What's the difference between captive and third-party brain providers in physical AI?

The brain-provider gradient: captive vs third-party

Physical AI brain providers split structurally across two integration models with different counterparty-risk profiles + cross-device-interoperability postures + competitive-dynamics implications.

Captive brain providers vertically integrate the AI substrate within the same corporate-state entity that builds the hardware embodiment. The AI and the platform ship together under the maker's full-stack control. The captive-brain positioning shows up in consumer-archetype humanoid platforms more often than industrial-archetype platforms per the consumer-vs-industrial humanoid sub-cohort architecture: consumer hardware that ships into homes needs end-to-end control over privacy + data handling + UX consistency + subscription-tier feature differentiation; vertical integration supports that cleanly.

Third-party foundation-model providers license or partner-integrate AI substrate that runs across multiple hardware platforms. The brain ships as infrastructure; the hardware ships separately; integration partnerships are the commercial mechanism. The third-party-foundation-model positioning dominates the brain-providers cluster per the brain-providers cluster framework; it shows up disproportionately in industrial humanoid integration claims and in cross-cohort foundation-model deployment.

The captive-vs-third-party gradient is editorial signal at the cohort-architecture layer. Trade-press coverage that frames humanoid robots as uniformly third-party-AI-platform-driven operates outside the verified brain-provider integration spectrum; trade-press coverage that frames captive-brain integration claims at scaled-commercial-deployment depth operates outside the verification posture appropriate to maker-stated captive-stack claims.

---

Captive brain: the verified canonical worked example

1X Redwood is the canonical captive-brain worked example. Registered at /brains/1x-redwood as a brain entity (NOT a humanoid hardware product per the audit-first verification scope), Redwood is 1X Technologies' foundation model that powers both 1X NEO consumer humanoid and 1X EVE industrial humanoid under a single corporate-state entity. The vertical integration positioning is the load-bearing structural distinction: 1X Technologies controls the AI substrate + hardware embodiment + commercial-model framing across both consumer + industrial product lines.

The captive-brain positioning enables structural advantages that third-party-brain integration doesn't replicate cleanly:

• Consumer-UX consistency: when the same maker controls AI + hardware + product-tier framing, consumer experience consistency across feature releases + safety improvements + privacy-handling defaults stays under single-corporate-state ownership. Third-party-brain integration introduces multi-party coordination overhead at every UX-affecting change.
• Privacy-handling control: sensor data flows + cloud-vs-edge processing + data-residency posture + retention defaults all sit under single-corporate-state policy ownership. Third-party-brain integration means data-handling policy gets negotiated across maker + brain-provider boundaries.
• Subscription-tier feature differentiation: gated feature unlocks tied to subscription tier require AI-substrate + hardware-platform coordination. Captive-brain integration unlocks features cleanly; third-party-brain integration requires per-integration negotiation.

The 1X Redwood positioning surfaces the consumer-archetype-captive-brain intersection editorially: per the consumer-vs-industrial humanoid sub-cohort architecture, consumer humanoid archetype operates more likely captive-brain integration; Redwood at 1X is the canonical worked example surfaced at the intersection.

---

Captive-brain candidates: cap-flagged pending verification depth

Several other humanoid platforms operate captive-brain positioning at the maker-stated level, but verification depth varies. Per cap-flag-as-trust-signal, the cap-flag discipline surfaces the verification posture honestly rather than papering over the gap.

Figure Helix on Figure 03: Figure operates the Helix VLA model as native captive AI substrate per Figure-stated framing. Per the editorial framing from the NVIDIA GR00T entity anchor, "Figure operates Helix VLA in-house foundation model"; Figure is competitor not deployment-partner relative to NVIDIA GR00T's third-party-foundation-model positioning. The captive-brain positioning is verified at the cluster-framework + competitive-positioning level; specific Helix architecture depth + training-methodology + deployment-scale claims operate at lower verification posture pending primary-source confirmation from Figure.

