Figure AI — Architecture Analysis

AI-First

Figure AI is the highest-valued private humanoid company in the world ( $39B post-money, Series C, September 2025 ^[1]). Founded 2022 in Sunnyvale by Brett Adcock with a team drawn heavily from Boston Dynamics, Tesla, and Apple; ex-OpenAI engineers joined later via the now-ended OpenAI partnership. Where Tesla bets on vertical integration and Boston Dynamics on locomotion R&D, Figure bets on an in-house foundation model (Helix VLA) and an early commercial deployment at BMW Spartanburg.

Body DOF

24

Figure 02 published

Hand DOF

16/hand

Figure 03 reveal

Height

165 cm

5'5"

Weight

60 kg

Figure 02/03

Current Price

$150–200K

Target $50K

Runtime

~5 hr

Stated, not benched

AI-first, hardware-second: Figure's public technical output is dominated by Helix VLA papers and video demos, not actuator specs or supplier announcements. This page reflects that asymmetry: rich on the software story, deliberately honest about hardware blanks.

No public teardown: Unlike Tesla (where Morgan Stanley, China Merchants Securities, and patent filings provide a near-complete picture), Figure has no published BOM analysis, no patent portfolio of comparable depth, and no investor filings from Chinese tier-1 suppliers. Spec gaps below are flagged “Not publicly disclosed — pending teardown” rather than guessed.

Helix architecture:Per Figure's February 2025 whitepaper ^[2], Helix uses a System 1 / System 2 split: a fast (200 Hz) 80M-parameter visuomotor policy on-board, plus a slower (~7 Hz) 7B-parameter VLM that reasons over scene + language. This is the first publicly-announced VLA model with full upper-body continuous control. Trained on Figure's own teleop corpus after the OpenAI partnership ended.

Approach

General-purpose, transformer-based AI, Helix VLA (in-house)

•Helix VLA: in-house vision-language-action model (OpenAI partnership ended)

•Whole-body transformer control

•Commercial deployment focus (BMW partnership)

•$39B valuation (September 2025, Series C)

•Silicon Valley talent density

Component-by-Component Teardown

Each Figure subsystem mapped to the closest commodity equivalent from our verified parts catalog. Where Figure has not disclosed a spec, the row is labelled Not Disclosed rather than filled in with speculation.

Subsystem	Figure Implementation	Public Spec	Commodity Equivalent	Price	Gap
Hip / Shoulder Actuator	Figure proprietary rotary — believed to be brushless DC + harmonic reducer (not publicly confirmed)	Torque/RPM specs not publicly disclosed; Figure 02 demos suggest peak torque comparable to Optimus class (100–180 Nm range)	Harmonic FHA-25C (108 Nm) or EYOU PH20-101 (182 Nm)	$660–$2,400	Not Disclosed
Knee / Leg Actuator	Figure proprietary — believed rotary (no linear-actuator patents on file matching Tesla's planetary roller screw approach)	Not publicly disclosed	T-Motor RMD-X8 (80 Nm) or CubeMars AK10-9 (50 Nm)	$420–$1,650	Not Disclosed
Hand / Finger Actuation	Figure 03 hands — 16 DOF per hand (publicly stated in Oct 2025 reveal). Tendon-driven design publicly described.	16 DOF/hand; tendon routing; tactile fingertip sensing claimed; per-finger DOF not broken down publicly	Dynamixel XL330 array (servo-per-DOF, 32 servos for both hands)	$65/servo (~$2,080 for 32)	Partial Gap
Compute	NVIDIA-based on-board compute (Figure has confirmed NVIDIA as a partner; specific module not publicly named)	Believed Jetson AGX Thor or Orin-class. Helix VLA inference runs on-board. Cloud offload for training only.	NVIDIA Jetson AGX Thor (2,070 FP4 TFLOPS, 130W)	$3,499	Close Match
Battery	Third-party Li-ion pack (vendor not publicly disclosed)	Runtime ~5 hours stated for Figure 02; pack capacity, cell chemistry, and pack designer not in public record	Victron Lithium 48V 50Ah (2.4 kWh)	$3,200	Not Disclosed
Vision	RGB cameras (count not publicly fixed; press images suggest stereo head + chest cameras)	4K-class RGB; no LiDAR mentioned in any public reveal; Helix uses vision tokens directly into the VLM	Intel RealSense D455 + Logitech C920	$400–$800	Partial Gap
Tactile Skin	Fingertip tactile sensing on Figure 03 hands (claimed)	Cell count, modality (capacitive / resistive / barometric), and resolution not publicly disclosed	No drop-in commodity equivalent — research-grade only	Research only ($2K+ per finger)	Not Disclosed
Frame / Structure	Figure-designed structural frame; materials not broken out publicly	Figure 02: 60 kg, 165 cm. Figure 03 weight unchanged. Photographs suggest aluminum + composite, no carbon-fiber claims.	Carbon fiber mid-tier or aluminum frame	$3,500–$22,000	Partial Gap

