"Earlier models could decorate. This class can design and fully implement"

Welcome back to AI Gamechangers, the newsletter where we talk to the people across the games industry doing practical, inventive things with AI. Each edition is a conversation with someone building at the sharp end.

This time we’re with Daniel Tabar and Branislav Siles, the co-founders behind Atomontage and its Virtual Matter tech. We get into what it means to rebuild an entire platform so that making a multiplayer game is as easy as talking to an AI, and why “no install, no sign-up” is a bigger unlock than it sounds.

A quick personal note before we get into it: the AI Gamechangers editor, Dave Bradley, has co-written a book with Kelly Vero, and it’s published by Bloomsbury in print on 3rd September. Myth-Made Machines is about how AI has been imagined in books, films, TV, games and philosophy over the centuries, from Greek myths and Gulliver’s Travels to Isaac Asimov and Detroit: Become Human. Find it here, and use code GLR BD8 for your 20% discount, or look for it in your preferred bookstore. Enjoy!

Daniel Tabar and Branislav Siles, Atomontage

Meet Daniel Tabar, co-founder of Atomontage and a game-maker with 25 years’ experience, from a teenage stint at NASA Ames to creating Cortex Command, the cult PC game that helped pioneer Early Access. He’s joined by co-founder Branislav Siles, who has built the Atomontage engine since 2004.

Their Virtual Matter tech makes worlds from tiny volumetric samples that persist, react and remember what players do. They recently rebuilt the whole platform so anyone can make, mod and remix multiplayer games by talking to an AI, in a phone browser, in seconds.

Top takeaways from this conversation:

1. AI lowers the barrier to entry but raises the ceiling, shifting the craft from implementation to creative direction.

2. A model-agnostic harness sends your edits to cheap, fast models and escalates only the most complex builds to the frontier models. When Anthropic briefly withdrew Fable, they simply rerouted and makers kept working: no single model is load-bearing.

3. Atomontage’s Virtual Matter tech means that user-generated content is shareable and playable instantly in a phone browser with no install or sign-up.

4. CPU-only servers keep play nearly free, with GPU cost metered only during creation.

AI Gamechangers: What can someone do with your platform today that they couldn’t when you showed it to journalists at GDC 2024?

Daniel Tabar: At GDC 2024, we showed the core medium: detailed, persistent worlds streamed across devices. Today, friends and colleagues can make, modify, and remix a multiplayer game through conversation, directly in a phone browser.

Anyone can tap a link on a phone and enter a multiplayer world in seconds, with no install or sign-up. That world may itself have been built from prompts in minutes. They can play it, change it while playing by talking to an AI agent, and remix the whole game into their own account to keep building it the way they want.

The next step is standalone export to itch.io and Steam, which we expect to enable later this year. There, the unique selling points of Virtual Matter will help these games really stand out.

Something else that is new: Every playable world currently published on the platform, including its art, game logic and physics behaviour, was made without a human writing a line of code. Two years ago that was a research demo. Now it’s a link you can tap and dive into immediately. To be clear, the AI assistance is totally optional - you can hand-craft the code and content as well if you wish.

AI Gamechangers: Why rebuild the whole platform?

Daniel Tabar: The old platform was built for a different era: desktop-first, high-touch onboarding, limited self-serve creation. It was not an end-to-end creation system for a non-technical maker. Once we understood that making and modding had to be as effortless as playing, in a phone browser, retrofitting the old platform was never going to get us there.

The second reason is that agentic AI changed the economics and viability of rebuilding. A small squad built the entire production platform codebase (not the voxel engine inside) from scratch in months, which would have taken us years the old way. We’re a live case study of our own thesis: tight teams with the right substrate and the right agents can do what used to need a sizeable startup or studio.

AI Gamechangers: What is Virtual Matter in plain terms, and what does it do that other 3D techniques and world models can’t?

Daniel Tabar: Imagine Minecraft’s blocks, but literally a million times smaller by volume. Everything in the world is made of tiny volumetric samples, so worlds have insides. You can dig into anything, break anything, grow anything, and it persists, multi-maker, multiplayer, streamed to a phone browser.

But the plainest way to say it is this: the world systemically reacts to and remembers you.

There’s a principle that goes back to Half-Life, that if the world ignores the player, the player stops caring about the world. Studios since have spent enormous money hand-authoring reactions to the handful of actions they could anticipate. Everything else the player does leaves no trace. You unload a rocket launcher into a wall in most games, and the wall is unbothered.

