FriendliAI Expands to San Francisco to Scale Frontier AI Inference for Open-Weight and Custom Models

Today we’re announcing the opening of our new San Francisco office at 20 Hawthorne Street, where we now occupy 7,000 square feet inside the historic Crown Point Press building, around the corner from the San Francisco Museum of Modern Art and the Moscone Center. The new space puts us at the heart of the Bay Area AI ecosystem and closer to the customers, partners, and developers building the next generation of AI applications.

FriendliAI’s 7,000-square-foot SoMa space anchors our U.S. expansion.

An inflection point for inference

Our expansion lands at an inflection point for AI inference. Two forces are driving the shift: AI agents — which plan, reason across many steps, and call tools on every turn — require five to thirty times more tokens per task than chatbots. That consumption compounds as agents move from pilots into always-on production workflows. Meanwhile, the latest open-weight models including Z.ai’s GLM-5.1, Moonshot AI’s Kimi K2.6, DeepSeek V4, and NVIDIA Nemotron 3 now match or exceed leading closed models like Anthropic’s Claude Opus at a fraction of the cost, and custom fine-tunes align even more tightly with enterprise use cases. Production-grade inference infrastructure has become the bottleneck — and the prize.

“San Francisco is the epicenter of AI innovation, and a deeper presence here lets us partner with the customers and developers shaping what comes next,” said FriendliAI CEO, Byung-Gon Chun. “The industry is no longer asking whether to build with AI — it’s asking how to run AI in production, profitably, at scale. FriendliAI, The Frontier AI Inference Cloud, was built for exactly that.”

How we got here

FriendliAI was founded by Professor Byung-Gon Chun and members of his research team at Seoul National University, where they pioneered continuous batching — the inference optimization technique that is now an industry standard. Today FriendliAI runs state-of-the-art open-weight and custom models at production scale with industry-leading throughput, latency, and reliability. Independent benchmarks from Artificial Analysis and OpenRouter rank FriendliAI as the top inference provider for models such as GLM-5.1 and Gemma 4 across output speed, latency, tool calling, and structured outputs. We partners with model creators on launch — most recently as a Day 0 partner for NVIDIA Nemotron 3 and Z.ai’s GLM-5.1 — and with cloud providers including AWS, OCI, and Samsung Cloud Platform on infrastructure to scale globally.

Customers including Twelve Labs and LG are already scaling with FriendliAI in production, and that momentum is translating into rapid business growth. FriendliAI is on a trajectory to grow revenue tenfold this year, with a goal of growing another tenfold the year after, as AI-native and AI-augmented SaaS companies migrate production workloads to its platform. The San Francisco expansion is built to support the trajectory: FriendliAI plans to significantly grow its U.S. team across go-to-market, partnerships, and engineering functions over the coming year.

“Inference is where AI economics are won or lost,” said our Chief Business Officer, Brian Yoo. “Every percentage point of GPU efficiency translates directly to margin, and every millisecond of latency translates to user experience. Putting senior commercial and engineering leadership on the ground in San Francisco lets us move at the speed our customers need as they scale.”

FriendliAI’s new SF office: a hub for the AI builder community

Our bright, loft-style space is also purpose-built as a hub for the AI builder community, hosting developer meetups, hackathons, and executive briefings on the practical realities of deploying inference at scale — from open-weight model deployments and GPU efficiency to multimodal and agentic workloads.

Stay tuned for hackathons, developer days, and event afterparties!