Super Renders Farm's 2026 VFX industry survey lists five simultaneous shifts: real-time engines crossing the final-pixel threshold, AI moving from denoise into the interior of the pipeline, Gaussian Splatting arriving as a first-class asset type, OpenUSD consolidating as the interchange standard, and cloud rendering becoming the default rather than overflow capacity. Four of those are in production in studios we work with this quarter. The fifth isn't, and a CTO who reads the survey as a buy list will mis-sequence the spend.
The real-time-plus-cloud combination is the change that has already landed, and the survey underweights it by listing the two separately. Unreal 5 with Movie Render Queue produces final pixels now; that part is uncontroversial. The interesting operational question is what real-time review looks like when the reference frame is a 16K dome and the artist sits in a different time zone from the venue. Gunpowder's work on the House of Parliament Super Bowl 2024 campaign was real-time review at burst scale two years ago: nine spots, seventy to three hundred artists in six weeks, new artists working in thirty minutes, across two anchor clouds. The hard part was never the render engine. It was identity, license counts, storage performance, and review bandwidth, in roughly that order. Studios that treated real-time and cloud as separate procurement conversations spent 2025 catching up.
OpenUSD has crossed the threshold the survey describes, with a caveat the survey doesn't quite name. The standard works; the 26.03 splat schema is real and Pixar, Adobe, Apple, Autodesk, and NVIDIA are pulling in the same direction inside AOUSD. The friction has moved one floor up. The question is no longer whether USD works for an asset. It's whether an asset graph survives a round-trip through three vendors with different OCIO configs, different ACES versions, and different opinions about layer composition. Most of the pipeline rebuild work we do this year is at that layer. The interchange format is solved. The asset-graph contract between studios is the work nobody has been paid to do yet, and it's where the next eighteen months of integration spend will go.
The messy middle
AI inside the interior of the pipeline is the shift the survey gets right and underplays. Denoising is a solved feature. Intel OIDN winning the 2025 Academy Technical Achievement Award closes that chapter. The interesting work is upstream and downstream: generative inpainting on plates, model-assisted rotoscoping, asset generation feeding into the lookdev loop, automated tagging on dailies. None of it is photogenic. None of it produces the "10x faster" headline that real-time does. Operationally, it's where most of the cost compression of the next two years comes from, and the studios doing it well share three traits. They have been treating model versioning as a pipeline concern since about 2023. They have asset residency clear enough that audit and rights conversations don't stall a delivery. And they have built storage that knows the difference between a generated frame and a human-touched frame, because their lawyers asked them to.
The studios just starting on this now are running into walls they didn't see, and most of those walls are storage and metadata problems wearing AI-shaped clothing. We've documented some of the underlying infrastructure on Google Cloud's blog on knfsd caching. The point that matters here is that the same primitive that makes a render fleet cost-effective is what makes a generative workflow auditable. Cache topology and asset residency end up being the same conversation, and "cloud as overflow" versus "cloud as default" describes two different cost structures rather than two different volumes. Studios that built proper caching and spot strategy get a render bill that scales sub-linearly with output. The ones that lifted-and-shifted their on-prem queue policy to cloud get the bill the survey's pessimists were warning about.
What's actually a year or two away
Gaussian Splatting is the one the survey treats as having arrived, and tool support has: Houdini 21's tech preview and the OpenUSD 26.03 first-class splat schema are the two that matter. Framestore's roughly forty final-pixel splat shots on Superman last year is a meaningful production data point. None of that translates to splats being daily currency in a typical pipeline. Node memory profiles change. Asset graph size goes up by an order of magnitude on a hero shot. Render-node selection logic has to learn the difference between a mesh asset and a splat asset, and the queue policy tuned for the former isn't tuned for the latter. The standard has shipped. Most shops haven't rebuilt the plumbing to use it. Plan for splats as a hero-shot capability in 2026 and 2027, then as a default asset type once the render-node tiers have caught up.
The survey is useful for what surveys are useful for: planning the next eighteen months. Four of the shifts are work to do now. The fifth is work to plan for. The order matters because budget conversations matter, and "real-time and cloud and AI and USD are all happening at once" is true but unhelpful as a buying signal. The buying signal is that real-time plus cloud is the spend that pays back first, USD work is the integration debt to clear next, AI is the operating discipline to start now if you haven't, and splats are the line item to keep an eye on. If you're a CTO reading the survey and trying to sequence the work, that's where I'd start. Your stack will disagree with mine in places. That's the conversation worth having.