Tesla's FSD-derived stack on Optimus: Tesla operates Tesla Optimus at consumer-aspirational positioning per the consumer-vs-industrial humanoid sub-cohort architecture; the Musk-stated FSD-derived stack framing positions Tesla's existing Full Self-Driving infrastructure as the captive-brain substrate for Optimus. Per cap-flag discipline, specific FSD-to-Optimus stack architecture claims + specific captive-brain training-methodology claims operate at lower verification posture pending Tesla primary-source confirmation. The Tesla automotive-OEM crossover archetype shares structural DNA with XPENG Iron; both operate captive-brain positioning at the maker-stated level with verification depth varying by claim granularity.

Sanctuary AI cognitive-architecture on Phoenix: Sanctuary AI operates a cognitive-architecture-positioned AI substrate on the Phoenix humanoid platform per Sanctuary-stated framing. Per cap-flag discipline, specific architecture-depth claims at named-substrate granularity operate at lower verification posture pending primary-source confirmation. The Canadian-headquartered Sanctuary positioning extends the captive-brain archetype geographically; verification depth on specific named-brain identity + technical architecture cap-flagged.

The cap-flag discipline matters editorially because the captive-brain claim space attracts maker-stated framings at marketing depth that don't reliably translate to primary-source-verified architecture claims. Editorially honest cap-flagging surfaces the gap.

---

Third-party foundation-model providers: the brain-providers cluster

The brain-providers cluster operates predominantly third-party-foundation-model positioning. Per the brain-providers cluster framework, the foundation-model tier specialization framework reads each provider at structurally distinct cohort-tier positions:

Physical Intelligence anchors the foundation-model-for-robotics tier. $5.6B confirmed valuation (NOT $11B; $10B/$38B belong to Project Prometheus separate Bezos lab per Agent A correction). Model lineage: pi0 + pi05_base open under Apache-2.0 + π0.6 + π0.7 closed. Deployment maturity: lab + limited-customer-pilots. Third-party integration partner across humanoid + adjacent platform integration.

NVIDIA GR00T anchors the foundation-model-for-physical-AI-general tier. License evolution: N1 noncommercial NVIDIA OneWay → N1.7 commercial-open. CRITICAL CAP-FLAG: NO verified production GR00T deployment exists anywhere as of mid-2026; aggregator narratives of "GR00T at scale" rejected per Agent A source-depth hygiene. Partner matrix verified: only Boston Dynamics + Unitree + Sharpa research-tier confirmed; Apptronik aggregator-inflated (primary brain is Gemini); Figure is competitor not deployment.

Wayve anchors the foundation-model-for-driving tier. $8.6B Series D February 2026 (largest UK AI Series D on record). Model lineage: AI Driver (production end-to-end driving model) + GAIA-3 (generative world model for simulation). Production deployment via Stellantis STLA AutoDrive partnership: commercial-tier deployment integration (rare in foundation-model brain-provider cohort).

Covariant RFM-1 anchors the foundation-model-for-robotic-manipulation tier. CRITICAL DISTINCTION: corporate-state vs model-state separation. Corporate-state: founders Pieter Abbeel + Peter Chen + Rocky Duan + ~25% staff Amazon-acqui-hired August 2024 into Amazon AGI team; Covariant continues as IP/legal entity. Model-state: RFM-1 has NO verified production deployment; verified-deployment Covariant model is legacy Brain via KNAPP warehouse-automation partnership, NOT RFM-1.

Dyna Robotics DYNA-1 anchors the foundation-model-for-dexterous-manipulation tier. Founders Lindon Gao + York Yang + Jason Ma per Agent A primary-source verification (NOT ex-Cobalt/Anthropic/Tesla framing aggregator coverage suggests). Pilot-not-scale deployment posture; service-industry settings (hotels + restaurants + laundromats + gyms) NOT factories. Closed weights + closed code + no published paper.

Skild AI operates integration-partner-forward positioning. Foundation-model-for-robotics adjacent; integration partners forward across humanoid + adjacent cohort customers.

Each third-party foundation-model provider operates structurally distinct cohort-tier specialization and structurally distinct deployment-maturity posture. The cluster-framework distinction matters editorially: trade-press coverage flattening foundation-model brain providers into uniform "AI for robotics" framing operates outside the verified cluster-tier specialization spectrum.

