FriendliAI's Inference Engine Already Powers SK Telecom and KT

The most expensive part of the AI stack isn't the model. It's the electricity and GPU time spent running it. For companies like SK Telecom and KT, the South Korean telecom giants, that cost is now a line item in the budget, not a research experiment. They are among the early customers of FriendliAI, a 2021-founded startup that is trying to turn academic research on inference optimization into a commercial business [SDxCentral, October 2025]. The bet is that enterprises will pay for a platform that makes their AI models cheaper and faster to run, not just easier to train.

From academic paper to production line

FriendliAI is an academic spinout from Seoul National University, and its technical wedge is rooted in a specific research contribution. The company states its founders pioneered continuous batching, a technique that groups inference requests to maximize GPU utilization, which has since become a standard optimization in the industry [Friendli.ai]. The founding team, led by CEO Byung-Gon Chun, a professor at SNU, and CTO Gyeong-In Yu, is building a commercial platform around this and other low-level optimizations like speculative decoding and custom GPU kernels [Friendli.ai]. This gives the company a clear, if narrow, technical moat: its product is not another model API, but an efficiency layer for running models, whether they are proprietary or from the open-source Hugging Face library.

The two-sided platform play

FriendliAI's product strategy has evolved into a two-pronged approach. The first is a straightforward inference cloud, offering serverless APIs and dedicated endpoints where developers can deploy and run models. The company claims this can deliver inference up to 7x faster and at 90% lower cost than baseline performance, though these are internal benchmarks [Friendli.ai]. The second, more recent move is InferenceSense, a platform launched in March 2026 designed to let GPU cloud operators monetize their idle capacity by plugging into FriendliAI's demand pipeline [Finance.yahoo.com, March 2026]. This positions the company not just as a tool for AI developers, but as an infrastructure layer for the broader compute ecosystem.

Seed Extension (Aug 2025) | 20 | M USD
Total Disclosed Funding | 26.7 | M USD

The company has secured capital to fund this expansion. In August 2025, it closed a $20 million seed extension round led by Capstone Partners, with participation from Sierra Ventures, Alumni Ventures, Korea Development Bank, and KB Securities [Friendli.ai, August 2025]. Public reports estimate total disclosed funding at approximately $26.7 million [Yespress.io]. The company has used this to build out its commercial team, appointing Brian Yoo, formerly of Moloco, as Chief Business Officer [Friendli.ai].

Traction in a specialized lane

Enterprise traction in the inference layer is harder to measure than user sign-ups, but FriendliAI points to a roster of initial customers that signal product-market fit in a specific, technical niche. Beyond SK Telecom and KT, the company's platform customers include AI startups Scatter Lab and Upstage [SDxCentral, October 2025]. These are not casual experiments; they are companies running AI in production where latency and cost directly impact the bottom line. The company has also forged strategic partnerships to expand its reach, including an alliance with Samsung Cloud Platform to offer inference on NVIDIA's latest B300 GPUs [Businesswire, April 2026].

Where the inference race gets crowded

For all its technical pedigree, FriendliAI is not operating in a vacuum. The market for efficient AI inference is heating up, and the competitive set is realistic and formidable. The company's ideal customer is a technical leader at a digital-native enterprise or a scaling AI startup,someone with a deployed model, a growing bill from a major cloud provider, and an engineering team capable of integrating a new serving layer. For this buyer, the alternatives are clear.

• Major cloud hyperscalers. AWS, Google Cloud, and Azure offer their own optimized inference services (e.g., SageMaker, Vertex AI). The sales motion here is about cost and performance beating the ease and inertia of the native cloud stack.
• Specialized inference startups. Companies like Anyscale (with Ray Serve) and Baseten offer similar promises of optimized model serving. Differentiation hinges on the depth of optimization, ease of use, and specific performance guarantees.
• In-house engineering. For the largest customers, building a custom serving layer using open-source tools like vLLM or TensorRT is always an option. FriendliAI must prove its managed service is more cost-effective than the internal developer time required to build and maintain a comparable system.

The company's answer to this competition rests on its academic roots and its focus on the full inference lifecycle, from serving to monetizing idle capacity. Its partnership to provide API access to LG AI Research's EXAONE 4.0 model also shows a strategy to bundle access to specific models with its performance optimizations [Einpresswire].

The next twelve months

The key milestones for FriendliAI will be commercial, not technical. The team has already demonstrated it can build the engine. Now it must prove it can sell it at scale, especially to customers outside its initial South Korean beachhead. Watch for two signals: an expansion of the customer list with more global enterprise names, and the traction of its InferenceSense platform with GPU providers. The path to a Series A will likely depend on showing that its optimization software can command premium pricing and secure long-term commitments, moving beyond project-based usage.

Sources

1. [SDxCentral, October 2025] Startup tackling AI inference bottlenecks bags $20M | https://www.sdxcentral.com/news/startup-tackling-ai-inference-bottlenecks-bags-20m/
2. [Friendli.ai] FriendliAI | The Frontier AI Inference Cloud | https://friendli.ai/
3. [Finance.yahoo.com, March 2026] FriendliAI Launches InferenceSense™ to Monetize Idle GPU Capacity | https://finance.yahoo.com/news/friendliai-launches-inferencesense-monetize-idle-130000648.html
4. [Friendli.ai, August 2025] FriendliAI closes $20M seed extension round | https://friendli.ai/blog/friendliai-raises-20m-in-seed-extension-round
5. [Yespress.io] FriendliAI company profile | https://yespress.io/friendliai
6. [Businesswire, April 2026] FriendliAI and Samsung Cloud Platform Forge Strategic Alliance | https://www.businesswire.com/news/home/20260415144530/en/FriendliAI-and-Samsung-Cloud-Platform-Forge-Strategic-Alliance-to-Power-Frontier-Model-AI-Inference-on-NVIDIA-B300-GPUs
7. [Einpresswire] FriendliAI closes $20M seed extension round to accelerate AI inference platform growth | https://www.einpresswire.com/article/843686113/friendliai-closes-20m-seed-extension-round-to-accelerate-ai-inference-platform-growth