The most expensive part of an AI model is not the training, but the serving. It’s the steady, kilowatt-hour-consuming hum of inference that turns a data center into a furnace. In South Korea, a new startup called AnyBridge is betting that the best way to cool that furnace isn’t a better GPU, but a smarter piece of software that treats every chip in the rack as a potential contributor.
Founded in 2026 by KAIST professors Jongse Park and Youngjin Kwon, AnyBridge is building what it calls a hardware-agnostic runtime and control plane. The idea is simple in concept, devilish in execution: create a software layer that can dynamically allocate AI inference workloads,whether for large language models or autonomous agents,across a heterogeneous mix of GPUs, neural processing units (NPUs), processing-in-memory (PIM) chips, and whatever comes next. The company claims this approach can cut inference costs by 40% to 70% compared to a GPU-only setup [Perplexity Sonar Pro Brief, retrieved 2026].
The academic wedge
The company’s origins are firmly in the lab. Both founders are professors in KAIST’s School of Computing, and CEO Jongse Park co-authored a 2026 research paper titled “LLMServingSim 2.0: A Unified Simulator for Heterogeneous and Disaggregated LLM Serving Infrastructure” [AI Post Transformers, June 2026]. This academic pedigree is the initial wedge. Their technology was selected as the subject of a Kakao AI project in early 2026 and won a related competition [Seoul Economic Daily, January 2026]. The technical advisory board includes Kunle Olukotun, a Stanford professor and co-founder of AI chip company SambaNova Systems, who is supporting the team’s global strategy [Seoul Economic Daily, January 2026]. For a startup selling deep technical infrastructure, this isn’t just pedigree; it’s the product’s first line of credibility.
Building the abstraction layer
AnyBridge’s core bet is that the future of efficient AI compute is diversity, not monoculture. Nvidia’s dominance is built on the GPU, but a growing array of alternative accelerators from companies like FuriosaAI in South Korea are targeting specific AI workloads with better performance-per-watt. The problem is that today’s AI software stack is largely built for the GPU. AnyBridge’s software aims to be the abstraction layer that sits between the AI application and the underlying silicon, making decisions in real time. It evaluates latency requirements, model size, and operating cost to route work to the most suitable hardware, all through a common interface [Perplexity Sonar Pro Brief, retrieved 2026]. The planned product, set for a formal launch in 2027, will be offered as a licensed, managed, or usage-based “token factory” service [Dealroom News, retrieved 2026].
The company’s early moves signal a pragmatic, partnership-driven path to market. In September 2026, AnyBridge signed an agreement with Korean AI chipmaker FuriosaAI to jointly develop an NPU-based enterprise AI agent platform, combining FuriosaAI’s second-generation “RNGD” accelerator with AnyBridge’s software [The Asia Business Daily, September 2026]. This is a critical proof point. It moves the company from a research project to a co-development deal with a strategic hardware player, validating its role as the essential software glue.
The funding and the roadmap
In October 2026, AnyBridge secured a seed investment from The Ventures [Asiae, October 2026]. While the amount was not disclosed, the company was also recommended for South Korea’s TIPS program, which is expected to provide up to 800 million Korean won (roughly $580,000) in R&D funding over two years [Asiae, October 2026]. This public-private backing gives the young company a long runway to focus on product development ahead of its 2027 launch.
The leadership team, while academically strong, presents a classic deeptech transition challenge.
| Role | Name | Background |
|---|---|---|
| Co-founder, CEO | Jongse Park | Associate Professor, KAIST School of Computing; Visiting Associate Professor, Stanford University [LinkedIn, retrieved 2026]. |
| Co-founder | Youngjin Kwon | Professor, KAIST School of Computing [Perplexity Sonar Pro Brief, retrieved 2026]. |
| Technical Contributor | Jaehyuk Huh | Professor, KAIST School of Computing [KAIST NEWS CENTER, retrieved 2026]. |
| Advisor | Kunle Olukotun | Professor, Stanford University; Co-founder, SambaNova Systems [Seoul Economic Daily, January 2026]. |
The table shows a team built for solving hard technical problems. The next test is building a commercial machine,hiring a sales and marketing function, navigating enterprise procurement cycles, and proving that their cost-saving claims hold up outside a lab environment. The risks here are not small.
- The performance gap. A 40-70% cost reduction is a massive claim. It will need to be demonstrated repeatedly on real, mixed-hardware clusters with production workloads, not just in simulations.
- The ecosystem sell. AnyBridge must convince enterprises to invest in and manage a diverse array of accelerators, which adds operational complexity. The value of the software must outweigh that overhead.
- The incumbent response. Major cloud providers and existing AI infrastructure software companies are not standing still. They will develop or acquire similar abstraction capabilities.
The bet, however, is grounded in a simple economic truth. If the cost of AI inference is the primary barrier to widespread enterprise adoption, then the company that best optimizes those joules wins. On paper, the math is compelling. Take a mid-sized AI inference cluster costing $1 million per year in cloud GPU spend. A 50% saving from better hardware utilization translates to $500,000 annually, a line item any CFO would notice. That’s the number AnyBridge’s enterprise sales team will lead with. To succeed, they don’t need to beat every cloud provider; they need to become the indispensable layer for companies like FuriosaAI, proving that their software is the key that unlocks the value of every new chip that comes to market.
Sources
- [Asiae, October 2026] The Ventures Invests in AI Inference Infrastructure Startup AnyBridge, Recommends It for TIPS | https://view.asiae.co.kr/en/article/2026102208280231813
- [Perplexity Sonar Pro Brief, retrieved 2026] AnyBridge company briefing
- [AI Post Transformers, June 2026] AI Post Transformers podcast page | https://podcast.do-not-panic.com/2026/06/index.html
- [Seoul Economic Daily, January 2026] KAIST Team Wins Kakao AI Competition with Multi-Accelerator LLM Software | https://en.sedaily.com/technology/2026/01/30/kaist-team-wins-kakao-ai-competition-with-multi-accelerator
- [The Asia Business Daily, September 2026] AnyBridge and FuriosaAI joint development agreement | https://jmagazine.joins.com/forbes/view/344078
- [Dealroom News, retrieved 2026] KAIST professors' AI inference startup AnyBridge lands seed round from The Ventures | https://dealroom.co/news/164880-kaist-professors-ai-inference-startup-anybridge-lands-seed-round-from-th/
- [LinkedIn, retrieved 2026] Jongse Park - AnyBridge | LinkedIn | https://www.linkedin.com/in/jongse-park-6a6929245/
- [KAIST NEWS CENTER, retrieved 2026] KAIST NEWS CENTER page for Jaehyuk Huh | https://kaist.ac.kr/newsen/html/news/?skey=college&sval=school+of+computing&list_s_date=&list_e_date=&GotoPage=1