FriendliAI

  Executive Briefing

FriendliAI is a Korean-founded AI inference cloud company that builds and operates a proprietary serving stack for open-weight large language models, offering serverless and dedicated GPU endpoints to developers and enterprises worldwide. The company is led by Byung-Gon ("Gon") Chun, founder and CEO, a Seoul National University professor on leave and former Principal Scientist at Microsoft Research, and co-founder and CTO Gyeong-In Yu, a Ph.D. graduate of Chun's SNU lab who shared the Best Ph.D. Dissertation Award from the department. Together, they co-authored the ORCA paper presented at OSDI 2022, which introduced iteration-level scheduling and continuous batching — the foundational queuing technique that directly inspired vLLM and has since been adopted across virtually every major inference framework in the industry. FriendliAI was founded to productionize those research insights, transforming a landmark academic result into an enterprise-grade inference cloud.

At the core of the product is the Friendli Inference Engine (formerly PeriFlow Engine, rebranded November 2023), a full-stack serving system built from scratch rather than layered on top of vLLM or Hugging Face TGI. By owning every layer — from CUDA kernel design to KV cache management to model parallelism — FriendliAI claims benchmark-leading throughput: 3.26x the throughput of vLLM FP8 on Qwen3 235B at 4-bit quantization, and rankings of first place on Artificial Analysis and OpenRouter for output speed and lowest time-to-first-token on multiple flagship models including GLM-5.1 and Gemma-4-31B.¹² The platform supports serverless pay-per-token access, reserved dedicated GPU endpoints across A100 through B200/B300 hardware, and a containerized self-hosted option for private clusters.

FriendliAI has raised approximately $26.5 million across two rounds — a seed of roughly $6 million in 2021 and a $20 million seed extension closed in August 2025 — while serving an estimated 25–30 large enterprise customers including LG AI Research (exclusive API provider for the EXAONE model family), Twelve Labs, Upstage, and Scatter Lab.³ In May 2026 the company opened a 7,000 sq ft office in San Francisco's SoMa district to expand its U.S. go-to-market presence, reflecting a strategic bet that the agentic AI wave — which consumes 5–30x more tokens per session than chatbot workloads — will drive the next order-of-magnitude growth in inference demand.⁴

The company occupies a distinctive position in the inference landscape: it can legitimately claim to have invented the algorithmic technique that underpins most modern LLM serving, it holds a granted U.S. patent on that technique (No. 11,442,775), and it operates a purpose-built engine rather than wrapping open-source tools. That technical pedigree, combined with enterprise compliance certifications (SOC 2 Type II, HIPAA, 99.99% uptime SLA) and day-zero partnerships for frontier open-weight model launches, gives FriendliAI a credible differentiation story against both the closed-API giants and the fast-growing open-source inference ecosystem it helped create.

  At a Glance

  Origins & Founding

FriendliAI's origins trace directly to a research group at Seoul National University. Byung-Gon Chun spent his academic career studying distributed systems and machine learning infrastructure: he earned his Ph.D. from UC Berkeley, his M.S. from Stanford, and held positions at Yahoo! Research Silicon Valley, Intel Research Berkeley, and Facebook (as a visiting research scientist) before joining Microsoft Research as a Principal Scientist and later becoming a professor at SNU. His lab at SNU produced Gyeong-In Yu as a Ph.D. student working on efficient neural network serving.

The two co-authored the ORCA paper — "Orca: A Distributed Serving System for Transformer-Based Generative Models" — which was accepted at OSDI 2022, the top operating systems and infrastructure conference.⁵ ORCA introduced iteration-level scheduling, a technique that allows a serving system to insert new requests into an ongoing batch at the boundary of each transformer forward pass, rather than waiting for all sequences in a batch to finish before accepting new work. This seemingly simple change unlocks dramatically higher GPU utilization under variable-length workloads, and the technique — under the name "continuous batching" — was independently adopted by vLLM shortly after the paper's publication and quickly became the de facto standard for LLM serving. FriendliAI holds U.S. Patent No. 11,442,775 on the underlying method.⁶

Chun founded the company in 2021, initially under the PeriFlow brand, with the thesis that the same research group that developed the foundational technique should be the one to productionize it at enterprise scale. The seed round of approximately $6 million, led by Capstone Partners, gave the team runway to build a production engine and go-to-market infrastructure. The core engineering team was seeded from Chun's SNU research group, giving the company deep technical continuity from paper to product. FriendliAI was selected for the Korea AI Startup 100 program in 2022, its first major industry recognition outside the academic community.