Hip / Shoulder Actuator

Not Disclosed

Figure

Figure proprietary rotary — believed to be brushless DC + harmonic reducer (not publicly confirmed)

Torque/RPM specs not publicly disclosed; Figure 02 demos suggest peak torque comparable to Optimus class (100–180 Nm range)

Commodity

Harmonic FHA-25C (108 Nm) or EYOU PH20-101 (182 Nm)

$660–$2,400

Figure does not publish actuator specs, supplier names, or torque tables. Press materials describe "electric actuators throughout" and "proprietary joint design." Architecturally Figure 02 appears closer to a conventional rotary stack than to Tesla's mixed rotary+linear approach. No teardown of a production Figure 02 exists in the public record. This is the largest spec gap in the entire stack.

Knee / Leg Actuator

Not Disclosed

Figure

Figure proprietary — believed rotary (no linear-actuator patents on file matching Tesla's planetary roller screw approach)

Not publicly disclosed

Commodity

T-Motor RMD-X8 (80 Nm) or CubeMars AK10-9 (50 Nm)

$420–$1,650

Unlike Tesla's well-documented planetary roller screw linear actuators (8,000 N), Figure has not published a knee architecture. Walking gait videos suggest a conventional rotary knee with reducer. If Figure uses harmonic drives at the knee, this is the most likely Suzhou Green Harmonic / Harmonic Drive Systems sourcing point — but Figure has named neither vendor.

Hand / Finger Actuation

Partial Gap

Figure

Figure 03 hands — 16 DOF per hand (publicly stated in Oct 2025 reveal). Tendon-driven design publicly described.

16 DOF/hand; tendon routing; tactile fingertip sensing claimed; per-finger DOF not broken down publicly

Commodity

Dynamixel XL330 array (servo-per-DOF, 32 servos for both hands)

$65/servo (~$2,080 for 32)

Figure 03 hands are tendon-driven, similar architectural class to Tesla Gen 3 (22 DOF) and Shadow Robot Hand. Figure claims 16 DOF/hand vs Tesla 22 — fewer total DOF, simpler routing. Actuator vendor for the tendon drives is not disclosed. Tactile sensing is claimed but resolution and cell count are not in the public record. This is the most-publicized component of Figure 03 but the spec sheet is still high-level.

Compute

Close Match

Figure

NVIDIA-based on-board compute (Figure has confirmed NVIDIA as a partner; specific module not publicly named)

Believed Jetson AGX Thor or Orin-class. Helix VLA inference runs on-board. Cloud offload for training only.

Commodity

NVIDIA Jetson AGX Thor (2,070 FP4 TFLOPS, 130W)

$3,499

Figure has stated publicly that the robot runs an end-to-end vision-language-action model (Helix) on-board. The Helix whitepaper (Feb 2025) describes a System 1 / System 2 split: a fast 80M-parameter visuomotor policy plus a slower 7B-parameter VLM reasoner. This compute profile fits Jetson Thor. Figure has not confirmed the specific module. NVIDIA is a Series B investor, so Thor is the natural assumption.

