What FriendliAI sells, and what it buys
Tokens and inference infrastructure in three shapes: serverless per-token endpoints over open models, Dedicated Endpoints for custom and fine-tuned models, and Friendli Container, the same engine running inside a customer’s own GPU cluster. 540,000+ Hugging Face models deploy one-click; 99.99% uptime SLA on geo-distributed infrastructure.
GPU capacity across regions and clouds. The design stays asset-light: FriendliAI sources the fleets behind its SLA rather than owning them, and the 2026 Samsung Cloud Platform alliance for NVIDIA B300 inference capacity is the visible tip. This is the clearest demand-side posture of any card on this ladder: FriendliAI is a perpetual GPU buyer.
Scale proof
The bet
Inference is a software problem. That's the wager, and real estate is the counter-position. Founder Byung-Gon Chun invented continuous batching (published as Orca at OSDI 2022, now industry standard in LLM serving), and the company bets that custom kernels, caching and speculative decoding squeeze enough extra tokens out of each GPU that Friendli can stay asset-light and rent capacity while competitors buy data centers. DeepInfra shares the rung and takes the opposite posture, which gives the tokens rung two ways to sell the same unit.
Natural counterparty
On the sell side: enterprises with custom or fine-tuned models (the LG EXAONE pattern) and APAC-rooted AI companies needing regional serving. On the buy side: capacity providers. Because Friendli procures GPUs across regions and price points continuously, it is a natural demand-side counterparty for every supply-side company on this ladder. The APAC overlap makes GMI Cloud the most obvious pairing.
Buyer fit
You serve a custom or fine-tuned model and want frontier-grade latency and cost without owning infrastructure, or you want the inference engine deployed inside your own GPU cluster.
You just need cheap catalog-model tokens at massive scale and don’t care whose engine serves them.
Five fields we track but don’t publish
Everything above is public and compiled from primary sources. The fields below are demand-side: what Friendli is looking for, verified in conversation and never published.