When to Modernize Your SCE: Build Now or Wait for Cloud Migration
Introduction
If you're responsible for a statistical computing environment in pharma, or if you're about to commission one, the infrastructure question will come up early: should we build on-prem, or should we start in the cloud?
It sounds like a technical decision. It is, partly. But it has real consequences for timelines, costs, and how quickly your teams can adapt to new requirements. This piece breaks down the trade-offs in plain terms, so you can have that conversation with your technical partners, or make the call yourself, with the right context.
The short version: on-premise infrastructure still has a role, but the nature of scientific computing is changing fast enough that cloud deserves serious consideration as a starting point, not an afterthought.
Let's start with what each option actually means.
On-premise is infrastructure that your organization owns or contracts out to a managed data center. Whether it sits in your building or in a partner facility, the hardware is dedicated to you. Your team (or your provider's team) is responsible for keeping it running, upgrading it, and scaling it when needs change. This includes private cloud setups, where the infrastructure behaves more like cloud but remains dedicated to your organization.
Cloud is shared infrastructure operated by a provider like AWS, Azure, or GCP. You don't own hardware. You consume compute, storage, and services on demand, and the provider manages the platform underneath.
Both have a place. The question is which one fits the kind of system an SCE needs to be today.
Where on-premise still makes sense
On-premise infrastructure works well for workloads that are predictable, stable, and long-running. Large-scale data storage where you know the volume and growth rate. Batch processing jobs that run the same way month after month. Environments where the compute profile hasn't changed meaningfully in years.
For these workloads, on-premise is often the most cost-effective option. You buy the hardware once, amortize it over its useful life, and the marginal cost of running another job is close to zero. If your SCE looked like this, and for a long time it did, on-premise was a rational default.
There is also a practical reality around data. In most organizations, historical and core datasets were designed to live on-premise. They are large, they have been there for years, and moving them wholesale is expensive and often unnecessary. A common pattern is to keep that data on-premise while cloud compute either connects to it remotely or works with a smaller operational subset that gets transferred for faster processing. This is one of the reasons hybrid architectures are becoming the default rather than a full migration to one side or the other.
Where it breaks down for SCEs
The problem is that SCEs are no longer stable workloads. The compute requirements are shifting faster than hardware procurement cycles can keep up with.
Teams now need multi-language support (R, Python, and SAS). They need hardware that can handle machine learning (and AI) workloads. They need environments that can scale up for a large submission and scale back down when it's done. They need iteration cycles measured in hours, not months.
This is what cloud infrastructure was designed for. Elastic compute. On-demand hardware. Fast provisioning. The ability to spin up a new environment configuration in minutes rather than waiting weeks or months for procurement and setup.
On-premise can technically deliver all of this, but it takes significantly more time and investment. The process of specifying, procuring, and provisioning new hardware can take months. By then, the requirements may have moved again.
Look at what some of the largest pharma companies are doing at the cutting edge. In June 2025, Novo Nordisk partnered with NVIDIA and the Danish Centre for AI Innovation to use the Gefion sovereign AI supercomputer for drug discovery, including molecule design and single-cell simulations. In October 2025, Eli Lilly announced a collaboration with NVIDIA to build the pharma industry's first in-house AI factory, a DGX-based supercomputer with over 1,000 Blackwell Ultra GPUs running on renewable energy.
These are impressive, but they are edge cases: massive strategic investments for very specific, high-performance workloads like molecular simulation and generative AI for drug discovery. They are still far from the day-to-day production workloads that most SCEs handle. Not every organization can pursue partnerships of this scale, and most don't need to.
The good news is that this class of compute is increasingly available through commercial cloud offerings. AWS EC2 UltraClusters provide on-demand access to thousands of GPUs through a pay-as-you-go model, no capital commitment required. AWS ParallelCluster lets you spin up full HPC environments without owning any hardware. Azure offers GPU-accelerated virtual machines for compute-heavy workloads. Hyperion Research projects the cloud HPC market will reach $11.5 billion in 2026, approaching half the size of the on-premise HPC server market. The GPU-as-a-service market more broadly is projected to grow from $6 billion in 2025 to over $160 billion by 2034.
