Multiply Labs Advances AI for Biologics Robotics on Anyscale

Multiply Labs builds robotic biomanufacturing systems designed to help the biologics industry manufacture precision medicines at scale. Their modular robotic clusters reduce manufacturing bottlenecks and minimize human contamination while automating industry-standard biomanufacturing processes. 

As Multiply Labs sought to eliminate biological contamination, most recently in vulnerable loading and unloading steps outside the robotics cluster, the team expanded its R&D to build physical AI systems capable of handling more complex scenarios and edge cases. This required investing in capturing and processing egocentric video datasets to power a range of approaches from imitation learning to emerging Vision-Language-Action (VLA) model architectures. 

To support the growing demand for multimodal data pipelines and distributed training workloads, the team needed access to premium GPUs such as A100s and H100s, which are often difficult to secure on-demand within a single region or cloud provider.

With Anyscale, Multiply Labs standardized distributed training pipelines that are:

• Portable across multiple cloud providers

• Deployed on-demand on premium GPUs

• Securely deployed within their own environment to protect sensitive manufacturing video data

LinkChallenges

As Multiply Labs built out new robotics R&D initiatives, the team needed a platform that could maintain strong data controls all while improving R&D development velocity on large scale training. To accomplish this, they needed to address three major challenges:

• Distributed training environments were difficult to standardize for researchers: Standing up and operating environments for distributed training runs, and having the necessary logging to troubleshoot required significant manual effort. This slowed R&D iteration and made it difficult for more engineers to run multiple training jobs in parallel. 
  

• Strict data controls were non-negotiable: Multiply Labs operates in highly IP-sensitive biomanufacturing environments. Video data from pharmaceutical manufacturing processes is highly confidential and must remain tightly controlled within their already approved cloud environments.
  

• Limited, unpredictable premium GPU supply: On-demand access to GPUs, specially NVIDIA H100 and A100, varied by region and provider but was critical to Multiply Labs’ growing research initiatives. As robotics R&D expanded, GPU availability constrained experimentation velocity. 

"We had these new robotic behaviors we needed to onboard onto our product to meet rising customer demand, but the AI infrastructure that got us here couldn’t scale to handle the more complex models we needed for the next phase."

Justin Shim | Advanced Robotics Software Engineer

LinkThe Solution

By partnering with Anyscale, Multiply Labs adopted a managed platform to reduce operational burden, maintain strict data controls, and deploy training workloads wherever GPU availability existed, without worrying about customizations per cloud provider.

With Anyscale, Multiply Labs is able to: 

• Improve developer experience and velocity for distributed training: Anyscale enables the team to run large-scale training on managed Ray clusters with a centralized developer workspace and job observability – allowing engineers to quickly build and debug experiments without managing clusters. 

• Maintain strong security and data controls with a Bring Your Own Cloud (BYOC) model: Multiply Labs deploys Anyscale within their own cloud accounts, keeping sensitive customer workflows and data tightly controlled while benefitting from managed orchestration and developer tools.

• Access premium GPUs with multi-cloud flexibility: With Anyscale, Multiply Labs engineers standardized infrastructure-as-code workflows to deploy Ray clusters through a single API surface across any cloud, enabling researchers to quickly run workloads wherever premium GPU capacity is available, including AWS, GCP, and Nebius.

"Using Anyscale as our multi-cloud Ray cluster manager, we’ve made our training runs fully portable across hyperscalers and specialized cloud providers, maximizing GPU availability. "

Varun Bhatia | Staff Robotics Software Engineer

LinkFrom cloud management to repeatable distributed training

Multiply Labs wanted distributed training to scale without constantly re-configuring their clouds. Ray provides a consistent distributed compute runtime that could not only scale training for multiple frameworks but also support scalable processing for other parts of the pipeline such as data processing. With Anyscale-managed Ray clusters, developers don’t have to handle cloud cluster bring-up, high-availability configuration, or any other cloud cluster lifecycle management — allowing the team to focus on building and iterating on the training workflow itself.

To further standardize operations, the team implemented an infrastructure-as-code (IaC) approach and worked toward a one-click style deployment experience. Standing up the environment became a simple, repeatable process rather than a manual sequence of setup steps. This reduced reliance on manual runbooks and made it easier to onboard engineers as R&D efforts expanded.

Centralized UI and observability complemented this shift. Engineers could monitor job lifecycle, inspect logs, and troubleshoot distributed runs in one place as they moved from small experiments to larger-scale training. As a result, ramp time improved for team members new to Ray, and iteration cycles accelerated across the organization. 

"By offloading Ray cluster operations to Anyscale, our engineers can concentrate on pushing the boundaries of physical AI rather than managing distributed compute infrastructure."

Varun Bhatia | Staff Robotics Software Engineer

LinkStrict data residency controls for AI using proprietary IP 

Multiply Labs works closely with pharmaceutical manufacturers, where process information is highly sensitive and tightly guarded. Maintaining strict controls over the data and the environments where the data was processed through training workflows was essential. 

Using Anyscale’s Bring Your Own Cloud (BYOC) deployment model, Multiply Labs runs distributed workloads directly within their own cloud accounts. This ensures that sensitive customer IP such as video data remains fully under their virtual private cloud, while still benefitting from managed orchestration, scaling, and observability. 

This architecture allows Multiply Labs to scale advanced robotics training without compromising data residency requirements or customer trust. 

"Our customers are extremely sensitive about their data. Anyscale lets us run distributed training in our own cloud environment, so we can keep our customers’ data in our own hands."

Justin Shim | dvanced Robotics Software Engineer

LinkMulti-cloud flexibility when premium GPU supply is constrained

Premium GPU supply, particularly for A100s and H100s, became a practical bottleneck as robotics R&D expanded at Multiply Labs. The team initially operated in a single AWS region, then expanded across additional regions and later added both GCP and Nebius to improve GPU access.

A key enabler of this expansion was Anyscale’s abstraction layer for deploying Ray clusters across any cloud. Using this single API surface, Multiply Labs was able to streamline automation through infrastructure-as-code. This eliminated manual runbooks and created a repeatable process for bringing new GPU environments online. As a result, cluster spin-up with H100s on Nebius used for training was reduced from 30 to 10 minutes with similar results for other GPU types across cloud providers.

By combining standardized deployment automation with Anyscale’s multi-cloud job orchestration, Multiply Labs can flex training workloads across providers as GPU supply shifts, enabling faster iteration on more advanced AI for robotics.

"GPU availability in the cloud is quite scarce and a primary bottleneck for us, as we needed access to premium GPUs like H100s on-demand. The ability to implement a multi-cloud solution has been instrumental in accelerating our AI development."

Varun Bhatia | Staff Robotics Software Engineer

LinkWhat’s Next for Multiply Labs

Multiply Labs is continuing to invest in robotics R&D to expand the adaptability and intelligence of its biomanufacturing systems. As workflows become more complex, the team is focused on enabling robots to rapidly learn new tasks, handle edge cases more robustly, and further reduce human intervention in sensitive manufacturing processes. 

This requires exploring more advanced learning architectures, requiring sustained access to scalable, distributed GPU training. With a standardized, multi-cloud foundation in place, Multiply Labs is positioned to iterate faster and accelerate the development of increasingly capable robotic systems for advanced biologics manufacturing.