Battery

Not Disclosed

Figure

Third-party Li-ion pack (vendor not publicly disclosed)

Runtime ~5 hours stated for Figure 02; pack capacity, cell chemistry, and pack designer not in public record

Commodity

Victron Lithium 48V 50Ah (2.4 kWh)

$3,200

Figure has stated 5-hour runtime targets in press materials. No public confirmation of cell vendor, chemistry, pack architecture, or BMS design. Most likely sourced from a Korean or Chinese tier-1 (LG Energy, Samsung SDI, CATL, BYD) but no announcement on record. Pending teardown.

Vision

Partial Gap

Figure

RGB cameras (count not publicly fixed; press images suggest stereo head + chest cameras)

4K-class RGB; no LiDAR mentioned in any public reveal; Helix uses vision tokens directly into the VLM

Commodity

Intel RealSense D455 + Logitech C920

$400–$800

Figure follows the vision-language-action paradigm pioneered by RT-2 / OpenVLA. Helix consumes camera frames directly into the VLM without an explicit depth/occupancy intermediate (per the Feb 2025 whitepaper). Camera count and sensor model are not on the public record. Architecturally this is closer to Tesla's vision-only doctrine than to Boston Dynamics' LiDAR-equipped Atlas.

Tactile Skin

Not Disclosed

Figure

Fingertip tactile sensing on Figure 03 hands (claimed)

Cell count, modality (capacitive / resistive / barometric), and resolution not publicly disclosed

Commodity

No drop-in commodity equivalent — research-grade only

Research only ($2K+ per finger)

Figure has shown video of fingertip tactile feedback influencing grip force in Figure 03 demos. No technical paper has been published. Most likely candidates: Contactile (Australian capacitive arrays), GelSight (vision-based tactile, MIT spinout), or Figure's own development. Not publicly disclosed.

Frame / Structure

Partial Gap

Figure

Figure-designed structural frame; materials not broken out publicly

Figure 02: 60 kg, 165 cm. Figure 03 weight unchanged. Photographs suggest aluminum + composite, no carbon-fiber claims.

Commodity

Carbon fiber mid-tier or aluminum frame

$3,500–$22,000

No Giga-Press analog. Figure manufactures at BotQ (Sunnyvale, CA) — a stated 12,000 units/year capacity facility. Frame is presumably machined and assembled rather than die-cast at Tesla's scale. Figure has not published a structural decomposition.

Gap Analysis Detail

Hip / Shoulder ActuatorNot Disclosed

Figure does not publish actuator specs, supplier names, or torque tables. Press materials describe "electric actuators throughout" and "proprietary joint design." Architecturally Figure 02 appears closer to a conventional rotary stack than to Tesla's mixed rotary+linear approach. No teardown of a production Figure 02 exists in the public record. This is the largest spec gap in the entire stack.

Knee / Leg ActuatorNot Disclosed

Unlike Tesla's well-documented planetary roller screw linear actuators (8,000 N), Figure has not published a knee architecture. Walking gait videos suggest a conventional rotary knee with reducer. If Figure uses harmonic drives at the knee, this is the most likely Suzhou Green Harmonic / Harmonic Drive Systems sourcing point — but Figure has named neither vendor.

Hand / Finger ActuationPartial Gap

Figure 03 hands are tendon-driven, similar architectural class to Tesla Gen 3 (22 DOF) and Shadow Robot Hand. Figure claims 16 DOF/hand vs Tesla 22 — fewer total DOF, simpler routing. Actuator vendor for the tendon drives is not disclosed. Tactile sensing is claimed but resolution and cell count are not in the public record. This is the most-publicized component of Figure 03 but the spec sheet is still high-level.

ComputeClose Match

Figure has stated publicly that the robot runs an end-to-end vision-language-action model (Helix) on-board. The Helix whitepaper (Feb 2025) describes a System 1 / System 2 split: a fast 80M-parameter visuomotor policy plus a slower 7B-parameter VLM reasoner. This compute profile fits Jetson Thor. Figure has not confirmed the specific module. NVIDIA is a Series B investor, so Thor is the natural assumption.

