How Korea’s FuriosaAI and Fal’s $1.3B Round Prove AI Inference Is the Next Battleground

South Korea’s FuriosaAI has just made the boldest play yet in the global AI inference chip war. In a Series D round that closed this fall, the startup raised $620 million, quadrupling its valuation to $2.9 billion in less than a year. The funding, led by Infinitum Partners and Korea’s National Growth Fund, positions FuriosaAI as a serious challenger to Nvidia in a market where power efficiency—not just raw speed—is becoming the deciding factor.

The money reflects a simple but critical insight: inference, the step where trained AI models generate answers for users, is where data centers now spend the most. FuriosaAI’s RNGD chip, which entered mass production in January, claims to deliver twice the users per kilowatt as a rival Nvidia card while matching its throughput. That efficiency pitch isn’t theoretical—it’s already powering deployments at Samsung SDS, LG Group, and Upstage, though all remain within Korea’s tech ecosystem. Breaking into global data centers will require proving it can scale beyond this home market, where state-backed investors and local enterprise demand are fueling its growth.

The round’s scale—one of the largest ever for a private chip company—also highlights a broader shift: investors are betting on credible challengers to Nvidia, even if the hardware isn’t yet deployed at hyperscale. FuriosaAI’s production targets (20,000 cards this year, doubling next) and its lock on fab capacity for advanced designs signal confidence in its roadmap. But the real test will be whether its efficiency claims hold up against Nvidia’s dominance in cloud and enterprise data centers, where vendor lock-in and software compatibility often outweigh hardware specs.

In a Series F round led by Bessemer, the company raised $1.3 billion, doubling its valuation to $16 billion in seven months. Fal’s specialty isn’t chips but infrastructure: it hosts and runs generative AI models for video, image, and audio, serving customers like entertainment studios and game developers. Its revenue jumped from an estimated $400 million annualized in March 2026 to $800 million by mid-year, a pace that forced it to acquire more compute capacity just to keep up.

What these two companies share is a focus on *specialization*—whether in hardware (FuriosaAI’s power-efficient inference chips) or software (Fal’s media-focused API). Both are betting that general-purpose AI infrastructure, dominated by Nvidia and cloud providers, can’t meet the rising demand for optimized, cost-effective inference. Fal’s $1.3 billion round included Nvidia as a strategic investor, a nod to the reality that even the most specialized players still rely on GPU hardware.

The tension between hardware and software specialization is the defining question of AI’s next phase. FuriosaAI’s valuation surge suggests that inference chips, long overshadowed by training accelerators, are now a high-stakes battleground. But its deployments remain concentrated in Korea, a market where state-backed capital and local demand can distort global relevance. Fal, by contrast, has proven that even a software-defined play can command eye-watering valuations if it solves a critical bottleneck—compute capacity—for a high-growth sector.

What to watch: FuriosaAI’s global expansion: Can it secure deals outside Korea, where Nvidia’s ecosystem is entrenched? Fal’s compute constraints: Will its revenue growth outpace its ability to secure chips, or will it become another victim of the AI hardware crunch? The inference hardware arms race: If FuriosaAI’s claims hold, expect more startups to test Nvidia’s dominance in data centers.

The message is clear: inference isn’t just the next frontier—it’s the current one. And the companies winning won’t be the ones with the flashiest training chips, but the ones that master the art of answering questions without breaking the bank.

Sources: gcn.com · gcn.com · marketwatch.com