Cloud is the democratized path to similar infrastructure. The same types of hardware that power those headline partnerships are becoming accessible on demand, without the capital outlay or the need for a strategic partnership to get there.
The security conversation
Security is the most common objection to cloud in regulated industries. It deserves a direct response: the idea that on-premise is inherently more secure than cloud is outdated.
Leading cloud providers invest billions in security infrastructure, more than most pharma companies' entire IT budgets. Gartner has projected that through 2025, 99% of cloud security failures would be the customer's fault, primarily through misconfiguration, not platform weakness. The risk is not that cloud platforms are insecure. The risk is that teams are not trained to configure them properly. That is a solvable problem. Training, guardrails, and well-designed security architectures address it directly.
Cloud providers maintain compliance certifications (SOC 2, ISO 27001, GxP-ready environments) that would be enormously expensive to replicate on-premise. For most organizations, cloud is now easier to secure well than on-premise, because the tooling and automation are built in.
Build portable, regardless of where you build
Here is the practical recommendation: even if cloud is not available to you today, build as if it will be tomorrow.
That means asking your technical team or partner to adopt cloud-native architecture patterns: containerization, infrastructure as code, automated deployments. These are not just technical buzzwords. They are engineering practices that make your platform portable, reproducible, and easier to maintain. (If you want to understand what each of these means in concrete terms, and why they matter for your SCE investment, I'm planning a follow-up piece that breaks them down for non-technical stakeholders.)
The industry has already moved in this direction. The CNCF's 2025 annual survey found that 82% of container users now run Kubernetes in production, up from 66% in 2023, and 65% run it across multiple environments specifically for portability. These are no longer emerging practices. They are the standard.
This is the architecture we apply when building BioVerse environments for pharma clients. Every layer is containerized and defined as code from day one, so the platform can run on-premise today and shift to cloud when the organization is ready, without a rewrite.
If you build on cloud-native patterns today, a future migration from on-premise to cloud becomes incremental: moving workloads one at a time, not rewriting the entire platform. If you build on older patterns, you are locking in an architecture that will need to be replaced from scratch when the move eventually happens.
Hybrid is the enterprise reality
The framing of "on-premise vs. cloud" is increasingly artificial. Most large organizations are not choosing one or the other. They are running both.
Gartner predicts that 90% of organizations will adopt hybrid cloud infrastructure by 2027. Approximately 83% of pharmaceutical companies already leverage cloud solutions in some form, with 40% fully cloud-enabled. The hybrid cloud market itself is projected to reach $348 billion by 2031.
For SCEs specifically, the pattern that works best in larger organizations is to keep stable, predictable workloads and core data on-premise, and run elastic, hardware-intensive, or rapidly evolving workloads in the cloud. For many large pharma companies, this is already the direction they are heading.
For smaller or newer companies, the calculus is simpler. The capital investment for on-premise infrastructure does not make sense when you are building from scratch. Cloud lets you start fast, scale as needed, and avoid the upfront hardware commitment that takes years to amortize.
Summing up
The decision is not really "build now or wait." It is "build smart, wherever you build."
If cloud is available to you, it is the right starting point for an SCE modernization. The elasticity, on-demand hardware access, and operational model fit what modern SCEs need.
If you don't know when your cloud migration is coming, and you're debating whether to start now, start. Even just designing and testing now, running proofs of concept will add a lot of valuable insight. Most of this foundational work transfers, and almost all of the process work and patterns will carry over nicely.
If on-premise is a hard constraint, build on cloud-native patterns anyway. Containers, infrastructure as code, automated validation. You will get a better platform today and an easier migration path when cloud becomes an option.
If you are a large organization with existing on-premise infrastructure, hybrid is likely your answer, and the industry data suggests you are already heading there.
The worst option is the one too many organizations still choose: building on legacy patterns because "that's how we've always done it," and deferring the cloud question entirely. Or worse, carrying those same legacy patterns into the cloud, running single oversized machines in an imposing but immovable HPC setup that does nothing but skyrocket costs. Every month of that kind of thinking is another month of technical debt that compounds.
The SCE you need to have is becoming a dynamic, multi-language, AI-ready platform whether you planned for it or not. Build the infrastructure to match. If you're at the point of making this decision, let’s talk.
