Fireworks AI

When PyTorch engineers get bored

In 2022, Lin Qiao was Head of PyTorch at Meta - which is roughly equivalent to being the caretaker of the most important piece of AI infrastructure in the world. PyTorch powers the research of nearly every major AI lab, runs inside Meta's own products, and underpins most of the AI papers published in academic journals. A comfortable perch, to say the least.

She left anyway. And she brought six colleagues with her.

The problem she kept running into wasn't a shortage of clever AI models. It was everything else. Companies across every industry wanted to deploy generative AI but had nowhere to start. No GPU infrastructure. No inference stack. No team with the depth to build it. The models existed. The capability to run them efficiently - at scale, at low cost, with production-grade reliability - was gated behind Meta, Google, and a handful of hyperscalers.

Fireworks AI's founding thesis was simple: democratize the infrastructure. Build the inference layer that every company needs but almost no company can build themselves. The name is a direct nod to PyTorch's flame logo - spreading the same fire that lit up Meta's AI stack across the rest of the industry.

Seven engineers who built the internet's AI backbone

The founding team is a collector's item. Five of the seven co-founders worked directly on PyTorch at Meta - the framework that runs inside nearly every serious AI research project and production system. The other two come from Meta Ads infrastructure and Google Vertex AI. This isn't a team that read about AI infrastructure. They built it.

Lin Qiao holds a PhD in Computer Science from UC Santa Barbara focused on distributed systems and machine learning. Before PyTorch, she worked at Facebook, LinkedIn, and IBM. Her co-founders bring comparably deep credentials: PyTorch core maintenance, Newsfeed ML systems, Google Vertex AI, and Meta Ads infrastructure at scale.

The full stack for AI in production

Fireworks AI is not a model company. It doesn't train foundation models or try to compete with OpenAI or Anthropic on raw capability. Instead, it solves a different problem: how do you take an open-source model and run it in production, fast, cheaply, and reliably, without hiring a team of GPU engineers?

The platform covers the full lifecycle - from model selection to fine-tuning to deployment to real-time inference at scale. The OpenAI-compatible API means developers can swap in Fireworks without rewriting their code. The catalog includes hundreds of models across text, image, audio, and multimodal workloads.

A funding history that keeps accelerating

Fireworks AI raised $25M in a Series A in early 2024. Seven months later, Sequoia led a $52M Series B at roughly a $552M valuation. Fifteen months after that, the company closed a $250M Series C at a $4B valuation - a 7x jump in a year and a half. The investor roster reads like a who's who of Silicon Valley's most plugged-in technologists.

Strategic investors include NVIDIA and AMD - both chipmakers betting that Fireworks will be the layer that puts their hardware to work for the most demanding AI workloads. Sheryl Sandberg (former Meta COO) and Frank Slootman (former Snowflake CEO) are among the personal angels backing the company.

What three years can look like

Running on every major cloud

Fireworks AI operates across NVIDIA, AMD, AWS, Google Cloud, and Oracle Cloud GPU infrastructure. In 2025, the company signed a Strategic Collaboration Agreement with AWS and achieved AWS Generative AI Competency Partner status. A multi-year deal with Microsoft Azure brought Fireworks into Azure Foundry as a native option.

On the model side, partnerships with Meta (Llama), Mistral AI, and Stability AI ensure the catalog stays current with the latest open-source releases. Databricks and MongoDB are both strategic investors and integration partners - Fireworks handles inference for ML workflows built on Databricks and RAG pipelines built on MongoDB.

Why one model will never be enough

The popular narrative in AI has been about chasing the biggest, smartest foundation model. GPT-4, then GPT-5. Claude Sonnet, then Claude Opus. Each release touted as the one that finally cracks general intelligence.

Fireworks AI is betting on a different future. Lin Qiao has argued publicly that the next phase of AI quality won't come from making models bigger - it'll come from making them more specialized. Hundreds of small, expert models trained on curated domain-specific data, each solving a narrow problem better than any general-purpose model can.

That future is convenient for a platform that already hosts hundreds of models and has built fine-tuning infrastructure capable of creating new ones cheaply. Fireworks isn't just serving the current AI market - it's building toward one where every company has its own model, its own inference stack, and its own AI product. Not a dependency on OpenAI's pricing decisions.

The Hathora acquisition

In March 2026, Fireworks AI acquired Hathora Inc., a real-time container orchestration platform spanning 14 regions across two bare-metal providers and four cloud environments. The move was framed as a talent and infrastructure acquisition - Hathora's engineering culture, described by Lin Qiao as focused on "every millisecond and every routing decision," maps directly onto the demands of cutting-edge AI inference.

The practical effect: sub-20ms routing decisions across a global network, applied to inference workloads that demand low latency at any location. For developers running AI products with global user bases, that's the difference between an AI assistant that feels responsive and one that doesn't.

From PyTorch to $4 billion

Six facts about Fireworks AI

What's been happening