The LBBOnline piece on generative AI cost compression is correct on the headline numbers. Cost-per-shot in post and VFX is coming down hard between now and 2030. Assembly edit, rotoscoping, layout, and the early iterations on a sequence are where the compression shows up first, because a generative pass replaces several hours of human work without a fight about authorship. Smaller teams are moving faster.
The headline number is still the wrong one to plan around. Cost-per-shot is a useful retrospective metric and a poor forward-planning unit, because it tells you nothing about how the shape of the work changes when you cut it. When the marginal cost of one more iteration drops by an order of magnitude, the iteration count goes up. Sometimes by a lot. Directors and clients have always wanted another version. The bill and the calendar stopped them. Take those out and the cadence resets. The interesting unit is shots-per-iteration, and how many iterations the pipeline survives in a week.
More iterations means more versions on disk, more egress between artist and render and review, more queue contention when three sequences all want the farm at once, more revisions sitting in storage that nobody has decided to keep or delete. None of this is sexy. All of it is non-negotiable. Storage costs depend on how disciplined your asset residency policy is. Render cost depends on whether your cache sits next to your compute or three hops away. Queue policy decides whether your producer can preempt a long-running job for a director's note without rebuilding the dependency graph by hand. Studios already doing this work absorb the new cadence cleanly. The ones who hadn't budgeted for it are running into pipeline ceilings they didn't expect to hit this year, on jobs they won fair and square.
The House of Parliament Super Bowl scale-up is the closest public example of a pipeline absorbing a new cadence at speed. 70 to 300+ artists in six weeks, new artists working in 30 minutes, nine commercials delivered through a single creative pipeline that hadn't existed eight weeks earlier. It worked because the storage layer, the identity layer, the license layer, and the review layer were all sized for the iteration rate the campaign was about to ask for, not for the iteration rate the studio had been used to. Our partner Sohonet was in the connectivity layer. The parts of the pipeline nobody photographs were the parts that decided whether it shipped.
The cache piece sits next to that. We documented the knfsd caching work on Refik Anadol Studio's 'Coral' with Google Cloud because the numbers were honest and the architecture is reusable. 15x render performance gain on a workload that previously couldn't iterate inside its window. Nobody at Gunpowder thinks knfsd is magic. The architectural point is that when iteration count goes up by an order of magnitude, the cost of moving the asset to the compute starts to dominate the cost of the compute itself. Caching close to the render fleet flips that ratio back. Without it, every extra iteration is paying egress and warm-up time on data the artist has already touched twice today. Treat caching at the render fleet as a planning input. The tooling preference comes second.
The LBBOnline piece notes the structural pressure on vendors to move toward cloud and AI-ready pipelines to keep pace with the new turnaround times. That pressure is real. In our experience it gets resolved last, because it shows up as a capex conversation rather than a creative one. Studios buy software faster than they buy infrastructure. Generative tools are cheap to trial and easy to roll back. A storage architecture rebuild is neither. Most of the operational pain in 2026 sits in that lag, and I don't see it closing by accident.
The shops that win in 2026 to 2027 are the ones who stop treating the iteration cycle as a creative-process question and start treating it as a capacity-planning question. Three things go on the agenda. Storage and asset residency, because hot data near compute is what makes iteration cheap, and cold data on the wrong tier is what makes the storage line ugly. Queue policy and render orchestration, because which job runs when, on which instance class, with what preemption rules, decides whether the farm survives a busy week. Versioning discipline, because every extra iteration generates artifacts somebody has to decide to keep, and a studio without a working policy on that is paying twice: once for the storage, once for the producer who has to find the right version on Monday morning.
None of that is a moonshot. It needs two days with the operations team, writing down what the new cadence actually looks like. The cost-per-shot story is real. The operational story underneath it decides who benefits.