---

Architectural significance: counterparty risk

The captive-vs-third-party gradient carries architectural-significance implications for the verification-posture discipline.

Captive-brain counterparty risk: the maker controls AI substrate + hardware embodiment under single corporate-state ownership. Failure modes (corporate-state distress; pivot away from brain-provider scope; acquisition by entity with different strategic priorities) collapse the AI substrate availability at the hardware-platform level. The hardware platform's continued usefulness depends on the maker's continued investment + commitment to the captive brain.

Third-party-foundation-model counterparty risk: the hardware-platform maker depends on the brain-provider's continued integration commitment + licensing terms + version-upgrade compatibility. Failure modes (brain-provider corporate-state shifts; licensing-term changes; integration-partnership renegotiation) introduce coordination overhead at the hardware-platform layer. But the third-party-foundation-model positioning enables hardware-platform makers to switch brain providers if the integration dependency becomes untenable.

The Covariant example operates as canonical worked example of corporate-state counterparty risk: Covariant founders + ~25% staff Amazon-acqui-hired August 2024 to Amazon AGI team; Covariant continues as IP/legal entity but the founder-cohort capacity moved to Amazon. The legacy Covariant Brain integration via KNAPP retains verified-commercial-deployment positioning; RFM-1 development direction post-acquisition operates at lower verification posture pending primary-source confirmation.

---

Architectural significance: cross-device interoperability

The captive-vs-third-party gradient carries cross-device-interoperability implications.

Captive-brain cross-device interoperability: the captive brain is tied to the maker's hardware embodiment. Third-party hardware platforms don't integrate the captive brain (by definition; that's what makes it captive). Cross-device interoperability across the captive-brain ecosystem requires the maker to expand into adjacent hardware platforms or to expose APIs at the brain-substrate level.

Third-party-foundation-model cross-device interoperability: the foundation-model brain is designed for cross-platform integration; multiple hardware makers can integrate the same brain provider. NVIDIA GR00T positioning targets cross-hardware-platform integration explicitly; Physical Intelligence pi0 + pi05_base Apache-2.0 open positioning enables cross-platform research integration; Skild AI integration-partner-forward positioning frames cross-customer integration as the commercial mechanism.

The cross-device interoperability gradient surfaces editorially when consumers and enterprise customers evaluate platform-lock-in posture. Captive-brain positioning locks consumers into the maker's full-stack ecosystem; third-party-foundation-model positioning lets the hardware platform integrate alternative brain providers if the original integration becomes untenable.

---

Architectural significance: competitive dynamics

The captive-vs-third-party gradient shapes competitive dynamics at the cohort-architecture layer.

Captive-brain competitive dynamics: makers compete on full-stack differentiation across AI + hardware + commercial model. The competitive moat is the integrated stack; competitors must replicate AI + hardware + commercial model to compete head-to-head. Switching costs for the consumer are high (locked into the full stack); switching costs for the maker are also high (vertical-integration investment is hardware-platform-specific).

Third-party-foundation-model competitive dynamics: foundation-model brain providers compete with each other for integration partnerships across hardware-platform makers; hardware-platform makers compete with each other on hardware + integration choice + product-tier framing. The competitive landscape stays more open at the brain-provider integration layer; new entrants can integrate existing brain providers without rebuilding the AI substrate.

The brain-providers cluster framework has documented the competitive-dynamics editorial significance across the third-party-foundation-model cohort: Physical Intelligence vs NVIDIA GR00T vs Wayve vs Covariant vs Dyna vs Skild AI operate at structurally distinct foundation-model tier specializations and structurally distinct integration-partnership positioning. The competitive dynamics at the brain-provider layer + the competitive dynamics at the hardware-platform integration layer compound editorially.

---

Hybrid and license-model gradient positions

Not every brain-provider integration falls cleanly into captive or third-party positioning. Several gradient positions occupy hybrid + license-model territory:

Hybrid licensed positions: some hardware-platform makers operate captive-brain primary substrate plus licensed-secondary integration. Apptronik's captive-brain positioning at the framework layer plus Gemini-based primary-brain integration per Agent A correction (Apptronik primary brain is Gemini, NOT NVIDIA GR00T) operates as worked example of hybrid integration depth.

