Nscale just locked in $3.5B of compute for humanoids—turning embodied AI into a tenant-class the “neocloud” playbook can underwrite

Nscale and Figure announced a strategic partnership to deploy the NVIDIA Vera Rubin platform for up to 100,000 GPUs. The companies described the economics as an initial $3.5B compute commitment with an intent to scale beyond $6B.

They also specified where and when the compute starts to show up operationally: initial deployment starting in the second half of 2027 in Barstow, Texas. Nscale will become a Figure shareholder and will act as Figure’s preferred compute provider.

Many AI compute deals are sold as “train once, fine-tune forever.” Humanoid robotics flips parts of that logic. Figure frames its robotics platform as compute- and data-bound for training its “Helix” models.

That matters because embodied-AI development tends to create a recurring loop: collecting data from systems, generating labels/simulation outcomes, re-training perception/control models, and repeating as the robot’s real-world capability improves. If the compute arrangement is truly aligned to that loop, it can behave more like an industrial workload (utilization over time) than a one-off training campaign.

In other words, the deal tries to convert robotics R&D into repeatable compute demand—which is exactly the type of utilization predictability infrastructure lenders and contracted-revenue investors care about.

To understand potential impact on Tesla Optimus and Agility, separate three layers: (1) the compute supply side (megawatts, GPU allocation, delivery timing), (2) the workload side (training versus inference versus robotics-control latency requirements), and (3) the tenant mix side (who occupies the capacity and what kind of demand profile they create).

This Nscale–Figure deal is strong on layer (1) and partially strong on layer (2) because it is anchored to a named GPU platform and deployment window. It is weaker on layer (2) for strict robotics “edge latency” specifics: the announcement emphasizes compute provisioning rather than where inference/control runs. That’s the key uncertainty for impacts on Tesla Optimus and Agility.

Even so, the existence of a large embodied-AI compute lock-up tightens the competitive space for who can credibly offer “robot-ready” training throughput, which can influence where other humanoid teams place training and model-iteration workloads.

The logic behind “anchor tenant” deals is straightforward: infrastructure builders want enough contracted demand to justify building out power and compute capacity, while customers want enough reliability to avoid paying “capacity premium” during scaling.

Nscale has already been associated with large contracted compute arrangements, and this new Figure deal extends the same structure into embodied robotics. What’s novel is not the existence of a contract—it’s the direction of pressure it creates: embodied-AI teams can now be compared to other AI tenants using a similar language of utilization and delivery schedules.

If Figure scales toward the stated intent of above $6B, embodied AI could become a capacity-stabilizing workload category—but only if their iterative training truly persists at the contracted magnitude through the ramp into 2027–2028.

This deal’s credibility is reinforced by the broader infrastructure posture Nscale described for its Monarch AI campus in West Virginia. Nscale said its Monarch Compute Campus has a path to up to 1.35GW of AI compute capacity via NVIDIA Vera Rubin and a power runway scalable to over 8GW.

Even without tying every number directly to Figure’s Barstow deployment, it matters for the anchor-tenant question because it shows Nscale is thinking in megawatt-scale terms rather than only in GPU-contract terms.

In the short term (days to quarters), the biggest moving parts are sentiment and the signaling of infrastructure readiness. A $3.5B commitment announced publicly tends to compress perceived “uncertainty” around whether embodied-AI work can secure large-scale compute allocation.

In the long term (1–3 years), the question becomes whether Figure (and other humanoid developers) actually sustain training demand at the committed scale as the Barstow ramp begins in 2H 2027. The most decision-relevant verification point is whether scale intent materializes into additional contracted compute (toward the stated >$6B direction) during the ramp window.

At the same time, impacts on Tesla Optimus and Agility hinge on workload partitioning: how much training goes to large centralized compute versus other paths. Without disclosed topology, any direct “tenant swap” conclusion would be speculative.