  History & Timeline

    2021–2022: Research commercialization and seed funding

FriendliAI was founded in Seoul in 2021, raised its approximately $6 million seed round led by Capstone Partners, and built its initial inference engine around the continuous-batching techniques Chun's lab was simultaneously publishing in academic form. The ORCA paper was presented at OSDI 2022 in July of that year, establishing the company's technical credibility with the AI infrastructure community and the paper's publication. The team enrolled in the Microsoft for Startups Founders Hub in November 2022, receiving $150,000 in Azure credits that supported early cloud infrastructure. By the end of 2022 the company had been named to the Korea AI Startup 100 for the first time.

    2023: Patent, rebrand, and U.S. expansion

In April 2023, CEO Byung-Gon Chun was invited to the White House Korea-US Business Roundtable, reflecting the company's growing profile. The company publicly launched the PeriFlow Cloud public beta in July 2023 — the first managed cloud offering built on its engine. That same month, on July 28, 2023, FriendliAI filed a patent infringement lawsuit against Hugging Face in the U.S. District Court for the Northern District of Delaware, alleging that Hugging Face's Text Generation Inference (TGI) tool infringed U.S. Patent No. 11,442,775 covering continuous batching and dynamic batch scheduling.⁶ In November 2023 the company rebranded comprehensively: PeriFlow became Friendli Engine, Friendli Container, and Friendli Dedicated Endpoints, with quantization and state caching added as new capabilities. The U.S. headquarters moved from Korea to Redwood City, California in late 2023, marking a formal westward pivot. The company was selected for Korea AI Startup 100 for a second consecutive year.

    2024: Product expansion and ecosystem integrations

January 2024 brought the launch of Friendli Serverless Endpoints — the pay-per-token API product — alongside recognition as an "Emerging AI+X Top 100" company in the NLP-based AI Platform category. In February, CEO Chun and CTO Yu were among only five companies invited to meet Meta CEO Mark Zuckerberg at a Developers Roundtable in Korea. The company demonstrated at NVIDIA GTC 2024 in March, launched Dedicated Endpoints in April, and became an official AWS Partner in April. CEO Chun presented at the Meta Llama Community Summit at Meta HQ in May 2024. The second half of the year brought a series of ecosystem integrations: Weights & Biases (June), Upstage partnership for Solar translation LLMs on Dedicated Endpoints (July), MongoDB for RAG pipeline deployment (August), and Weaviate vector database (September).

    2025: Patent settlement, Hugging Face partnership, and growth funding

On January 8, 2025, FriendliAI and Hugging Face reached a settlement: a cross-licensing agreement was signed allowing both companies access to each other's patent portfolios, and the case was dismissed without costs or fees to either party.⁶ The same month, FriendliAI launched a Hugging Face Hub integration enabling one-click deployment of models from the Hub to Friendli Endpoints. In July 2025, LG AI Research named FriendliAI as the exclusive API distribution partner for the EXAONE 4.0 model family, a significant enterprise anchor. The $20 million seed extension closed on August 28, 2025, led by Capstone Partners with new investors Sierra Ventures, Alumni Ventures, KDB Investment, and KB Securities joining.³ At that point the company reported approximately 25–30 large enterprise clients and projected 6–7x revenue growth for full-year 2025 versus 2024.

    2026: San Francisco expansion and next-generation hardware

FriendliAI opened its 7,000 sq ft San Francisco SoMa office at 20 Hawthorne Street on May 11, 2026, establishing its operational U.S. headquarters in proximity to the developer ecosystem and enterprise customers.⁴ Concurrently, the company announced a strategic alliance with Samsung Cloud Platform (Samsung SDS) to deploy NVIDIA B300 GPU clusters, extending its hardware fleet to the latest generation silicon. CEO Chun has stated targets of approximately 10x revenue growth in 2026, followed by another approximately 10x in 2027, driven by agentic workload expansion.³

  What They Offer — Products & Platform

FriendliAI's commercial offering is organized into three products, each targeting a different deployment context along the build-versus-manage spectrum.

