Artificial intelligence inference startup Fireworks AI Inc. has acquired Hathora Inc., a real-time compute and server orchestration platform focused on multiplayer games and AI inference.

Chief Executive and co-founder of Fireworks AI Lin Qiao told SiliconANGLE in an exclusive interview following last week’s announcement that the company sought out Hathora not for its gaming customer book, but as a talent-and-infrastructure acquisition. The purchase is meant to strengthen Fireworks’ existing global compute orchestration layer for inference and training.

Qiao made a deliberate connection between gaming infrastructure and the next phase of AI infrastructure. Most gamers will forgive certain trade-offs when playing, such as slightly reduced graphics or occasional frame drops, but the moment there is lag or gameplay is disrupted by delays, they notice.

“That discipline, the obsession with every millisecond and every routing decision, is exactly what AI inference needs,” Qiao said.

Hathora introduced itself in 2022 and officially launched its platform at the Game Developers Conference in 2023. In the intervening years, it built a container orchestration platform spanning 14 regions, two bare-metal providers and four clouds. It powered server infrastructure for live gaming titles including Splitgate 2, Stormgate and Predecessor, then recently expanded into real-time AI workloads with a voice model marketplace.

The velocity at which the industry is changing is driving delivery of AI models, both inference and training, to new heights.

“This acquisition is helping Fireworks accelerate our development of a global computation platform … our software architecture has a compute layer,” Qiao added. “We require global orchestration, routing, disaster recovery, auto scaling, very low-latency and high-performance setup.”

From compute orchestration to agentic AI

Acquiring Hathora is part of Fireworks’ broader through-line for how training and inference will evolve as agentic AI becomes the norm. Right now, most architectures run atop large language models as “brains,” but they require extensive upfront training, which is expensive and then remains static for long periods before the next training cycle.

“Right now, we are just heavily indexed on text-based gen AI,” added Qiao. “But in the future, I think multimodality is how eventually it’s going to land and with those agents starting to thrive. And agent talk with agents; the velocity of those interactions is going to pass how human talk with humans … and then speed is even more critical.”

The bigger thesis is that this could eventually allow Fireworks to redefine its position in the market. Qiao said Fireworks is not just an inference provider; it also delivers pretrained models, post-training capabilities and an automation layer, with inference serving as the visible output of customization rather than the entirety of the business.

The strategy is that the most valuable data is trapped in private vaults and its application specific, so the future will not belong to a single universal or generic model, but to “millions of models,” each constantly and continuously customized to fit purpose and use case.

She distinguished this from the current trend toward AGI, or artificial general intelligence, the so-called “holy grail” of AI in which AI could match or surpass human intelligence. In the case of custom, use-case-specific AI models, each model would be continuously tuned to its particular role so that it always had an expert understanding of its persona, rather than acting as an extremely large generalized model serving as an “everyman.”

The first company that proves itself capable of continuous fine-tuning of models and then providing inference on these models will be a game-changer. That means constant orchestration and low latency at scale. Fireworks intends to be that game changer and that means bringing on talent and infrastructure from a firm such as Hathora.

“We actually are automated customization,” Qiao said. “That’s what we’re building, not just inference. Inference is basically for us to show results of all this customization.”

Image: SiliconANGLE/Microsoft Designer