AI gets the headline. Infrastructure does the work. Every panel and trade headline this year is about the model, the prompt, the generative tool that turns a sentence into a shot. The thing that decides whether any of that reaches a delivery on a Friday night is the layer underneath it, and almost nobody writes about that layer because it isn't photogenic. Cache topology doesn't make a sizzle reel. It's also where the promises on the conference slides either become true or stay slogans.
The shape of the production stack is changing fast enough that the infrastructure question is the live one. Virtual production sits at roughly $3.67 billion in 2026, and per Mordor Intelligence and SNS Insider, LED volumes are expanding more than 30% a year and are on track to overtake chromakey before 2031. Mordor's read is that cloud collaboration is widening the opening for managed-service firms. That tracks with what I see in render trends: per the 2026 surveys from iRender and Super Renders Farm, UE5 is producing final pixels, AI-assisted denoise and upscale are standard, USD is the interchange layer, and cloud render has become the default for cross-border teams rather than overflow capacity. None of those four shifts works without somebody making the boring layer underneath behave.
Where the boring layer earns its keep
Here is what that layer looks like when it works. For Refik Anadol's Coral installation at Davos in 2023, we rendered 100 million coral images on up to 250 T4 GPUs across Google Cloud Spot VMs. Three weeks. The same job on LA studio servers ran six. That's not a render-engine win, it's an infrastructure win: the spot pricing made the GPU count affordable, and a knfsd cache sitting between on-prem assets and the cloud fleet gave us up to 15x the performance you'd get pulling cold every time. The render engine was the same render engine. The difference was whether the fleet could be fed.
The cache story scales further than that. For the House of Parliament Super Bowl 2024 work, a cloud-based knfsd cache delivered up to 150x performance on the right access pattern. When you're moving 5 million rendered frames and 2 PB of data, the cache stops being an optimization and becomes the thing standing between you and a missed air date. Sohonet has written up how that cloud build came together from the production side. On the storage layer, WekaFS cut the storage spend on that job by 80%. The headline tool was the creative. The reason it shipped was a cache and a filesystem nobody in the audience will ever hear about.
Scale and access pattern decide everything at this tier, and the trade-offs are real. Spot capacity cut the Coral bill hard because the workload tolerated restart. On a four-hour simulation, a spot eviction is a producer call, not an engineering one, and the 15x cache number assumes a working set that fits the cache. Get the access pattern wrong and you've paid for infrastructure that's reading cold anyway. For Urs Fischer's "CHAOS," we ran 2.6 TB of memory, 1.5 million render cores, and 400 GB/s of sustained throughput across a 250-node knfsd cluster and delivered in one week. That throughput figure is the number that matters, because it's the one that breaks first when you scale an artist team and forget the storage tier underneath it.
The careers question hiding inside the infrastructure question
There's a jobs story tangled up in all of this, and it's worth being straight about. Per Metaintro and AVG Guild, around 118,500 US film, TV, and animation jobs, roughly 21% of the workforce, are projected to be consolidated or replaced by generative AI by 2026, and LA County has lost about 41,000 film and TV jobs in three years. Those are real people and a real number. The same sources track new roles coming the other way: AI animation supervisor, generative content director, AI pipeline specialist. The infrastructure layer is a big part of what decides which side of that number a crew ends up on.
Dean Yurke, who has spent 30 years in Hollywood VFX and now works in virtual production, puts the careers logic plainly: "If you want to get a job, specialize in something, but if you want to have a career, generalize." His read on the tools is the operator's read. AI, in his words, "is brilliant at copying things it's already stolen." Ask it for original work - his example is a weird robot and a spaceship that have never existed - and you still need the skilled human. Kevin Vandermarliere's UE5-plus-AI virtual production walkthrough lands in the same place: the new AI techniques speed up the workflow, and the workflow still starts with shooting live-action footage, then rotoscoping or keying it. The model is a faster tool inside a pipeline that a crew still has to run.
That's the version of this era I'd bet on. The generative tools are getting genuinely good, and they need a render fleet, a cache, a storage tier, and a real-time review path to be useful on a real delivery. Whoever owns that backbone owns the part that scales, and the skilled crews who run it are the ones who stay employable when the headline tool changes again next year. We don't sell a model. We're the media-and-technology partner that builds the layer the models run on, so the work that gets the applause can actually ship.
Make the technology invisible, so the creators can focus on creating. The cache, the spot strategy, the throughput, the review bandwidth - get those right and the AI in the headline does the thing the headline promised. Get them wrong and you've bought a very expensive way to render cold. That's the whole job, and it's the part nobody puts on the panel.
