How Aether retired a customer's seven-figure CAE renewal

Abstract

We describe a paid two-quarter pilot at a mid-sized aerospace prime in which Aether was evaluated head-to-head against the customer's incumbent commercial CAE stack across twelve real production cases. The cases were drawn from the customer's actual workload and covered structural FEA, CFD, electromagnetics, explicit-dynamics crash, and coupled fluid-structure. Seven cases had measured wind-tunnel or wind-on-rig data; five had production telemetry. Aether matched or beat the incumbent stack on accuracy in ten cases, was within 3% on the other two, and completed the campaign in 38% of the wall-clock time at iso-accuracy. The customer retired a $1.8M three-year renewal at the end of the quarter and consolidated their licence stack down to a small residual count. We document the methodology, the results, the cases that did not go our way, and what the customer changed about their workflow as a result.

1. The setup

Our customer — referred to as the prime — was about to sign a $1.8M three-year CAE renewal across structural, fluids and electromagnetics. The renewal covered the leading commercial structural FEA tool, two commercial CFD packages, a full-wave electromagnetics suite and an explicit-dynamics crash code, split across two established vendors. The team had used the stack for over a decade. Methodology was mature. Engineers were trained. The decision was inclined to renew on momentum.

The internal champion at the prime had read our impinging-jet benchmark publication and proposed a paid head-to-head pilot. We agreed on twelve real production cases — drawn from the prime's actual current workload, not from a demo set — with a written win condition agreed in advance. If we matched or beat the incumbent on at least eight of twelve at iso-accuracy with no statistically significant regressions, the prime would not renew.

We assigned a forward-deployed engineer to pair with the prime's CAE lead for the duration of the pilot. Pilot duration was capped at two quarters from kickoff to renewal decision. The pilot fee was credited against year-one platform licence cost if the prime proceeded.

2. Case selection

Cases were chosen by the prime's CAE lead, not by us. The mandate was: cases that look like our normal workload, not cases that are easy or that show the new tool in its best light. Five aerodynamic studies on a fixed-wing platform, three crash and impact simulations on a structural test article, two electromagnetic compatibility runs on the same airframe, and two coupled fluid-structure cases on a rotor stage.

Seven of the twelve cases had measured experimental data — either wind-tunnel results from a major transonic facility, wind-on-rig measurements from the prime's own rig, or measured antenna-range data. Five cases had production telemetry from existing airframes in service rather than tunnel-grade measurements. Mesh budgets and time-step budgets were matched between the two stacks. Solver settings were tuned independently for each tool by the engineers familiar with that tool — we did not handicap the incumbent.

3. Measurement methodology

For cases with experimental data we used direct comparison — lift and drag polars against tunnel runs, peak strain against gauge data, antenna patterns against measured ranges. The metric was deviation of the predicted integrated quantity from the measured one, normalised by the measurement uncertainty where the prime had documented it.

For cases without measured data we used the prime's own acceptance criteria from their existing CAE methodology — yes, the criteria the prime was already using to qualify the incumbent stack's results. This is intentionally a high bar: we were asked to be qualified by the same methodology that the incumbent had passed for a decade.

We blinded the result tabulation. The prime's CAE lead tabulated outcomes per case before knowing which tool had produced which result; we provided the prediction with a hash and revealed the source after the lead had marked pass/fail per case. This eliminated one obvious source of bias.

4. Headline numbers

Of the twelve cases, Aether produced a closer agreement with measured data or with the prime's acceptance criteria in ten. Two cases came in within 3% of the incumbent stack, with no statistically significant winner. Zero cases were won by the incumbent at greater-than-3% margin. The win condition (eight of twelve at iso-accuracy) was met by margin.

Campaign wall-clock — the total time from project kickoff to final result for all twelve cases — was 38% of the equivalent campaign on the incumbent stack. The wall-clock improvement was concentrated in the coupled fluid-structure cases, where the incumbent flow required two solvers and an inter-tool import step that Aether subsumed into a single solve.

5. Where Aether did better

The clearest wins came from the coupled fluid-structure cases on the rotor stage. Aether solved fluid and structural together on a single mesh; the incumbent stack required two solvers, an inter-tool import step, and a manual time-step matching pass. The model not only finished faster — it produced a closer agreement to the rig measurements at the trailing edge, where the inter-tool import previously introduced a small but consistent offset that the prime had been carrying as a known quirk.

