Infinite interactivity: the promise and challenge of GenAI video games

As generative AI hype has swept across the tech industry, game developers are seeing the potential to create whole new types of experiences for players.

“We can do games that were not possible before,” said Ramón Iglesias, founder of San Francisco-based AI games studio Clementine. “GenAI just gives you a new design space to work with: that's where the excitement comes from.”

He and others working in the space are using generative AI to develop experiences that they hope will deliver more interactive, adaptive and engaging virtual worlds. The potential is significant: the global market for AI in games predicted to grow from $3.3 billion in 2024 to $51.3 billion by 2033.

But these infinitely interactive worlds come with a fundamental problem: every player interaction costs real money. The economics are stark: ongoing inference calls to cloud-hosted models turn each player action into a billable expense. That’s forcing developers to explore new business models, while also trying to prove that the technology can really help create experiences that people actually want to play.

“You can think of it as vibe coding your monster”

LA-based Jam & Tea Studios’ first game, Retail Mage, casts players as wizards working in a “mage mart,” where their job is to service magical customers, winning as many five-star reviews as possible.

“It's kind of late-stage capitalism comes for Diagon Alley,” said Aaron Farr, co-founder and chief technical officer at Jam & Tea.

What makes Retail Mage different from traditional job simulator titles is that it uses GenAI models to let players interact with objects in unlimited ways. Where a player might have previously been able to pick up a book and read it in a traditional game, in Jam & Tea’s world, a player can do pretty much anything they like with it: rip out a page to write an IOU note, or use the book to stand on to reach a high shelf.

“In our game, players can type or say what they want to do, and we use machine learning to translate that player intent into ways we can represent that in a three-dimensional game world. That's really where a lot of the magic of the game happens,” said Farr. “It’s like the simulation is playing back at you.”

Clementine, meanwhile, is developing a game called Allyon, where players use natural language to train Pokémon-like monsters that respond to commands in a more realistic and less limited way than was possible before GenAI.

“From a tech perspective, we're converting natural language into code in real time, and that is what allows us to give lifelike traits to the monster: it feels like it understands what you're saying, and reflects that in the world in a very open-ended way,” said Iglesias. 

Making inference make sense

While developers are hoping players will be delighted by these kinds of limitless interactive experiences, which Farr describes as giving people “infinite tactical agency,” they also don’t come cheap.

Both Jam & Tea and Clementine’s games rely on cloud-based AI models, meaning that every time a player triggers an event that requires a GenAI response, it creates an inference call that costs money. Studios need access to a flexible compute provider that can deal with spikes in player activity, and that they’re grappling with a new business model for games.

“The availability of that compute is much better today than it was, but it can still be an issue, and the expense of that compute is significantly higher [than is required for regular online games],” said Farr. “Those costs have to come out of studios’ same budget effectively, which is already a challenge: it's not like we had a lot of extra budget sitting there.”

Jam & Tea has developed proprietary machine learning technology to reduce the costs of inference calls, while Clementine has fine-tuned and distilled models that are smaller and cheaper to run.

“We still have to push the server-side costs down quite a bit to make this happen, and we think we can make it happen in a reasonable timeline,” said Iglesias.

But while studios are optimizing their tech to make their games cheaper to run at an inference level, Farr believes that GenAI games might need to be financed differently, with a possible return to subscription-based models that underpinned early online games like World of Warcraft.

“Will we see a return of subscriptions? Maybe what we see is players are willing to essentially pay for intelligence,” said Farr. “I think it's a good challenge for the industry: are we creating games that are novel enough that players are willing to pay for them in a different way?”

The big players enter

Some GenAI game developers are pushing for a different approach, aiming to get their titles to run on-device, thereby swerving inference costs.

San Francisco-based Studio Atelico is developing a card-battling game where players use GenAI to create new cards with new abilities on the fly. Co-founder and CEO Piero Molino said that his team has built “small, specialized models” that mean the game will be able to run on-device in “the vast majority of cases”.

But even Molino, who plans to license Studio Atelico’s technology to other developers wanting to build GenAI games that can run primarily on the edge, said that some complex game mechanics and multiplayer experiences will need to run in the cloud.

Massive studios like Activision Blizzard (Call of Duty) and Supercell (Clash of Clans) have announced plans to experiment with GenAI to supercharge content creation, signaling broader industry interest. If ambitious developers looking to push GenAI to its limits are successful in making the numbers add up, we might also be about to enter a world where games are more immersive and engaging than ever.