Serverless Endpoints (Model APIs) are the pay-as-you-go tier, giving developers API access to a curated catalog of popular open-weight models with no infrastructure management. Billing is per token, there is no cold-start charge for established models, and the product supports text, vision, audio, and video modalities together with tool calling, structured outputs, and streaming. The serverless tier also includes compatibility with the Anthropic Messages API format. More than 570,000 Hugging Face models can be deployed to Friendli Serverless Endpoints via the Hub integration, making the effective catalog far larger than the curated list.⁷

Dedicated Endpoints provide reserved GPU capacity billed per second, with no startup-time charges and guaranteed throughput, for teams that need predictable low latency, high concurrent request handling, or the ability to serve custom fine-tuned models. Hardware options span A100 80GB at $2.90/hr, H100 80GB at $3.90/hr, H200 141GB at $4.50/hr, and B200 180GB at $8.90/hr. Dedicated Endpoints are the appropriate tier for production workloads where latency SLA and throughput consistency matter more than per-unit cost. Teams can bring custom model checkpoints and deploy them without managing the underlying serving infrastructure.

Friendli Container is the self-hosted option, packaging the Friendli Engine as a container image that operators can run on their own private GPU clusters. Pricing is custom and available through direct sales engagement. This tier is aimed at enterprises with data-residency requirements, regulated industries, or existing GPU investments they wish to utilize with a production-grade serving layer.

All three tiers are backed by a 99.99% uptime SLA, SOC 2 Type II certification, and HIPAA compliance — certifications that position FriendliAI for healthcare, financial services, and government workloads that cannot accept the compliance posture of the open-source alternatives.

  Technology & Infrastructure

The Friendli Inference Engine is a ground-up implementation of a large language model serving system, deliberately not built on top of vLLM, Hugging Face TGI, or NVIDIA TensorRT-LLM. The engineering team's rationale is that owning the full stack provides "full control over every layer," enabling optimizations that wrapper-based approaches cannot achieve without forking and maintaining upstream codebases. This architectural choice reflects the founders' position: having invented the continuous-batching technique, they consider the full-stack engine the natural expression of that research lineage.

Key technical capabilities of the Friendli Engine include:

Continuous / iteration-level batching — the ORCA-derived scheduling technique that processes requests at the granularity of individual transformer forward passes, enabling high GPU utilization even under heterogeneous sequence lengths and arrival rates. This is the foundational optimization from which the company's performance claims flow.

Custom CUDA kernels — the engine implements custom GPU kernels for attention and other compute-intensive operations, enabling performance characteristics that diverge from stock PyTorch or CUDA library baselines.

Smart KV caching — the key-value cache management includes a "host KV cache" mechanism that allows the effective context window to exceed GPU VRAM limits by offloading cache to CPU memory, enabling longer-context workloads without proportional GPU cost.

Speculative decoding and adaptive speculative decoding — techniques that use a smaller draft model to propose token sequences that a larger verifier model then validates in parallel, reducing effective latency per output token for appropriate workloads.

MoE-aware routing and expert-activation scheduling — optimizations specifically designed for mixture-of-experts architectures (Qwen3-235B, DeepSeek-V3.2, GLM-5.1, K-EXAONE-236B), where expert routing patterns create load-imbalance challenges that generic serving frameworks do not address.

Online quantization — INT8 and INT4 quantization applied at serving time with, per the company's claims, minimal quality degradation, and with published benchmarks showing 3.26x throughput gain over vLLM FP8 baseline on Qwen3-235B at 4-bit in a 500-input/4000-output token scenario.¹

Tensor and pipeline parallelism — multi-GPU and multi-node distribution for models that exceed single-card memory.

The hardware fleet covers NVIDIA A100 80GB, H100 80GB, H200 141GB, B200 180GB, and (as of 2026) B300 GPUs. Infrastructure is multi-cloud, running across AWS, Oracle Cloud Infrastructure (OCI), and Samsung Cloud Platform, with geo-distributed data centers. The Samsung SDS alliance announced in May 2026 adds dedicated B300 cluster capacity. Published benchmarks report 4x higher throughput on H100 versus A100 at equivalent latency on Llama 2 70B, and 1.5–1.8x lower latency under varying load.⁸

  Model Catalog & Performance

FriendliAI's serverless catalog is curated around high-demand open-weight models from the major frontier open-weight model families, with a particular emphasis on day-zero availability for new model releases. The company has established a pattern of launching inference access for new models on or near their public release dates, including NVIDIA Nemotron 3 Super (as a Day 0 partner), GLM-5.1, EXAONE 4.0 (as exclusive API provider for LG AI Research), and MiniMax-M2.5.

Current catalog highlights with indicative serverless pricing:

All pricing is indicative as of mid-2026 and subject to change. The company also supports 570,000+ Hugging Face Hub models via one-click deployment, though those are typically served on Dedicated Endpoints rather than as standardized serverless SKUs.

