The structural issue

Hardware cost per core has fallen steadily for decades. Per-core licensing means your simulation cost has not followed it down. That divergence — not the absolute price — is what drives engineering teams to reconsider owning their stack.

This article is about the economics rather than the engineering. If you are trying to justify a decision internally, these are the costs that belong in the model and the ones usually left out.

The costs on the invoice

  • Base solver licence, usually annual.
  • HPC or parallel packs — the cost that scales with cores and dominates at scale.
  • Physics modules — combustion, multiphase, acoustics priced separately.
  • Pre- and post-processing, sometimes licensed apart from the solver.
  • Annual escalation, which compounds over a multi-year horizon.

The costs that are not on the invoice

Hidden costHow it shows up
Idle hardwareCluster capacity you own but cannot license to use
QueueingEngineers waiting for a free seat instead of working
Simulations not runDesign questions left unanswered because the run is too expensive
Automation ceilingsOptimisation loops limited by concurrent licences, not compute
Cloud restrictionsTerms that prohibit or reprice execution off-premises
Embedding prohibitionCapability you cannot ship inside your own product

The most expensive of these is the one nobody measures: simulations that were never run. When each additional case has a marginal licensing cost, engineers self-censor. Design exploration narrows, and the loss appears as a worse product rather than as a line in a budget.

The ownership side, honestly

It would be easy to present ownership as cost-free after the build. It is not:

  • Development — a one-time engineering investment sized by physics scope.
  • Maintenance — realistically 15–20% of build cost annually.
  • Validation and re-validation — every physics change needs fresh evidence.
  • Knowledge retention — documentation and a second engineer, or you have a single point of failure.
  • Opportunity cost — engineers maintaining a solver are not doing engineering analysis.
  • Compute, which both options pay equally.

Ownership converts a licence fee into an engineering salary. That is a good trade at some scales and a poor one at others — and the crossover depends far more on core-hours than on frustration.

Building a model that survives scrutiny

  1. Measure actual core-hours by analysis type for the last year. Usually far more concentrated than expected.
  2. Identify the dominant repeated case — the candidate for ownership, not your whole workload.
  3. Price licensing over five years including escalation and any HPC packs.
  4. Price ownership over five years including build, annual maintenance, and validation.
  5. Add the unmeasured costs — estimate the simulations not being run today.
  6. Apply a risk margin to the build. If ownership only wins by a narrow margin, licensing wins in practice.

Before modelling anything, negotiate. Multi-year commitments, group-wide volume, and simply asking what the number would be at double your commitment routinely produce meaningful movement. Organisations frequently build the business case for ownership without ever testing whether the licence price was fixed.

The hybrid that usually wins

Full replacement is rarely the right structure. What works for most engineering organisations:

  • Keep a small number of commercial seats for exploratory work, unusual physics, and one-off cases.
  • Own the solver for the single repeated analysis that consumes most of your core-hours.
  • Run automation and optimisation loops on the owned solver, where licence limits currently cap you.

You capture most of the saving on the workload that drives it, without taking on breadth you cannot maintain or validate.

When to stop and stay commercial

  • Your analysis mix is genuinely varied rather than dominated by one case.
  • You have no in-house capability to maintain numerical software.
  • Core counts are modest enough that licensing is not the binding constraint.
  • Customers or regulators specify particular commercial tools.
  • The programme horizon is short.

Want this modelled against your real usage — including an honest answer if licensing wins? Send us your core-hours and analysis mix. See our CFD solver service, the TCO comparison, and the build cost breakdown.

Frequently asked questions

Because most commercial pricing treats parallel capacity as a separate product from the solver itself. The solver licence lets you run; HPC packs let you run wide. That means the cost of a simulation grows with the resolution you need, which is the opposite of how hardware costs have moved.
Frequently, and surprisingly few organisations try. Multi-year commitments, academic or startup programmes, and volume across a group all create room. Ask what the price would be at double your current commitment before assuming the list structure is fixed.
For bursty workloads, often yes — you pay only when solving. For sustained heavy use, owned hardware usually wins. But check the licensing implications first: some licence terms restrict cloud execution or price it differently, which can reverse the comparison entirely.