BatteryNot Disclosed

Figure has stated 5-hour runtime targets in press materials. No public confirmation of cell vendor, chemistry, pack architecture, or BMS design. Most likely sourced from a Korean or Chinese tier-1 (LG Energy, Samsung SDI, CATL, BYD) but no announcement on record. Pending teardown.

VisionPartial Gap

Figure follows the vision-language-action paradigm pioneered by RT-2 / OpenVLA. Helix consumes camera frames directly into the VLM without an explicit depth/occupancy intermediate (per the Feb 2025 whitepaper). Camera count and sensor model are not on the public record. Architecturally this is closer to Tesla's vision-only doctrine than to Boston Dynamics' LiDAR-equipped Atlas.

Tactile SkinNot Disclosed

Figure has shown video of fingertip tactile feedback influencing grip force in Figure 03 demos. No technical paper has been published. Most likely candidates: Contactile (Australian capacitive arrays), GelSight (vision-based tactile, MIT spinout), or Figure's own development. Not publicly disclosed.

Frame / StructurePartial Gap

No Giga-Press analog. Figure manufactures at BotQ (Sunnyvale, CA) — a stated 12,000 units/year capacity facility. Frame is presumably machined and assembled rather than die-cast at Tesla's scale. Figure has not published a structural decomposition.

What Would It Cost to Build a Figure-Equivalent?

Running the closest commodity BOM through the HUMA Build Compiler — targeting 165 cm, 60 kg, 20 kg payload, 5 hr runtime, pro compute, 56 total DOF (24 body + 32 hand).

DOES NOT COMPILE

6 errors6 warnings

Total Cost

$38,565

Figure Current Price

$150–200K

Total Mass

37.2 kg

Peak Power

18,704.45 W

A commodity Figure-equivalent costs $38,565 vs Figure's current customer price of $150,000–$200,000 (target: $50,000 at volume). Commodity hardware now beats Figure's long-run target on price alone. The defensible moat is Helix, not the chassis.

✗

Electrical

Main-bus voltage mismatch: battery is 48V but 1 main-bus actuators expect different voltage

Dynamixel XM430-W350-T expects 12V

✓

Electrical

Power budget OK: 3553W effective peak vs 7200W battery

Actuator peak 11359W × 30% + 145W non-actuator

!

Communication

Multiple actuator protocols: EtherCAT, CAN, UART — requires a multi-protocol gateway

✓

Communication

Compute (NVIDIA Jetson AGX Thor (T5000 Module)) bridges to actuator bus via motor controller

!

Communication

EtherCAT actuators detected with standard Linux compute. EtherCAT requires a real-time kernel (PREEMPT_RT patch). Stock Jetson Ubuntu will have jitter issues without RT patches. Consider: (a) apply PREEMPT_RT to Jetson, (b) use a dedicated EtherCAT master (Beckhoff), or (c) switch actuators to CAN protocol.

✗

Mechanical

Missing hip or knee actuators

✓

Mass Budget

Mass budget OK: 39.2 kg / 60 kg (65%)

✗

Software

Compute too weak for pro tier: 0 TOPS < 200 TOPS required

Full VLA models (Helix, GR00T N1.5, π0.5)

!

Thermal

Thermal budget tight: 802W simultaneous heat vs 30W passive (27× — active cooling required)

Harmonic Drive FHA-25C ×4: 324W each (η=73%); T-Motor RMD-X8 Gimbal Motor ×2: 302W each (η=82%); CubeMars AK10-9 QDD Actuator ×2: 230W each (η=80%); EYOU PH11-101 Harmonic Actuator ×2: 72W each (η=70%)

!

Thermal

Low-efficiency actuator: Dynamixel XH540-W270-T at 60% — monitor joint temperature in continuous use

!

Runtime

Runtime marginal: 3.4h estimated vs 5h target

Battery 2400 Wh ÷ avg 713W (actuators 568W at 5% duty + 145W always-on). Power management or reduced duty may be needed.

!