Performance standings: Artificial Analysis has ranked FriendliAI first for output speed and lowest TTFT on GLM-5.1 as of 2026.² OpenRouter lists FriendliAI as top-rated for tool calling and structured outputs with the lowest error rates, and as one of the most-used providers for coding inference.⁹

  Pricing & Performance Position

FriendliAI's pricing philosophy targets two distinct cost arguments: against closed-model APIs (OpenAI, Anthropic) and against self-managed open-source serving. Against closed providers, the company claims 50–90% cost savings by substituting open-weight models via Friendli APIs — reflecting both lower per-token rates and the efficiency premium of its engine. To support migrations, the company has run a "$50,000 switch campaign" offering inference credits to teams moving from closed providers. Against self-managed vLLM or TGI deployments, FriendliAI claims a 20–40% cost improvement from needing roughly 50% fewer GPUs to serve equivalent throughput — a claim grounded in the 2–3x throughput advantage the Friendli Engine reports over vLLM on comparable hardware.

Dedicated Endpoint hourly rates, billed per second:

There are no startup-time charges on Dedicated Endpoints, which is a meaningful cost advantage versus alternatives that bill a minimum warmup period. Serverless billing is pure pay-per-token with no seat fees, minimum commitments, or idle charges.

Published performance benchmarks (from the company and third parties):

• 3.26x higher throughput vs. vLLM FP8 baseline on Qwen3-235B at INT4, 500 input / 4000 output token scenario.¹
• 4x higher throughput on H100 versus A100 at equivalent latency on Llama 2 70B.⁸
• 1.5–1.8x lower latency on H100 versus A100 under varying request load.⁸
• Ranked #1 by Artificial Analysis for output speed and TTFT on GLM-5.1.²
• Ranked #1 by OpenRouter for output speed and TTFT on Gemma-4-31B-it, and highest rated for tool calling and structured output reliability.⁹

As with all vendor-published benchmarks, independent replication is advised for production procurement decisions.

  People & Leadership

Byung-Gon ("Gon") Chun is the founder and CEO of FriendliAI. His career spans Yahoo! Research Silicon Valley, Intel Research Berkeley, Facebook (visiting research scientist), and Microsoft Research (Principal Scientist), and he holds a Ph.D. from UC Berkeley and an M.S. from Stanford. His academic work on distributed systems and ML infrastructure culminated in the ORCA paper, which he co-authored with Yu. He has been invited to the White House Korea-US Business Roundtable (April 2023), was among five companies invited to meet Mark Zuckerberg at a Korea Developers Roundtable (February 2024), and presented at the Meta Llama Community Summit at Meta HQ (May 2024).¹⁰

Gyeong-In Yu is co-founder and CTO. Yu earned his Ph.D. in Computer Science and Engineering from Seoul National University under Chun's supervision, receiving the department's Best Ph.D. Dissertation Award. He is the lead author of the ORCA paper and holds two U.S. patents on dynamic batching for transformer inference. During his Ph.D., he completed internships at Microsoft AI & Research and Microsoft Research Asia.¹¹

Changyup Kim serves as VP of Business Development and Sales, leading enterprise and partnership go-to-market efforts.

The company does not publicly disclose headcount, but has been expanding U.S.-based engineering, partnerships, and go-to-market staff in connection with the San Francisco office opening in May 2026.

  Funding, Ownership & Business

FriendliAI is an independent company. Total reported funding across all rounds is approximately $26.5 million.³ (Some sources cite $25 million; the discrepancy likely reflects the $6 million seed and $20 million extension summing to $26 million, with rounding differences in reported figures.)

Seed round (late 2021): approximately $6 million, led by Capstone Partners.

Seed extension (August 28, 2025): $20 million, led by Capstone Partners, with Sierra Ventures, Alumni Ventures, KDB Investment, and KB Securities as new investors.³ Valuation was not disclosed; CEO Chun confirmed it exceeds the valuation at the time of the prior raise. No Series A has been announced as of June 2026.