The aerodynamic cases were a more modest win. Aether produced lift and drag polars within 0.6% of measured values across the angle-of-attack range; the incumbent produced 1.1%. Both are competitive results in absolute terms; the margin matters mostly because the prime's iteration cycles compound the difference into hundreds of design-exploration runs per programme.

On the crash cases, Aether matched the incumbent on peak deceleration and on integrated impulse, and beat it modestly on local strain concentration in the regions the prime's gauges sampled. The explicit-dynamics regime is where the incumbent had the most maturity; we expected to lose this category and were pleasantly surprised to draw.

6. Where it was a wash

Two cases — an EMC study at low frequency and a high-Reynolds external flow on a familiar geometry — came back within 3% of the incumbent baseline, with no clear winner. For these, the prime's existing methodology and engineer hours dominated; either tool would have done the job, and the difference between them was inside the measurement uncertainty.

These cases are useful in two ways. First, they grounded the pilot — we did not claim a sweep. Second, they argued against ripping out the incumbent stack for everything; the prime kept a residual licence count for cases like these where the cost of switching was not justified by the marginal accuracy gain.

7. Where the incumbent had advantages we lost

Two areas where the incumbent stack retained an advantage even when Aether produced better headline numbers.

Specialised post-processing. The incumbent's post-processing tool had two decades of careful UI engineering and a particular set of canned visualisations that the prime's engineers had memorised. Aether produced the same results with a different visualisation workflow that required relearning. The prime's senior engineers needed roughly two weeks before they could iterate at the same speed they had been on the incumbent.

Solver-specific custom user material subroutines. The prime had a library of custom user-material subroutines they had written and validated for the incumbent's structural solver. Porting these to Aether was non-trivial; we built a translation layer that handled most of the surface but the prime ultimately wrote new versions for a small subset of materials where the underlying physics had been hand-tuned over the years.

8. What the customer did

Three things, in this order.

First, the prime did not renew. The renewal letter was sent back unsigned. The incumbent's account team escalated; the prime's CFO held firm because the technical evaluation was clear and the renewal cost was material.

Second, the prime consolidated their CAE team's licences down to a small residual seat count for the niche tools Aether doesn't yet cover — primarily the specialised EMC analysis tool and a small set of custom material flows that proved too costly to migrate inside the pilot window. The residual licence cost is approximately 12% of the previous total.

Third, the prime redirected the freed budget to a new hypersonics R&D programme. The programme is now running on Aether — using the platform's hypersonic CFD capability, which became commercially available shortly after the pilot concluded. The CAE methodology developed during the displacement pilot has been re-used for the new programme; the prime's engineers are now experienced operators rather than evaluators.

9. What we learned

Three things we learned about how a CAE-displacement pilot should be run.

First, blinding the result tabulation matters. The prime's CAE lead's first informal pre-pilot view was that the incumbent was probably better, given the decade of methodology investment; the blinded outcome was the opposite. Without blinding, we believe the result would have been less clean — not because of dishonesty, but because of unconscious expectation effects on case-by-case judgement calls.

Second, pair forward-deployed engineers with senior engineers from the customer rather than with the customer's evaluation lead. The senior engineers carry the methodology knowledge that matters during edge cases; the evaluation lead is necessary for kickoff and decision but not optimal for case-by-case judgement during the pilot.

Third, the post-processing UI matters more than we expected. We have invested substantially in the studio surface since this pilot; the prime's engineer feedback on the visualisation workflow drove a substantial product-design pass.

10. What this is not

It is not a proof that Aether displaces every incumbent at every customer. The cases here came from one customer's workload mix; another customer with a different mix may see different results. The pilot showed parity-plus on this specific portfolio of cases; we explicitly do not claim general superiority.

It is also not a guarantee that every subsequent pilot will run this cleanly. Several subsequent pilots have had mixed results — usually because the customer's workload includes niches that our platform does not yet cover well, or because the customer's existing methodology has features that are hard to reproduce without redoing the engineering work. We document those publicly when the customer permits.

We publish this case study because the methodology and the numbers are reproducible. The cases (anonymised geometries), the meshes, and the agent traces are available on request under NDA. The prime has reviewed and approved this publication.