AI in VR: What's Real Right Now, and What's Still Hype
"AI plus VR" has been a conference slide for about a decade. Most of it has been a picture of a brain with some glowing lines coming out of it.
But a few things really have shifted over the last year or two, and one of them changed how we price our work. So here's my sort: what has shipped, what it means if you're a business, and what's still a slide.
1. Making 3D things without a 3D artist
This is the big one, and it's the least glamorous.
Historically, the expensive part of any custom VR project wasn't the code. It was the stuff – every chair, sign, tree and machine in the scene had to be modelled by a person, by hand, at an hourly rate. A branded room with fifty objects in it was fifty small jobs.
Generative 3D tools have taken a serious bite out of that. You describe an object, or hand over a reference photo, and get back a usable 3D model with adjustable detail and materials. At AWE 2026, CGTrader launched a service bundling several generative 3D models behind one subscription – text or image in, exportable asset out, with polygon counts and materials you can dial before you export.
It is not magic and it doesn't replace a good 3D artist for hero pieces. What it does is kill the long tail. The forty background objects nobody looks at closely now take an afternoon instead of a fortnight.
Why you'd care: this is a large part of why we can now quote a custom branded VR experience at a number a mid-sized business will happily sign off. It's the single biggest reason custom VR development got cheaper rather than better.
2. Characters that answer back
VR training has always had the same weakness. You script a difficult customer, the trainee learns the script, and from the second run onwards they're rehearsing a dialogue tree rather than practising a skill.
Generative AI characters change that, because there's no tree. At AWE 2026, Aequilibrium showed a VR training product built on generative agents rather than scripted branches – the character improvises, responds to tone, and doesn't give the same answer twice.
For anything involving a conversation – de-escalation, complaints, sales objections, clinical communication, interviews – that's the difference between a demo and a training tool.
Why you'd care: if your induction or training involves a person handling another person, this is the part of AI in VR that has a real business case today. If your training is procedural – operate this machine, follow this sequence – you don't need AI characters at all, and you shouldn't pay for them.
3. Capturing real places properly
Photogrammetry has been around for years and it's always been a compromise – great from the angles you shot, strange from anywhere else, and enormous to download.
Gaussian splatting has largely solved the "strange from anywhere else" problem, and the work now going into it is about size. Gracia AI showed a 4D version at AWE 2026 – volumetric video rather than a still space – with compression claimed to cut bandwidth by over 90%.
Ninety percent is the number that matters. Realistic captured spaces have always looked great on a workstation and been unusable over a venue's Wi-Fi. Compression is what moves this from a research demo to something you can put behind a link.
Why you'd care: it's the direction of travel for property, venue and site walkthroughs. It doesn't replace what we do with 360 virtual tours today – that's a mature, reliable, affordable way to put a space online right now – but it's what the premium end of that starts to look like.
4. The headset understanding the room
The quieter change is AI running on the headset itself rather than in a data centre.
Meta's Horizon OS 2, which started rolling out in February 2026, added offline on-device voice dictation on the Quest 3 and 3S – speech to text with no connection required. The same update lets windows snap to walls the headset has detected on its own.
Individually, small features. Together they're a pattern: the headset is starting to understand the physical room and the person in it without phoning home. That matters for speed, for privacy, and for anywhere with bad Wi-Fi – which, if you've ever exhibited at a convention centre, is most places.
What's still hype
There's a lot of it about, so here's where I'd be sceptical.
"Describe your world and AI builds it." You can generate objects well. Generating a coherent, well-lit, navigable, performant space with a purpose is still design work done by people. The demos are real; the finished products are not.
AI as the reason to do VR. If a proposal leads with AI rather than with what your audience will do in the headset, be wary. AI is a cost and capability input. It isn't the reason anyone puts a headset on.
The metaverse, again. Worth noting that Meta removed Horizon Worlds from VR headsets entirely in June 2026, moving it to mobile. The company that spent the most money on the social-VR-world idea has moved on from running it in a headset. Treat "persistent AI-populated worlds" pitches accordingly.
So what does this mean if you're a business?
Practically, three things.
Custom VR costs less than the quote you got two years ago. If you priced a branded VR experience in 2023 or 2024 and put it in the too-hard basket, the number has moved. That's mostly asset generation and partly the shift to browser-based delivery.
Conversation-based training is now worth a look. Procedural training in VR has been solid for years. Conversational training has only recently become properly useful.
Nothing about the hardware got harder. The headsets are the same ones you'd hire for a party. None of this requires a new device, a subscription, or a platform decision.
Frequently asked questions
Has AI made VR development cheaper?
Yes, mainly through generative 3D asset creation. The expensive part of a custom VR project was always modelling every object in the scene by hand. Generative tools now handle the bulk of background assets from a text description or reference image, which removes a large share of the labour from a typical build.
Can AI generate a whole VR environment automatically?
Not to a usable standard. AI generates individual 3D objects and textures well. Assembling them into a coherent, well-lit, navigable space that performs on a standalone headset and serves a purpose is still design work done by people. Demos that appear to do this rarely survive contact with a real brief.
What are AI characters in VR training?
They're conversational agents that replace scripted dialogue trees. Because responses are generated rather than pre-written, a trainee can't memorise the correct path, and the character can respond to tone and phrasing. This makes them useful for de-escalation, complaints handling, sales objections and interview practice.
What is Gaussian splatting?
It's a technique for reconstructing a real place in 3D from photographs or video that holds up when you move away from the original camera positions, which traditional photogrammetry struggles with. The main barrier has been file size, and recent work has focused on compression – Gracia AI demonstrated a 4D version at AWE 2026 claiming over 90% bandwidth reduction.
Does AI in VR need an internet connection?
It depends what's doing the work. Generated assets are baked into the experience before it ships, so they need nothing at runtime. Conversational AI characters generally need a connection. Some capability now runs entirely on the headset – Meta's Horizon OS 2 added offline on-device voice dictation for the Quest 3 and 3S in February 2026.
Do I need a new headset to use any of this?
No. The Meta Quest 3 and 3S run all of it, and they're the same headsets used for events and hire. There's no separate AI device, subscription or platform commitment involved.
Is the metaverse still a thing?
Not in a headset, in the form it was originally sold. Meta removed Horizon Worlds from Quest headsets in June 2026 and moved it to mobile. Standalone VR apps and experiences are unaffected and continue as normal – it's the persistent social-world idea specifically that has been scaled back.
Should AI be the reason we do a VR project?
No. AI is an input that affects what a project costs and what it can do. The reason to do a VR project is what your audience will do in the headset and what you want them to take away. If a proposal leads with the AI rather than the experience, that's worth questioning.
Thinking about a project?
If you looked at custom VR a couple of years ago and the number didn't work, it's worth another conversation – quite a lot has moved since. And if it still doesn't stack up, we'll tell you that too.
Contact page, or email info@virtualrealityhire.co.nz.