Every engineering leader running an AI-accelerated simulation program has had some version of the same conversation. The board wants faster discovery cycles. The CFO wants a defensible ROI story. The platform team wants to ship. And somewhere in the middle, a senior researcher quietly mentions that the workflow has been broken for three weeks because a dependency upgrade silently changed the behavior of a pretrained model.
This is the part of the cost structure that nobody put in the business case.
Open source remains essential infrastructure for AI in scientific computing. It accelerates early exploration, supports interoperability, and gives technical teams the inspection rights they need to trust what a model is doing. None of that is in question. What is in question—increasingly, and rightly—is the assumption that “no license” means “low cost.” In AI-accelerated simulation, where workflows are already complex, and the answers feed into expensive downstream decisions, that assumption tends to fall apart fastest.
See also: AI Without Governance Is Just Faster Risk
The Pressure Changing the Math
Two things are converging at once. The first is that AI is now operational. McKinsey’s 2025 survey of global organizations found that 88% are using AI in at least one business function, but only about one in three has actually scaled AI across the enterprise. The gap between “in use” and “at scale” is where most of the hidden cost lives—as well as most of the abandonment.
Gartner was blunter about that abandonment risk, projecting that at least 30% of generative AI projects will be dropped after proof of concept, citing poor data quality, weak controls, escalating costs, and unclear business value. None of those reasons are licensing-related. All of them are operating-related.
The second shift is that simulation has moved from a periodic, batch-oriented activity into something closer to real-time decision support. Materials, chemistry, and process teams increasingly need answers on the same cadence as the product decisions those answers inform—days, not quarters. That changes what the underlying stack has to deliver. A workflow that works once, in a notebook, run by the engineer who built it, is not the same thing as a workflow that produces reliable answers on demand for a team of researchers who don’t have time to debug it.
Treating simulation as developer infrastructure, rather than as a research tool, is the shift that makes the real cost structure visible.
See also: How AI Is Forcing an IT Infrastructure Rethink
Where the Engineering Tax Actually Accrues
An open simulation stack typically combines an atomistic engine, Python workflow libraries, one or more pretrained machine learning interatomic potentials, cloud or cluster provisioning, notebooks, visualization, and storage. Stitched together, that architecture is good for exploration. The trouble starts when the requirement shifts: from one researcher proving a method works, to a team needing repeatable results across projects, with access control, version tracking, benchmarking, and onboarding for people who aren’t computational specialists.
At that point, the unpaid work begins to pile up. Someone owns dependency updates and the regression testing that follows. Someone owns GPU and driver compatibility. Someone owns reproducibility, making sure that the same inputs, run six months later, produce the same outputs. Someone owns documentation, because the original author has moved to a new project. Someone owns the question of which pretrained model is acceptable for which use case, and why, and how that gets revisited as new models appear in the literature every few months.
None of that work is visible on a license line. All of it is paid for in engineering time.
The cost becomes legible when teams measure engineering time honestly. A 2024 developer experience study run with Atlassian found that developers lose roughly a day a week to internal friction—dependency issues, environment setup, broken tooling, the slow accumulation of context-switching cost. DORA’s 2024 DevOps research, looking at the same problem from the other end, found that well-built internal platforms measurably improve individual productivity, team performance, and organizational outcomes.
If a “free” stack is consuming a day a week of high-value specialist time, the license was never the variable that mattered.
Governance is Now Part of Runtime
The second hidden cost is governance, and it has gotten harder, not easier, as open source has moved deeper into critical systems. The Linux Foundation found that organizations systematically underinvest in security practices, formal governance, and comprehensive strategy even as open source becomes core infrastructure. The same study found that 71% of organizations expect sub-12-hour response times for mission-critical open-source workloads—a service level most volunteer maintainers were never going to provide.
The vulnerability data tells the same story. Black Duck’s 2025 open-source security and risk analysis, drawn from audits of more than a thousand commercial codebases, found that 86% contained known vulnerabilities, 81% contained high- or critical-severity ones, and 91% contained components more than four years out of date. Those numbers do not argue against open source. They argue against pretending it manages itself.
In simulation environments, this matters more than it might in other domains. The toolchains pull in scientific libraries, ML frameworks, cloud services, and a long tail of transitive dependencies—any one of which can shift the behavior of a downstream model without anyone noticing for weeks. Tracking that surface area is part of the ongoing cost of operating the platform. So is responding to it. The U.S. NSA, CISA, and ODNI have explicitly tied software bill-of-materials practices to vulnerability management for exactly this reason: in real-time-adjacent systems, transparency about what’s running may be a compliance issue, but it’s also a core part of operations.
Data Quality and Reproducibility Belong on the Cost Line
Simulation teams also inherit a data burden that rarely appears in headline price comparisons. Every meaningful simulation generates structures, trajectories, forces, energies, metadata, and provenance—and all of that needs to be preserved if the results are going to be reusable, auditable, or defensible six months later. The FAIR data principles (findable, accessible, interoperable, reusable) exist precisely because scientific results without traceable inputs, model versions, and assumptions decay in value fast.
In a research setting, that decay is acceptable. In a production R&D setting, where decisions about candidate materials feed into multi-million-dollar experimental programs, it isn’t. A simulation result you can’t reproduce, can’t trace to a specific model version, or can’t validate against a known benchmark is not actually saving anyone time. It’s producing optionality that the organization can’t act on.
This is also where the real-time framing matters. The point of investing in AI-accelerated simulation is not to produce answers faster in isolation; it’s to compress the loop between a hypothesis and a decision the business can act on. That compression only works if the answers can be trusted on the cadence they arrive.
Measure the Right Thing
The cleanest way for engineering leaders to think about total cost of ownership in this space is to forget the license question entirely and treat simulation the way they would treat any other piece of developer infrastructure. TCO is the full lifecycle: acquisition, integration, configuration, ongoing operation, maintenance, governance, retirement. License fees are usually the smallest line in that stack.
The return, similarly, should be measured the way platform investments are measured everywhere else. Not by license avoidance, but by decision latency—the time from a new question to a trustworthy answer. By the number of candidates a team can credibly screen before expensive validation begins. By how often results are reusable across projects rather than rebuilt from scratch. By how much specialist time is spent doing actual science versus maintaining the substrate the science runs on.
Measured that way, the choice in front of most engineering leaders becomes less about longer open versus proprietary. It is the right operational mix: open components where exploration and interoperability matter most, managed infrastructure where reliability and governance carry the load, and a clear-eyed accounting of where engineering time is actually going. The teams that get this right won’t be the ones that paid the least for software. They’ll be the ones that paid the closest attention to the cost of delay.