The business model is usage-based: pay-per-token for serverless, per-second for dedicated GPU, and custom contract for Friendli Container. The company was not yet profitable as of August 2025 by the CEO's own account, but described strong gross margins and efficient scaling. CEO Chun has stated revenue projections of approximately 6–7x growth in 2025 versus 2024, approximately 10x in 2026, and another approximately 10x in 2027 — forward-looking estimates from the CEO, not audited figures, and subject to significant uncertainty.³

The company enrolled in the Microsoft for Startups Founders Hub in November 2022, receiving $150,000 in Azure credits, and became a formal AWS Partner in April 2024. Infrastructure is run across AWS, OCI, and Samsung Cloud Platform, though whether FriendliAI owns or rents the underlying GPU capacity is not publicly disclosed.

  Customers & Partnerships

FriendliAI's enterprise customer base was reported at approximately 25–30 large organizations as of August 2025.

Notable customers:

• LG AI Research — FriendliAI's most prominent enterprise anchor. FriendliAI serves as the official inference partner for the EXAONE model family, including the 236B MoE model, and became the exclusive API distribution partner for EXAONE 4.0 in July 2025.¹²
• Twelve Labs — video understanding inference at scale.
• Scatter Lab — Korean AI chatbot company.
• Upstage — Solar translation LLMs deployed on Dedicated Endpoints (announced July 2024).

Cloud and infrastructure partnerships:

• AWS (AWS Partner Network member since April 2024)
• Oracle Cloud Infrastructure (OCI)
• Samsung Cloud Platform / Samsung SDS (strategic alliance for NVIDIA B300 GPU clusters, announced May 2026)¹³

Ecosystem integrations:

• Hugging Face Hub (January 2025) — one-click deployment of 570,000+ Hub models to Friendli Endpoints.⁷
• Weights & Biases (June 2024) — experiment tracking and observability for inference workloads.
• MongoDB (August 2024) — RAG pipeline deployment integration.
• Weaviate (September 2024) — vector database integration for retrieval-augmented generation.

Marketplace and developer presence:

• OpenRouter — listed as a top-rated inference provider; most-used for coding inference, tool calling, and structured outputs.⁹

  Competitive Position

FriendliAI competes across multiple layers of the inference value chain simultaneously, facing different competitors depending on the buyer's starting point.

Against open-source serving frameworks — vLLM, Hugging Face TGI, NVIDIA TensorRT-LLM — FriendliAI argues that its proprietary engine delivers materially better efficiency and that self-managing open-source infrastructure requires engineering investment that outweighs the cost of its managed offering. The ORCA lineage gives FriendliAI a legitimate claim to foundational IP that vLLM itself built on, and the 2025 cross-licensing settlement with Hugging Face (following the TGI patent suit) underlines the seriousness of that IP position.⁶

Against closed-model API providers — OpenAI, Anthropic, Google — FriendliAI positions as the open-weight alternative with 50–90% cost savings, data-residency optionality, and the ability to serve custom fine-tuned models. This is a fundamentally different product proposition rather than direct competition on identical models.

Against managed inference API providers — Together AI, Fireworks AI, Replicate, Groq, Baseten, Hyperbolic — FriendliAI's differentiation rests on three claims: technical pedigree (ORCA inventors), proprietary full-stack engine rather than vLLM/TGI wrappers, and enterprise-grade reliability and compliance (SOC 2 Type II, HIPAA, 99.99% SLA). Day-zero support for frontier model launches (GLM-5.1, NVIDIA Nemotron 3 Super, EXAONE 4.0) and exclusive partnerships (EXAONE API distribution) also differentiate the catalog.

The principal risks for FriendliAI are capital scale (well-funded competitors like Groq with custom silicon, or Together AI with deep VC backing), the pace of improvement in open-source serving frameworks (which continuously narrow the performance gap), and the possibility that frontier model performance shifts back toward closed providers in ways that reduce demand for open-weight serving infrastructure.

  Notable Events

Patent lawsuit against Hugging Face (2023–2025). On July 28, 2023, FriendliAI filed suit against Hugging Face in the U.S. District Court for the Northern District of Delaware, alleging that Hugging Face's Text Generation Inference (TGI) tool infringed U.S. Patent No. 11,442,775 covering continuous batching and dynamic batch scheduling for transformer inference.⁶ The case was among the first patent enforcement actions in the LLM serving infrastructure space. On January 8, 2025, the parties reached a settlement: a cross-licensing agreement granting both companies access to each other's patent portfolios, with the case dismissed without costs or fees to either side. The specific financial terms of the cross-license were described as confidential.⁶

White House and Meta engagements (2023–2024). CEO Byung-Gon Chun was invited to the White House Korea-US Business Roundtable in April 2023. In February 2024, Chun and CTO Yu were two of only five companies' representatives invited to a Developers Roundtable with Meta CEO Mark Zuckerberg in Korea. Chun presented at the Meta Llama Community Summit at Meta HQ in May 2024 — engagements that reflect FriendliAI's standing within the open-weight model ecosystem.¹⁰