DOF

BOM has 46 actuators for 56 target DOF — some joints may use multi-axis actuators or passive compliance

✗

infrastructure

No DC-DC converter or PSU producing the 5 V rail (load 23 W)

Loads on this rail: Dynamixel XL330-M288-T ×32, Intel RealSense D455 Depth Camera ×2, Logitech C920 HD Webcam ×2, Xsens MTi-3 AHRS ×1, ATI Mini40 F/T Sensor ×2. Add a 48→5 V converter (e.g. Pololu D36V50F-series, Mean Well SD-100C, Murata UWE/OKL).

✗

infrastructure

No DC-DC converter or PSU producing the 12 V rail (load 133 W)

Loads on this rail: Dynamixel XM430-W350-T ×2, NVIDIA Jetson AGX Thor (T5000 Module) ×1. Add a 48→12 V converter (e.g. Pololu D36V50F-series, Mean Well SD-100C, Murata UWE/OKL).

✗

infrastructure

No DC-DC converter or PSU producing the 24 V rail (load 10 W)

Loads on this rail: Dynamixel XH540-W270-T ×2. Add a 48→24 V converter (e.g. Pololu D36V50F-series, Mean Well SD-100C, Murata UWE/OKL).

i

Availability

All 14 parts have supplier URLs — stock & lead-time NOT verified

In-House vs Partner-Led

Where Tesla is vertically integrated by default, Figure's architecture is software-vertical, hardware-partnered: Figure owns the AI brain end-to-end and assembles the robot in-house at BotQ, but pulls compute from NVIDIA, cloud from Microsoft, and most component vendors remain undisclosed.

Figure Owns / Makes In-House

Helix VLA (vision-language-action model)

First publicly-announced VLA with full upper-body continuous control. System 1 / System 2 split: 80M-param visuomotor policy + 7B-param VLM reasoner. Trained in-house at Figure after the OpenAI partnership ended in early 2025. This is now Figure's primary technical moat.

Robot mechanical design + system integration

Figure 02 / Figure 03 industrial design, kinematics, electronics integration done in-house in Sunnyvale. Brett Adcock has stated the team designs every joint and PCB.

BotQ manufacturing facility

Figure-operated facility in Sunnyvale, CA. Stated capacity 12,000 units/year. Final assembly done by Figure rather than contract manufacturer. No CM partnership has been announced.

Tendon-driven hand mechanics (Figure 03)

Figure 03 hand redesign reduced from earlier servo-per-joint approach to tendon-driven 16 DOF/hand. Mechanical design Figure-internal; actuator + tendon supplier not disclosed.

Robot data collection + training pipeline

Figure operates teleoperation rigs at BMW Spartanburg and internal labs. Helix is trained on Figure's own demonstration corpus. Dataset size and composition not publicly disclosed.

Partners / Confirmed Relationships

NVIDIA — compute + simulation

NVIDIA Capital led portions of Series B (2024). On-board compute is NVIDIA Jetson-class (Thor or Orin). Isaac Sim / Isaac Lab almost certainly used for sim-to-real, though Figure has not published its simulation stack.

Microsoft — cloud + Azure OpenAI

Microsoft co-led the $675M Series B (Feb 2024). Training compute via Azure. Microsoft holds an indirect stake through OpenAI as well.

OpenAI — former model partner

OpenAI Startup Fund participated in Series B. Figure used GPT-4V for high-level reasoning during the partnership (announced Mar 2024). Partnership ENDED early 2025 — Figure brought reasoning in-house with Helix. Public split has been amicable but the IP reversion terms have not been disclosed.

BMW — first commercial customer

Figure 02 deployed at BMW Spartanburg, SC (announced Jan 2024). Insertion of sheet-metal parts into fixtures. Figure 03 in structured pilot at the same site (2025–2026). BMW is also a strategic investor. Figure has stated BMW deployment is 4× faster and 7× more accurate than earlier Figure generations.

Tier-1 strategic investors (LG, Salesforce, T-Mobile, Qualcomm, Intel)

Series B and C captable includes corporate strategics across electronics, telecom, and chips. None has been publicly tied to a specific component sourcing relationship — these read as financial + future-distribution bets, not BOM commitments.

