On Wednesday 3 September 2026, Nscale put a London press release on the wire that Figure’s next training cycle will ride Nvidia’s Vera Rubin platform through Nscale’s cloud. The headline numbers are hard to miss: up to 100,000 GPUs, an initial US$3.5 billion compute commitment, an intent to grow past US$6 billion, and first racks targeted for Barstow, Texas, in the second half of 2027. Nscale becomes Figure’s preferred compute provider and a shareholder; the equity check size stays undisclosed. FIRGELLI reads it as a serious power and silicon bet for Helix, not as a substitute for the continuous-torque and wrist-payload rows buyers still do not have.
Brett Adcock’s quote in the release is blunt about the bottleneck: data and compute, not a new joint family. Helix, Figure’s on-robot and training stack, is the product of that bet. Jensen Huang’s line in the same announcement sketches the loop Nvidia wants buyers to remember—train on Vera Rubin through Nscale, validate in Isaac Sim, then run inference on GPUs inside the robot. That is a coherent systems story. It is still a systems story about racks and tokens, not about hip continuous torque after a sixty-minute loco-manipulation shift.
What the deal actually locks, and what it does not
Locked, if the companies execute: a multi-year preferred-compute relationship, a named GPU platform, a named Texas site, and a dollar floor on compute. The PR Newswire and Nscale pages both put Vera Rubin and the H2 2027 Barstow start on the same sentence. AI Insider’s Greg Bock summary matches those figures and adds the supply-chain aside: the parties will “explore” putting humanoids into Nscale’s own logistics. Explore is not a purchase order. Explore is not a published takt time.
Not locked: a public continuous-payload card for the biped you would actually put on a line, a per-joint continuous and peak torque table, hand DoF and tactile configuration as a buyer SKU, thermal derate curves, battery swap logistics under factory duty, or a delivered-unit count by serial range. Our Figure 02 catalog row still sits at 1700 mm, 70 kg, 25 kg payload, 35 DoF, and five-hour runtime claims with enterprise/pilot pricing. Figure 03 sequencing work at BMW Spartanburg is real industrial activity this desk already tracks; it still does not publish the continuous-payload and intervention-rate package a cell designer needs before wiring a deposit.
Compute is not continuous torque
Humanoid buyers keep getting sold the same confusion. More GPUs improve imitation learning, world models, and teleop-to-autonomy transfer. They do not raise gearbox continuous ratings, fix wrist backlash under side load, or cool a shoulder actuator after an hour of tote work. If Helix gets smarter on Vera Rubin, the actuator package still has to survive the duty. FIRGELLI’s bar does not move because the training cluster got larger. Ask for the joint map first. Ask for the wrist payload at full reach second. Ask for case temperature at hip and shoulder under a named duty third.
Actuator families still decide whether a station works. Harmonic-drive hips that overheat under continuous gait, quasidirect-drive knees that lose torque after thermal foldback, and tendon or in-finger hands that stretch under side load are not fixed by a larger training budget. A robot that can imitate a BMW body-shop sequence in sim still has to hold continuous wrist force at full reach on the floor, with a published intervention rate and a service interval on the wrist pack. Those numbers are how you size a cell. A GPU count is how you size a rack hall.
The timeline matters for procurement calendars. Second-half 2027 is not a 2026 factory hire. Anyone writing a 2026 cell plan around this release is writing a model-training story into a hardware delivery schedule. Compare that to what already ran today on this desk: CJ Logistics put two named ROBOTIS AI Worker dual-arm units onto a live Olive Young packing station in Yongin with a simple two-finger gripper, and still withheld takt and thermal. Faraday Future put price tags on Futurist and Master without a usable continuous-torque sheet. SoftBank’s reported 1X talks and XPENG’s Iron funding wave earlier this season followed the same pattern: capital and compute first, datasheet later. Figure’s Nscale pact is bigger money and clearer silicon, and it still sits in the same honesty bracket on the metal.
What a factory or lab buyer should ask next
Ask which Helix generation this cluster is sized for, and whether on-robot inference stays on a discrete Nvidia module or moves with Rubin-class silicon. Ask how much of the US$3.5 billion is reserved capacity versus prepaid tokens, and what happens to priority if Nscale’s other tenants spike. Ask for the Barstow energization date, not the marketing half-year. Ask whether BMW Spartanburg—or any other named plant—will publish continuous wrist payload, intervention rate per shift, and mean time between actuator service under the same Helix stack this cluster will train. Ask whether the Nscale supply-chain “explore” clause will ever name a tote mass, a cycle time, and a humanoid SKU.
FIRGELLI’s read on 3 September is simple. Figure secured a multi-year Vera Rubin path through Nscale at a US$3.5 billion starting compute commitment, with upside past US$6 billion and first racks aimed at Barstow in H2 2027. That is a real constraint removed for Helix training if the GPUs land on time. It does not publish continuous joint torque, wrist payload under reach, thermal derate, or delivered biped volume. Hold Figure to the same bar as every other humanoid on this desk: named duty, measured shift, and a joint map you can size actuators against—not a rack count.