Virtual Matter acknowledges everything, including what nobody planned for, because the record isn’t a script; it’s the Matter itself. An entire ecosystem and life cycle of plants and animals can emerge from very simple behaviours like growing, foraging, hunting, eating, pooping, and breeding. Come back tomorrow, the world has evolved in surprising and organic ways. You can read its history directly from the half-eaten corpses and signs of a struggle.

“The craft shifts from implementation to direction. Systems design becomes tuning emergence you can feel in minutes rather than days. Two years ago we expected makers to learn our Lua tools. Natural language is now the default entry point”
Daniel Tabar

That combination is difficult to achieve with the dominant alternatives without layering on separate systems. Polygon models are fragile hollow shells that need expert tooling and have hard limits on detail. Gaussian splats are beautiful but baked and ethereal: you can’t collide or simulate against them alone. Signed distance fields either don’t capture fine detail, or become expensive when they do. World models generally lack durable, inspectable state, so it’s hard to motivate building anything in them.

Virtual Matter remembers all the materials and details, at depth, until things are altered again.

AI Gamechangers: Why is mobile browser, no install, no sign-up a real unlock, and what did it take technically?

Daniel Tabar: Mobile’s dirty secret is that the install is one of the biggest drop-off points. Every share of a game normally costs the recipient an app store visit, a download, an account. Our worlds travel as links directly into the browser: someone drops a URL in a group chat, and their friends are inside the same world in seconds, on whatever phone they have. That turns every Discord server, WhatsApp group, and iMessage thread into a distribution channel we do not have to build.

Technically, it took a decade of compression R&D, a full WASM client, and browser restrictions finally lifted enough by the mobile platform owners who prefer everything to go through their app stores. Virtual Matter streams progressively, like 2D video does, but with our 3D data, so the world starts playable before it finishes arriving. Our servers are CPU-only, which is what makes instant multiplayer economically scalable.

AI Gamechangers: Does CPU-only server-side still hold with an AI agent involved?

Daniel Tabar: Yes, with a clean separation. The worlds themselves, simulation, streaming, multiplayer state, still run entirely on CPU-only bare metal. That claim holds, and it’s core to our cost structure. Once standalone export launches, Atomontage no longer hosts that title; makers can run their own dedicated servers.

The AI agent is a separate workload: frontier model APIs, plus high-end inference hardware Dell and NVIDIA sent us after seeing the demos, which we use to cut cloud AI costs through routing to open models. So the expensive GPU work happens only when someone is creating with the agent, metered by credits, while playing a world stays as cheap as it ever was. Keeping those two cost curves separate is deliberate: play has to be nearly free for this to work at scale.

“Fable 5 raised the ceiling on the deep work: long-horizon, multi-step building where the agent plans a whole game system, writes it, tests it in the world, and iterates without a human unsticking it”
Daniel Tabar

Any power maker can also use their own AI subscriptions and harnesses to build Virtual Matter worlds through our native client on whatever platform they prefer: Windows, Mac, Linux. Then they can take over and use their subsidised subscription plans from the model providers directly, and don’t have to touch our credits and associated costs at all.

And again, it’s also completely optional to use any of the AI assistance in making content or code - powerful editors are right there if you prefer to craft everything by hand.

AI Gamechangers: Can you share any examples built, and any numbers?

Daniel Tabar: The rebuilt browser platform soft-launched with half a dozen playable multiplayer worlds, all made without a human touching code: an FPS with an Easter-egg-hunt twist, a self-playing RTS title where everything is destructible, an ecosystem sandbox where you can grow plants to feed bunnies that then feed predatory cats, and more. My favourite is a gritty 1-on-1 FPS rooftop brawler where every shot leaves a bullet hole that’s still there next round, so the arena you fight in is one your earlier fights built.

The most honest test so far wasn’t ours. At a recent team sprint, a colleague’s kids, 11 and 13, built a working Formula-style racing game and a physics-driven alien-invasion game with no prior experience of the platform and only minimal guidance. When things broke, we could fix them on the spot. That’s exactly what this phase is for.

We’re deliberately gating access in waves where we’re hitting our daily sign-up caps, and the cohort is still too early for meaningful public metrics. The numbers we care about right now are retention and conversion among early makers, not headline sign-ups.

AI Gamechangers: Why Fable 5 specifically, and what can a player do with a prompt mid-game that earlier models couldn’t?

Daniel Tabar: When we last spoke, AI could help import an object into Virtual Matter. Today it can plan, implement, test and revise a complete game system inside the live world - both content and code.

Our harness is model-agnostic by design: different jobs route to different models. But Fable 5 raised the ceiling on the deep work: long-horizon, multi-step building where the agent plans a whole game system, writes it, tests it in the world, and iterates without a human unsticking it.

