The Runtime Gap: What 25 Years Inside Game Engines Taught Andrew Bowell About Where AI Actually Matters
Andrew Bowell spent 15 years at Havok and a decade at Unity, where he was the first product hire and built the product organization that captured 50%+ market share ahead of their $13.7B IPO. He now runs Iconic, a startup building engine technology for AI-native interactive experiences.
We sat down with him to talk about what the industry gets wrong about AI in games, why on-device inference changes everything, and whether players actually want any of this.
Here are three things that stuck.
1. The AI paradigm shift in games isn't in production.
The gaming industry's AI adoption story so far is almost entirely about making things faster and cheaper on the development side. Asset generation, code assistance, dialogue pipelines. Bowell's argument is that this is productivity tooling, not a paradigm shift, and that the real untapped surface is at runtime, in the player experience itself.
What is runtime? Fair question.
In game development, runtime is everything that happens while the player is actually playing the game. It's the live, real-time execution layer. The contrast is with build-time or production-time, which is everything that happens before the game ships, such as writing game code, creating art assets, designing levels, recording dialogue, testing etc.
His framing is worth paying attention to. Havok rode CPU power to make games feel physically real. Unity rode accessible GPU and mobile hardware to make game development accessible. The next wave, he argues, rides machine learning compute to make game worlds dynamic and responsive in ways that pre-scripted systems can't achieve.
The specific capability Iconic is chasing is conversational AI characters with memory, persona, and narrative coherence, running live during gameplay. This is not designed as something studios can bolt on to existing titles, but as a design primitive that the game is built around from day one.
Bowell is blunt about this distinction. Integrating voice into existing games as an add-on doesn't work nearly as well as designing the game around the capability from scratch. The way I read it, he encourages studios to go AI native. Dive in. Not dip their toes.
This is a useful filter for evaluating every "AI in games" pitch you see. Ask whether the AI is making the production pipeline faster, or whether it's changing what the player actually experiences. The former is valuable and measurable but incremental. The latter, if it works, creates a new category. And that’s a big if.
2. Who pays for the AI when a million players play at once?
Bowell frames on-device AI as the single most important technical problem Iconic is solving, and the reasoning is almost entirely economic.
Run an LLM in the cloud for a game that hits a million concurrent users, and the token costs become existential. Add privacy concerns around voice data leaving the device, model drift as foundation models update underneath your game, and latency requirements for real-time conversation, and the cloud path gets really ugly, really fast.
Iconic's first title, The Oversight Bureau, runs its full AI stack on a Steam Deck. That required aggressive constraints, fine-tuned small models, and a trade-off with the art team over GPU allocation. But the result is a game that has zero ongoing inference costs per player and zero dependency on external model providers.
The Oversight Bureau | Official Announce Trailer
The hardware reach problem is real, though. Only about 30% of Steam users currently have a GPU capable enough (roughly a 4080) to run the full stack. Bowell acknowledges this openly and points to hybrid on-device/cloud architectures as a bridge, but his conviction is clear: the long-term winner in AI-native gaming will be whoever cracks performant on-device inference at scale.
What Bowell means is that right now, not everyone's hardware is powerful enough, so you might need to split the work. Run some of the AI on the player's device and offload the rest to the cloud, which leads to existential problems if the game gets lots of players without monetizing well enough to cover the token costs. But long term, the company that figures out how to run the full AI stack on the player's own machine, fast enough and cheap enough, wins.
The implication for studios evaluating AI integration is stark. If your AI features depend on cloud inference (Inference is when an AI model takes an input and produces an output), your cost structure scales linearly with player engagement. That's a game-as-a-service cost model bolted onto what might otherwise be a premium title. It constrains your business model before you've even shipped.
So the way I understood it, holistic AI integration works only with live service games that monetize through microtransactions. That is the only way to cover inference costs until someone comes up with how the game’s AI stack can be run on the player’s computer.
3. Is Iconic also destined for the game engine graveyard?
Stadia built custom tech with Google's infinite resources and died. Amazon spent over a billion on Lumberyard. Improbable raised $855 million for a metaverse engine and went quiet. The track record of "experienced people build a new game engine" is genuinely terrible. And the success of Unity and Unreal Engine is debatable, with both companies being scaled up by auxiliary businesses rather than their engine offering.
Bowell doesn't dodge this. His counter is positioning is that Iconic isn't trying to compete with Unity or Unreal on rendering, physics, or general-purpose game development. Those are mature, million-line codebases with massive ecosystems. Instead, Iconic is building a component layer, an AI runtime and authoring toolset, that sits on top of existing engines.
The strategy is closer to what Havok did with physics middleware than what Unity did as a full engine. Solve one hard problem extremely well, prove it works by shipping your own games on the technology, then license it to studios that don't want or can’t build it themselves.
Whether this works depends on two things Bowell’s team can't fully control yet:
First, whether AI-native game design produces experiences that players actually seek out. The Oversight Bureau got positive Steam reviews despite mandatory AI disclosure, which is encouraging. But one indie title is not market validation.
Second, whether the on-device inference stack can reach sufficient hardware breadth before the funding runway ends.
The honest read on Iconic is that the thesis is sound but unproven at commercial scale, which is exactly what you'd expect from a pre-revenue startup. The technology is truly interesting. Yet the question is whether the market window is open long enough for a compact and fully remote team to prove it before the incumbents absorb the same capabilities into their engines.
If you got excited about Iconic’s mission and want to join the frontier of AI game engine development, Iconic is hiring. The team is welcoming current and future AI programming icons.
Andrew Bowell is the founder of Iconic and previously led product at Unity from its earliest days through IPO. Iconic is currently hiring in London. You can find The Oversight Bureau on Steam.