Parkway VC — financial lead

Parkway VC led the Series C ($1.5B+, Sep 2025) at $39B post. Largest single financial investor by check size.

Neumann's Replicability Lens

~70%

P% (Producibility)

BotQ assembles at 12K/yr stated capacity; component manufacture is outsourced

~60%

M% (Manufacturability)

Assembly-grade — no Giga-Press analog, no die casting, hand-built structural elements

~15%

C% (Commodity Replicability)

Hardware is more commodity-replicable than Tesla. Helix VLA is the defensible asset, not the chassis.

Funding & Captable

Figure has raised $2.5B+ cumulative across three rounds, reaching a $39B post-money valuation in September 2025 ^[1]. The captable concentrates strategic capital from chips (NVIDIA, Intel, Qualcomm), cloud (Microsoft), telecom (T-Mobile), software (Salesforce), consumer electronics (LG), and the commercial customer itself (BMW).

2023Series A

$70M

Lead: Parkway

2024Series B

$675M

Lead: Microsoft, OpenAI Startup Fund, NVIDIA @ $2.6B

2025Series C

$1.5B+

Lead: Parkway VC @ $39B

Why $39B matters: At this valuation Figure trades at roughly $260M of paper-equity per shipped robot (using the ~150 units/2024 figure). The company is priced on the bet that Helix becomes the operating system of every humanoid, not on the present hardware. This is the diametric opposite of Tesla's vertical-integration thesis — and explains why Figure's public technical output emphasizes AI papers over BOM teardowns.

Figure vs Tesla Optimus — Side by Side

The two companies sit on opposite axes of the same market. Tesla owns the hardware floor; Figure is racing to own the software ceiling.

Dimension	Tesla Optimus	Figure 02 / 03
Strategy	Vertical integration: design + build + train + deploy	Partnership-led hardware, in-house AI
Hand DOF	22/hand (Gen 3, tendon-driven, 23 forearm linears)	16/hand (Figure 03, tendon-driven)
Locomotion actuators	Mixed rotary + linear (planetary roller screw, 8,000 N)	Believed all-rotary; specs not disclosed
Compute	Custom AI4 / AI5 chip (Tesla-designed, Samsung/TSMC fab)	NVIDIA Jetson Thor-class (commodity)
Foundation model	End-to-end NN, 48 co-trained sub-networks, FSD-derived	Helix VLA (System 1 / System 2 split)
Training data moat	8.2B vehicle miles + Cortex (~67K H100s)	Teleop corpus (size undisclosed) + Microsoft Azure compute
Manufacturing	Giga Press, Fremont line (Model S/X conversion)	BotQ Sunnyvale, 12K units/yr stated capacity
First customer	Internal (Tesla factories) — external pilots 2026	BMW Spartanburg (live, Jan 2024 onward)
Valuation signal	Part of $800B+ Tesla parent	$39B private, AI-style multiple
Long-run unit price	<$20K target	$50K target (current $150–200K)

Where Tesla is ahead

•Hardware vertical integration (no commodity dependency)
•Hand DOF (22 vs 16) and forearm linear-actuator architecture
•Pre-existing FSD vision moat (8.2B miles transferred to robot)
•Manufacturing scale (Giga Press, Fremont conversion)
•Long-run unit cost target (<$20K vs $50K)
•Custom silicon (AI5 vs commodity Jetson)

Where Figure is ahead

•External commercial deployment (BMW live; Tesla still internal)
•First publicly-announced VLA with full upper-body continuous control
•AI talent density from OpenAI / DeepMind alumni network
•Investor diversity (BMW, NVIDIA, Microsoft, OpenAI, LG, Salesforce)
•Software releases cadence (Helix paper, video demos every quarter)
•Lower hardware lock-in — easier to swap commodity components

Open Questions for the Portal

Items where Figure has not made the public record. We surface these here so visitors know what we know we don't know — and so the community can fill in gaps as teardowns appear.

Q1.

Who supplies the rotary actuators?

