon generalizes beyond Phoenix.

The industry will be watching whether Sanctuary can name a customer, publish a commercial Carbon licensing offer, and show the same performance across a second and third robot platform. Until then, the 99.5% number is the most credible signal the company has produced in some time — and it came from a factory floor, not a stage.

---

## Key Takeaways

- **99.5% success rate, 2.54-second cycle time** on a wire-plugging task for a Tier 1 automotive supplier — Carbon ran on a third-party robotic arm, not Phoenix.
- **Sanctuary is explicitly pivoting** Carbon as a licensable AI layer for existing industrial robots, separating the "robot brain" from its own humanoid hardware.
- **CEO Daniel Friedmann**, formerly chairman and a long-tenured aerospace executive, is leading the shift — signaling a commercial, not purely technical, reorientation.
- **Reported total funding** of approximately C$140–$148.6 million (per Bloomberg and CB Insights) makes a hardware volume race against Figure AI or Tesla structurally untenable.
- **Key unknowns remain**: the automotive customer is unnamed, Carbon's standalone pricing is undisclosed, and cross-platform generalization beyond this single task has not been demonstrated publicly.
- The result is the most concrete evidence to date that humanoid AI control systems can perform meaningfully on non-native hardware — a finding that complicates the full-stack humanoid thesis.

---

## Frequently Asked Questions

**What is Sanctuary AI's Carbon system?**
Carbon is Sanctuary AI's AI control system, originally developed as the intelligence layer for its Phoenix humanoid robot. It handles task planning, reasoning, and motor control. Sanctuary is now positioning Carbon as a standalone, licensable software layer that can run on third-party robotic arms, not just its own hardware.

**What did Sanctuary AI's factory test actually prove?**
On June 17, Sanctuary reported that Carbon achieved a 99.5% task success rate on a wire-plugging task for a global Tier 1 automotive supplier, at a 2.54-second cycle time. The test ran on a competitor's robotic arm, validated against the customer's live production benchmarks — not an internal lab setting.

**How does Sanctuary AI's funding compare to Figure AI?**
Bloomberg and CB Insights report Sanctuary's total funding at approximately C$140 million to $148.6 million. Figure AI closed more than $1 billion in a single Series C round in September 2025 at a $39 billion post-money valuation, per Figure's own announcement. The capital difference is an order of magnitude.

**Is Sanctuary AI abandoning Phoenix?**
Not explicitly. Phoenix has reached at least an eighth generation, with the latest platform described as built for higher-quality data capture. However, the company's clearest commercial signal right now is Carbon as a cross-platform software license, not Phoenix hardware sales. Whether that constitutes a pivot or a parallel track depends on what Sanctuary announces next on the Phoenix roadmap.

**How does Carbon compare to Nvidia's GR00T approach?**
Both target the intelligence layer rather than the hardware layer. Nvidia's Isaac GR00T N1.7 is a vision-language-action model distributed as open weights under Apache 2.0. Carbon is a proprietary system Sanctuary intends to license commercially. The key structural difference: Nvidia has no competing humanoid body to cannibalize; Sanctuary does, which creates a more complex internal incentive structure.