“Virtual Matter consists of ordered, compressed volumetric samples, a structure generative models handle well, much as they handle pixel grids and audio samples. That is one reason images and audio became AI-native earlier than interactive 3D. We think of it as giving interactive 3D the file format the AI era was missing”
Branislav Siles

Now Opus 5 does certain things even better than Fable, and the various flavours of GPT-5.6 also have their pros and cons. The landscape is shifting by the week, so it’s important to have a flexible system that can balance quality, cost, and speed for the best building experience.

Concretely, the ability to quickly construct a game world out of voxel primitives is a tangible difference between the models. Fable was a big jump, and Opus 5 recently took another step in spatially understanding and crafting, like a detailed house with a chimney, windows, windowsills, and every other meaningful detail.

Then, truly detailed pieces can be made with generative models that produce 3D meshes from text prompts or images that we finally turn into Virtual Matter, and our agent does all that end-to-end as well. It even cuts up a character into segments and animates the parts for various walk cycles, jumping, crouching, etc. This wasn’t possible just months ago.

What that means mid-game: a player can say “make this dinosaur decompose over time and grow plants my bunnies can eat” and watch it happen in the live world, behaviour, art and simulation together. Earlier models could decorate. This class of model can design and fully implement. Quick edits still route to lighter models, which keeps most interactions fast and cheap.

AI Gamechangers: Why premium models rather than smaller, cheaper ones, and how do you keep latency and cost sane? And what if Anthropic withdraws Fable again?

Daniel Tabar: Quick edits route to fast, cheap models. Only deep building escalates to the frontier tier. Credits are priced to track underlying compute cost, which lets us adjust routing and pricing as model economics change, and a growing library of reusable, community-rated artefacts means more requests get served from that content cache. This lets the system reuse and mod proven components rather than generate everything again.

Because the harness is model-agnostic, we rerouted, and everyone kept building. No single vendor is load-bearing for us, and open-source models provide credible fallback over time.

AI Gamechangers: Why are ordered, progressively streamed samples better for AI than meshes or Gaussian splats?

Branislav Siles: JPEG was a breakthrough compression format with its progressive levels of detail (for those of us old enough to remember the Netscape era). Standard image formats like it created a huge, interoperable corpus of pixel grids long before generative models arrived. Interactive 3D never developed an equally consistent, editable representation.

Meshes can have different levels of detail, but creators usually have to build or generate separate versions ahead of time, and their triangles still have no single order an AI can rely on.

“Mobile’s dirty secret is that the install is one of the biggest drop-off points. Every share of a game normally costs the recipient an app store visit, a download, an account. Our worlds travel as links directly into the browser: someone drops a URL in a group chat, and their friends are inside the same world in seconds, on whatever phone they have”
Daniel Tabar

Gaussian splats can look very realistic, but they mostly store how something looks. They do not naturally store what is solid, what is inside an object, how its parts connect, or how it should behave when it moves, breaks, or collides.

Virtual Matter consists of ordered, compressed volumetric samples, a structure generative models handle well, much as they handle pixel grids and audio samples. That is one reason images and audio became AI-native earlier than interactive 3D.

Because the format stores volumetric material data rather than only a surface, interiors and composition remain explicit through edits and simulation. It avoids relying on separate high-resolution meshes, textures, and collision representations, and can stream the detail each device needs and can handle. It also matches how people naturally reason about physical objects: volumes made from materials. We think of it as giving interactive 3D the file format the AI era was missing.

AI Gamechangers: You frame creation as vision and taste with no tech or art skills needed. What does that mean for professional developers, and how does AI-native creation change the craft?

Daniel Tabar: I’d flip the framing. Yes, the floor drops: someone with no tools background can make something real, the way phones made everyone a photographer. But phones didn’t end professional photography. They made taste the differentiator, and the pros with taste thrived.

However, Virtual Matter raises the ceiling even more than it lowers the floor. A professional designer can test dozens of game mechanic ideas in an afternoon that used to require separate implementation cycles, in a medium with genuinely unexplored design space. There’s a whole category of design problems that mostly disappear: you no longer have to anticipate and hand-author every reaction, because the world records what players do by default. For decades, game budgets have gone into hand-authoring signs that the world noticed the player. Getting that acknowledgement by default changes what designers spend time on.

The craft shifts from implementation to direction. Level design becomes shaping possibility spaces. Systems design becomes tuning emergence you can feel in minutes rather than days. The professionals who treat this like early Unreal, a new medium to master before anyone else has, are going to make the defining games of it. That is why our next key hires include world-class artists and designers: as implementation gets cheaper, creative direction becomes a larger constraint.