Figure has not named a harmonic-drive partner, motor vendor, or contract assembler for joints. Candidates by elimination: Harmonic Drive Systems (Japan), Suzhou Green Harmonic (China — also Tesla supplier), Sumitomo, or fully Figure-internal. Pending teardown.

Q2.

What battery cells and pack designer?

Pack supplier and cell chemistry not on record. Likely candidates: LG Energy (Series C strategic), Samsung SDI, CATL, BYD, or Northvolt. 5 hr runtime at humanoid power profile (~500 W mixed) implies ~2.5–3 kWh — consistent with a 48–52 V Li-ion pack.

Q3.

Which Jetson module powers Helix on-board?

Figure has confirmed NVIDIA compute but not the SKU. Helix System 1 + System 2 split (80M policy + 7B VLM) fits Jetson AGX Thor. Could also be a multi-Orin configuration. Has implications for power budget and thermal headroom.

Q4.

Tactile sensor cell count and modality?

Figure 03 fingertips show tactile feedback in demos but no spec sheet exists. Candidates: Contactile capacitive arrays, GelSight vision-based, custom barometric. Tactile resolution drives manipulation envelope.

Q5.

Helix training corpus size?

Number of teleop demonstrations, hours of robot operation, and synthetic augmentation factor are all undisclosed. Tesla publishes 10,000× synthetic-per-real; Figure has published nothing comparable.

Q6.

Contract manufacturer for high-volume scale-up?

BotQ stated capacity is 12K/yr. Beyond that, Figure has not announced a Foxconn / BYD / Flex-class CM. Required for the 100K+ unit ambition implied by the $39B valuation.

Q7.

OpenAI partnership reversion terms?

Partnership ended early 2025 but the IP, training-data, and equity arrangement at termination is not public. OpenAI Startup Fund retains a Series B stake. Helix is described as Figure-built.

Q8.

Figure 03 weight, exact DOF count, peak torque?

Figure 03 reveal (Oct 2025) showed redesigned hands (16 DOF) but did not republish a full body spec table. Whether Figure 03 changes body DOF from Figure 02's 24 is unclear.

Why This Matters for Kit Builders

Figure validates AI-first

A $39B valuation on ~150 shipped robots tells you the market is priced on which company owns the foundation model, not which ships the most chassis. Kit-builders should treat Helix-class VLA models as the core deliverable and the hardware as commodity packaging.

BMW de-risks deployment

Figure on a BMW Spartanburg line is a public, ongoing existence proof that humanoids can do useful, repetitive work in a structured factory. That validates the whole category — and gives kit-builders a benchmark for what “good enough” means in the first commercial niche.

The hardware blanks are the opportunity

Figure's hardware is far more commodity-replicable than Tesla's. NVIDIA Jetson, off-the-shelf rotary actuators, third-party Li-ion: every layer except the brain is substitutable. Open-stack kits compete more directly with Figure 02 than with Optimus Gen 3.

Sources & Footnotes

Figure AI Series C announcement, September 2025 — $1.5B+ raised at $39B post-money, led by Parkway VC. Captured in companies.ts:figure-ai.funding and graph.ts:figure.
Figure AI “Helix: A Vision-Language-Action Model for Generalist Humanoid Control” whitepaper, February 2025. System 1 / System 2 architecture, 80M policy + 7B VLM. graph.ts:helix-vla.
BMW Group press release, January 2024 — Figure 02 deployment at Spartanburg, SC. Subsequent updates from Brett Adcock founder communications throughout 2024–2025 reporting 4× speed and 7× accuracy improvements over earlier generations.
Series B announcement, February 2024 — $675M led by Microsoft, OpenAI Startup Fund, NVIDIA, with participation from Intel, LG, Samsung, Jeff Bezos, and others. companies.ts:figure-ai.funding[1].
Component-level spec gaps reflect public-record absence as of the Figure 03 reveal (October 2025) and the Series C announcement (September 2025). No production-unit teardown has been published by Munro & Associates, Morgan Stanley, China Merchants Securities, or equivalent third parties at time of writing.