Open-license third-party positions: Physical Intelligence pi0 + pi05_base Apache-2.0 open licensing operates at structurally distinct positioning from closed-source third-party-foundation-model integration. Open-source positioning enables hardware-platform makers to integrate without licensing-fee dependency but introduces shared upstream maintenance dependencies.

Research-only positions: brain providers operating exclusively at research-tier integration without commercial-deployment partnerships occupy a structurally distinct position at the brain-providers cluster framework. NVIDIA GR00T research-tier partners (Boston Dynamics + Unitree + Sharpa) operate at this gradient position; commercial-deployment translation operates at lower verification posture pending primary-source confirmation.

The gradient positions matter editorially because the captive-vs-third-party framework operates as the foundational distinction but the hybrid + license-model + research-only positions surface meaningful editorial distinctions at the cohort-architecture layer. Per cap-flag-as-trust-signal, the gradient positioning gets surfaced honestly at the verification depth where it operates.

---

Framework operating at the brain-provider integration layer

The captive-vs-third-party gradient demonstrates the framework operating recursively at the brain-provider integration layer. Per DEPLOY's how-deploy-verifies methodology editorial, the framework operates at form-factor-cell granularity across cohorts; the brain-provider integration layer is a form-factor-cell within the broader brain-providers cluster + cross-cluster humanoid platform integration positioning.

The same verification-posture discipline applies at the brain-provider integration layer:

• Verified: brain entity exists at registry surface; captive-vs-third-party positioning anchored at primary-source-verified depth (1X Redwood captive at /brains/1x-redwood; PI third-party at /companies/physical-intelligence + multi-model lineage at /brains/pi0 + /brains/pi05; NVIDIA GR00T third-party at /brains/gr00t-n1 with license-evolution model lineage).
• Stated: maker-stated positioning at primary-source verification depth (Figure Helix captive per Figure communications + competitor framing in NVIDIA GR00T partner-matrix exclusion).
• Claimed: maker-stated captive-vs-third-party positioning asserted without independent primary-source verification at architecture-depth granularity (specific captive-brain architecture claims at Figure Helix; Tesla FSD-derived stack on Optimus; Sanctuary cognitive-architecture on Phoenix).
• Absence: brain-provider integration relationship would matter editorially but isn't verifiable at primary-source depth (specific Apptronik primary-brain integration depth beyond Gemini-NOT-GR00T correction; specific cross-Optimus FSD architecture mapping).

The framework reads the captive-vs-third-party gradient honestly across the verification depth where each claim operates. Editorial discipline operates at the gradient-position layer just like it operates at the form-factor-cell layer + the cohort-tier layer.

---

Why this gradient matters

The captive-vs-third-party gradient matters editorially because it shapes the verification posture appropriate to brain-provider integration claims across cohorts. Trade-press coverage flattening the gradient into uniform "humanoid AI" or "robotics foundation model" framing misses the structurally distinct counterparty-risk + cross-device-interoperability + competitive-dynamics implications that institutional partners (insurance underwriting + data sharing permissions + cross-device interoperability + standards body validation) evaluate at primary-source-verified depth.

The framework operating at the brain-provider integration layer + recursively at the captive-vs-third-party gradient + recursively again at the hybrid + license-model + research-only gradient positions demonstrates the discipline reproducing at every layer of the cohort architecture. The 1X Redwood captive-brain editorial anchor surfaces the canonical worked example at the consumer-archetype-captive-brain intersection; the foundation-model brain-providers cluster operates the third-party-foundation-model spectrum across cohort-tier specializations.

For the canonical captive-brain editorial anchor, see what is 1X Redwood. For the canonical brain-providers cluster framework + third-party foundation-model spectrum, see the brain-providers cluster. For the consumer-vs-industrial humanoid sub-cohort architecture where the captive-vs-third-party gradient operates at cross-archetype layer, see consumer vs industrial humanoid archetypes. For the methodology editorial canonical reference, see how DEPLOY verifies.