In short, two years ago we expected makers to learn our Lua tools. Natural language is now the default entry point, while professionals can still exercise deep and complete control.

AI Gamechangers: What’s the business model?

Daniel Tabar: Three layers, in the order they arrive.

Today, makers pay a subscription plus metered AI credits. Playing remains free; credits cover the model calls and asset generation that create variable compute cost. When makers export standalone games, we take a small engine royalty only after the game begins earning significant revenue, so our upside is tied to theirs - modelled on Unreal Engine’s royalty structure.

“Every playable world currently published on the platform, including its art, game logic and physics behaviour, was made without a human writing a line of code”
Daniel Tabar

At platform scale: a marketplace where makers sell and remix each other’s work and players spend inside worlds, with the platform earning a share of transactions. The long-term shift is toward participating in maker success, while subscriptions and metered creation cover the compute-heavy work today.

AI Gamechangers: How much have you raised, and are you raising again?

Daniel Tabar: We’re backed by Makers Fund, Hartmann Capital, plus a group of angels who built this industry: Cliff Bleszinski, key creative figure behind Unreal and Gears of War; Hilmar Veigar Pétursson, CEO of Fenris Creations (formerly CCP Games); Tommy Palm, CEO of Resolution Games and of Candy Crush fame, among others.

As PocketGamer.biz reported in 2024, we had raised $4.5 million at that point. We are not disclosing an updated total today, and we can’t comment on fundraising activity outside formal announcements.

Our focus right now is the soft launch and proving that makers return, publish, and keep building.

AI Gamechangers: What do the next six to 12 months look like for you?

Daniel Tabar: Three things. First, widening the waves: we onboard makers in controlled batches, learn from each one, and open further as the platform hardens.

Second, we expect to enable standalone itch.io and Steam exports later this year as product readiness and IP work conclude. We will test the pipeline with our own sample titles: prototype rapidly on the platform, then validate which concepts attract players before investing in full production.

Third, the ecosystem: the first games made entirely by outside makers, a growing community on our new Virtual Matter Discord, and an asset library that makes every next game faster and better to build than the last.

Readers can start with a blank world at VirtualMatter.dev, or try the Lovable-like prompt-first experience at VirtualMatter.ai to make their game ideas come alive today!

Why not subscribe? It’s free, and you’ll receive interviews with industry insiders and studio leaders in your inbox. We have a number of conversations underway, including some important meetings at Gamescom, so watch this space - up next, TILKI.

Further down the rabbit hole

Some useful news, views and links to keep you going until next time:

• The Mobile Games Awards are revealed live in Cologne next week. There is an AI category. Voting has long since closed, but you can see who’s in contention here.

• Luke Dicken, who ran Take-Two’s AI team until it was dissolved in April, argues nobody has actually measured what generative AI returns for games businesses. Speaking to The Side Quest, he says, “it’s not the silver bullet that the headlines and the press releases want you to believe it is.”

• A former Icelandic minister has joined Fenris Creations (the EVE people) to lead AI partnerships with universities, laboratories and policymakers.

• Shortly after its $19.1 million funding round, Liquidnitro hired ex-Tencent Games CTO for AI-enabled game production.

• The Sandbox launched a second AI game jam on 12th August, opening its new AI-led Studio tools to more creators after a boutique alpha in July. LAND NFTs won’t be required to build experiences, but will unlock premium publishing perks and a bigger revenue share when Studio launches properly in Q4 2026.

• David Kaye, founder of F4 Fund, isn’t mincing words on LinkedIn: “Do not pitch me unless AI is absolutely at the core of what you are doing.” Ignoring AI, he argues, is now disqualifying for VC funding.

• Saber Interactive is embroiled in an AI writing row after former lead writer Stella Sacco said she’d been replaced by ChatGPT. Saber denied replacing writers outright, while confirming AI is used for dialogue in an “experimental” mode. CEO Matt Karch’s combative response to Sacco hasn’t helped matters.

• Cascadeur 2026.2 is out, with Additive Animation Layers arriving in alpha after being one of the tool’s most-requested features. There are also new Easing controls for interpolation, better collision cleaning, and Chinese, Japanese and Korean language packs. We interviewed the CTO in May.

• Razer’s betting big on AI gaming intelligence, teaming up with NUS Singapore on a new research lab purpose-built for real-time, adaptive gameplay.

• Look out for Myth-Made Machines: The Stories Behind Artificial Intelligence, available now as an ebook and in hardcover from 3rd September: