Physix Frontier
Robotics firms buying compute mirrors structural biology's GPU purchases
Viewing the $3.5 billion compute agreement between Figure and Nscale as robot news is easy; viewing it as infrastructure news is where it gets interesting. It's like buying servers in a lab: on the surface, it's chip counts; underneath, it's power, cooling, data feedback loops, and validation cycles.
I've been trying out AI inference nodes recently. Previously running AlphaFold3 on GPU clusters gave me a direct feeling: getting a model to run and producing stable experimental results are two different things. Figure plans to deploy up to 100,000 NVIDIA Vera Rubin chips, and Nscale externally states high contract revenue with an average term of 5.7 years. This structure resembles cloud vendors signing long-term contracts, packaging GPUs, data centers, power, and networks into predictable cash flows. The ~$45 billion agreement between Anthropic and Nscale is similar, leasing about 460 MW and locking in six years. Compute is starting to be priced like electricity.
But robots aren't proteins. AlphaFold2 proved in Nature 2021 that multiple sequence alignments and structural modules can push sequences to high-precision structures; AlphaFold3 expanded to complexes in Nature 2024. Its strong performance on biological data relies not just on compute, but on PDB, cryo-EM, MSA, physical constraints, and wet-lab validation. The gap between wet and dry experiments can't be filled by just adding more cards. Figure is buying compute for training and inference budgets for Vision-Language-Action models, but real fault tolerance comes from grasp failures, lighting changes, ground friction, battery decay, and field network jitter. Watching scenario competitions before, I felt similarly: sprints look good, but the floor of scenario competitions is what's valuable.
Two routes emerge here. Nscale builds AI cloud infrastructure, binding long-term contracts to large model and robotics companies, earning scale effects from capacity, power, and capex. Figure builds an embodied intelligence closed loop, outsourcing compute but keeping data, action policies, and deployment in-house. The former is like a public GPU cluster: elastic and general-purpose; the latter is like a specialized protein prediction pipeline: narrow tasks, but every link needs tuning. Performance on biological data often depends on data formats, parallelization, caching, and evaluation; robots are the same. 100,000 chips won't all sit on the robot body; most are for training, simulation, and batch inference. The body itself needs low latency, small models, and stable execution.
Papers like RT-2 transfer vision-language models to action tokens, and PaLM-E tries connecting large models to robot states. They show models can unify representations, but action environments are messier than protein structures. Protein prediction has relatively standard inputs/outputs; robots face unstructured worlds. Thus, the larger the training compute, the greater the need for high-fidelity simulation, teleoperation data, and field evaluation. Otherwise, you're just scaling errors.
I care more about the quality behind the contract numbers. The revenue sounds huge, and the 5.7-year average provides a cash flow narrative, but the core risk of AI clouds isn't whether there are GPUs, but whether power lands, liquid cooling delivers, and chip generations rapidly devalue old capacity. I just touched on converting mining racks to inference yesterday; cheap hardware doesn't mean cheap operations. Maintenance, heat dissipation, and stability are the barriers. If Figure's money buys long-term stable robot work, it's not buying cards, but a data feedback pipeline. If not, it might just delay trial-and-error costs to the next IPO.
When those 100,000 Vera Rubins light up, whether robot companies are closer to a verifiable prediction engine like AlphaFold, or closer to an expensive simulation server room, is a question worth watching more than the fundraising amount.
📌 This article is compiled from Bloomberg Tech, original source: https://www.bloomberg.com/news/articles/2026-09-03/nscale-backs-robot-firm-figure-alongside-3-5-billion-cloud-deal
All rights reserved. This is a compilation and independent analysis based on public reports.