Shell

Capacity When the Business Needs It: How Shell Is Rethinking Hybrid HPC with Qumulo

Shell is evolving its global high-performance computing environment to make data and compute more flexible across on-premises infrastructure and AWS, helping teams respond faster to changing business and computational demands.

The Business of Computational Agility

For Shell, high-performance computing is not simply an infrastructure service. It supports work that can sit directly on the path to major business decisions.

A significant portion of Shell’s HPC demand supports subsurface imaging, including seismic data processing, reservoir simulation, and drilling optimization. Seismic acquisition can generate petabytes of data that Shell processes through computationally intensive algorithms to build images of geological structures miles beneath the Earth’s surface. The results help scientists better understand subsurface conditions when evaluating potential assets.

Approximately 85% of Shell’s HPC demand supports subsurface imaging, according to Michael Gujral, Shell's Global HPC Engineering & Operations Manager. Another 10% supports AI workloads, including supervised machine learning, generative AI techniques, model fine-tuning, and AI for science. The remaining workloads include computational fluid dynamics, weather modeling, and reservoir simulation.

The scale is substantial, but raw performance is only part of the equation.

For Shell, the increasingly important question is whether the business can access the required computational resources when it needs them. Thousands of GPU cores, petabytes of memory and storage, power efficiency and sustainability, supply-chain constraints, and infrastructure lead times all influence how quickly computational work can move forward.

A system with excellent price-performance characteristics has limited value if the business must wait months for the infrastructure required to use it. The value decisions these systems drive far outweigh the value of the hardware itself. 

That reality pushes Shell toward a different HPC model: one where compute capacity becomes more dynamic, and data does not depend on proximity to a physical location.

The Challenge: Compute Is Becoming Dynamic. Data Must Follow.

Traditional HPC environments are largely built around infrastructure that an organization owns, installs, and operates for years. Capacity planning means forecasting future demand, purchasing infrastructure, and building enough headroom to accommodate expected growth.

Cloud computing introduces a fundamentally different possibility.

Instead of determining years in advance exactly how much infrastructure it will need, Shell can combine its existing on-premises HPC foundation with cloud resources that provide additional capacity when and where it is needed. This extends the life and functionality of an existing system without constraining operations.

Shell has operated HPC workloads in a hybrid AWS environment since 2022, but making hybrid computing work effectively requires more than access to cloud compute.

The challenge was the data.

With innovation delivered by Qumulo, Shell engineers and scientists teams may use applications both on-premises and in the cloud. Individual users may move between environments. Different disciplines may collaborate on the same projects. Applications, repositories, databases, and computational resources may reside in different places.

Where before these operations required expensive and time consuming data replication operations, now the data never moves from the desired persistent location.

The old approach created friction. It introduced additional infrastructure, processes, duplicate datasets, operational complexity, and substantial expense. 

All of these constraints are eliminated.

“What our end users want is what we all want from any software or platform. We don’t want to have to think about where our data lives. With Qumulo Cloud Data Fabric, users simply access their data as if it were local, whether it’s across the hall or in the cloud.”

Michael Gujral
Global HPC Engineering & Operations Manager, Shell

The Solution: Bring Data to the Opportunity

Qumulo is the backbone of what Shell calls its technical storage landscape, supporting an architecture that spans its on-premises and cloud consumption zones.

Qumulo Cloud Data Fabric extends that model by helping Shell rethink how users and applications access data across those consumption zones, optimizing for overall performance, cost, and time-to-value. The benefit is threefold: (i) eliminate data transfer workflows, saving time; (ii) reduce operating expense with elastic management of compute resources, reducing fixed expenses; (iii) increase overall capacity to perform business-critical analysis, both increasing value and decreasing scheduling burdens. 

Consider a multidisciplinary team where some applications run on-premises while others run in AWS. The team should not need separate working datasets simply because applications run in different locations.

Chasing Compute Without Chasing Data

Accelerated computing represents a significant portion of Shell’s HPC requirements. Gujral estimates that 80 to 90% of Shell’s seismic imaging workload is accelerated, while its AI workloads also rely heavily on accelerators.

That makes GPU availability a strategic consideration.