San Francisco office opening (May 2026). The company formally established a 7,000 sq ft operational U.S. headquarters at 20 Hawthorne Street, SoMa, San Francisco, on May 11, 2026, signaling a decisive move to build U.S. commercial presence commensurate with the seed extension proceeds and revenue growth trajectory.⁴

  Outlook & Roadmap

FriendliAI's stated strategic thesis for the next two to three years is that the transition from chatbot-style AI to agentic AI represents the defining infrastructure inflection of the current cycle. Agentic workloads — where AI systems operate autonomously over extended tasks, calling tools and generating intermediate reasoning — consume an estimated 5–30x more tokens per session than conversational chatbots, creating a step-change demand for inference capacity. FriendliAI positions itself as the "picks and shovels" layer of this wave: the inference infrastructure that agentic frameworks and applications run on, regardless of which open-weight model wins at the application layer.

Specific focus areas:

• Agentic workload optimization — serving infrastructure tuned for the long-context, high-throughput, multi-turn patterns of agentic systems.
• U.S. market expansion — the San Francisco office is the operational foundation; the company plans to add engineering, partnerships, and go-to-market headcount in the U.S. through 2026–2027.
• Multimodal expansion — extending beyond text to vision, audio, and video serving, already underway in the catalog.
• Next-generation hardware — B200 and B300 GPU support is already live or in deployment, positioning FriendliAI at the hardware frontier.
• Open-weight model ecosystem — continued day-zero partnership strategy for frontier open-weight launches, creating catalog freshness as a competitive moat.
• Revenue scale — CEO-stated targets of approximately 10x revenue growth in 2026 and another approximately 10x in 2027 are aspirational projections and should be treated as such; no Series A or subsequent funding round has been announced as of June 2026.

The open-weight model quality trajectory is FriendliAI's central market tailwind: as models like Llama, Qwen, DeepSeek, and EXAONE narrow the gap with closed-API alternatives, the addressable market for open-weight inference infrastructure expands proportionally.

---

  References

1. FriendliAI Qwen3-235B Benchmark
2. Artificial Analysis — GLM-5.1 Providers
3. Crunchbase — FriendliAI raises seed extension
4. BusinessWire — FriendliAI expands to San Francisco
5. ORCA paper — OSDI 2022
6. TechCrunch — Hugging Face settles patent suit with FriendliAI
7. FriendliAI — Hugging Face Hub integration
8. FriendliAI — Friendli Engine on NVIDIA H100
9. OpenRouter — FriendliAI provider page
10. Cerebral Valley — FriendliAI profile
11. Gyeong-In Yu — personal site
12. FriendliAI — News
13. FriendliAI — Samsung Cloud Platform collaboration

  References

1. FriendliAI Qwen3-235B throughput benchmark — friendli.ai/blog/Qwen3-235B-benchmark. ↩ ↩² ↩³

2. Artificial Analysis ranking for GLM-5.1 output speed and TTFT — artificialanalysis.ai/models/glm-5-1/providers. ↩ ↩² ↩³ ↩⁴

3. Seed extension details and revenue projections (CEO-stated, not audited) — Crunchbase / Bloomberg. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

4. San Francisco office opening — BusinessWire, May 11 2026. ↩ ↩² ↩³

5. ORCA paper — Yu et al., OSDI 2022 — usenix.org/conference/osdi22/presentation/yu. ↩

6. Patent suit and settlement — U.S. Patent No. 11,442,775; settlement January 8 2025 — TechCrunch. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

7. Hugging Face Hub integration (January 2025) — friendli.ai/news. ↩ ↩²

8. H100 vs A100 throughput and latency benchmarks — friendli.ai/blog/friendli-engine-llm-serving-nvidia-h100. ↩ ↩² ↩³

9. OpenRouter provider rankings and usage data — openrouter.ai/provider/friendli. ↩ ↩² ↩³ ↩⁴

10. CEO engagements and background — Cerebral Valley profile; bgchun.github.io. ↩ ↩²

11. CTO background — gyeongin.github.io. ↩

12. LG AI Research partnership — friendli.ai/news. ↩

13. Samsung Cloud Platform strategic alliance — friendli.ai/blog/friendliai-collaborates-with-samsung-cloud-platform. ↩