Shell maintains dependable on-premises base capacity and active capacity in AWS. From there, it can explore additional options such as AWS Capacity Blocks, multiple AWS Regions, and potentially more dynamic compute resources.

Therefore, Shell benefits from a nimble consumption model, utilizing low cost, available resources without pre-staging.

Shell has already begun running projects across multiple AWS Regions. But as Gujral points out, multi-region computing quickly comes back to data. Expensive GPUs provide little business value when they sit idle, waiting for the data needed to do the work.

Shell therefore sees making data available to compute as key to expanding its computational options.

From Infrastructure Forecasting to Business Decisions

Shell’s longer-term vision extends beyond hybrid infrastructure.

Gujral describes a “10X” challenge as the organization's North Star: finding ways to dramatically reduce the wall-clock time required to complete computational work.

Simply purchasing ten times as much hardware is not reasonable. Shell requires solutions that adapt and scale without linear investment. 

They are breaking traditional capacity-management dependencies and the reliance on forecasts. Organizations no longer need to estimate what they will need years in advance, acquire infrastructure, and hope demand develops roughly as predicted.

Gujral compares that model to building a highway: by the time the new road is finished, another traffic issue may already be forming.

The more dynamic infrastructure model allows Shell to add the equivalent of lanes as demand rises and remove them as demand falls.

That changes capacity planning from a primarily infrastructure-driven decision into a business-driven one.

How much compute does the project need? How quickly does it need results? How long will that capacity be required? What is the business value of completing the work sooner?

Data needs to be ready to support those decisions.

For the model to work, Gujral says data availability must become an assumption: the data is there, hydrated, and performs as required.

Business Outcomes

Greater freedom to use available compute

A hybrid architecture gives Shell more options for where computational work can run. As cloud capacity becomes available across different locations and consumption models, Shell can explore those resources without treating data location as an immovable constraint.

More seamless collaboration across environments

Cloud Data Fabric helps teams maintain access to shared project data even when users and applications span on-premises and cloud infrastructure. The goal is a consistent data experience, not separate datasets and transfer workflows for every environment.

A path toward more dynamic capacity

Shell can combine predictable base capacity with more flexible cloud resources, allowing infrastructure decisions to more closely follow computational and business requirements.

Reduced dependence on long-term infrastructure forecasts

Instead of relying entirely on multi-year forecasts and fixed infrastructure investments, Shell is working toward an environment where capacity can respond more quickly as portfolios, projects, or computational requirements change.

Better utilization of expensive accelerated compute

Making data available where compute exists can help reduce the risk of valuable GPU resources waiting for datasets to arrive, an increasingly important consideration as accelerated infrastructure becomes more expensive and difficult to secure.

The Takeaway: Make Data Predictable So Everything Else Can Change

Shell’s HPC environment is specialized, but the infrastructure problem it is addressing is not.

A media company may need rendering capacity. A pharmaceutical company may need GPUs for drug discovery. A manufacturer may run engineering simulations. A financial services organization may need compute for risk analysis. A research institution may need temporary capacity for a major project. An enterprise AI team may need access to whichever GPU infrastructure becomes available.

The applications are different. The architectural challenge is remarkably similar.

As compute, applications, teams, and business priorities become more dynamic, data cannot stay tied to a single location without creating friction.

The broader lesson from Shell is therefore not about seismic imaging or oil and gas. It is about decoupling where data is stored from where work can happen.

Organizations can maintain the infrastructure that makes sense on-premises while using cloud services where they provide greater availability, scalability, or agility. Teams can adopt cloud gradually rather than waiting for a perfect migration point. Infrastructure leaders can increasingly make capacity decisions based on current business needs rather than forecasts made years earlier.

That flexibility may become particularly important as AI increases demand for GPUs and other scarce infrastructure.

The goal is not necessarily to move everything to the cloud. Nor is it to keep everything on-premises.

It is to give the business more choices.

Shell’s experience also demonstrates why ideas can move between industries. A data architecture built for one workload may solve a completely different problem elsewhere.

For Shell, that principle is helping shape the next stage of hybrid HPC: keep the data experience predictable, so the infrastructure and the business can become more dynamic.