Fireworks AI

Fireworks AI is an American artificial intelligence infrastructure company that sells access to optimised inference — the step of actually running a trained model — through a hosted cloud platform. It does not build its own chips; instead it optimises how open-weight models execute on GPUs and rents that performance to developers and enterprises by the token.^[1]

History

Fireworks AI was founded in October 2022 by seven engineers who had worked together at Meta, led by co-founder and CEO Lin Qiao, who previously ran the PyTorch and Caffe2 teams as Meta's Senior Director of Engineering. The founding thesis was that the industry had poured resources into training ever-larger models while serving them quickly and cheaply in production remained unsolved.^[2]

The company grew with the wider shift of AI spending from training to inference. Sequoia Capital and NVIDIA backed it in a 2024 round that brought total funding to roughly US$552 million. In October 2025 Fireworks announced a US$250 million Series C co-led by Lightspeed Venture Partners, Index Ventures and Evantic Capital with continued participation from Sequoia, valuing the company at US$4 billion; annualised revenue had already passed US$280 million.^[1] By November 2025 it reported more than 10,000 customers — including Uber, Shopify, DoorDash, Notion and Upwork — and over 10 trillion tokens processed daily.^[3] In July 2026 the company said it had surpassed a US$1 billion annualised revenue run rate and was serving more than 40 trillion tokens per day, with total funding reaching about US$1.8 billion.^[3]

Key Concepts and Technology

The core product is model serving: Fireworks hosts open-weight models such as Llama, Qwen, DeepSeek, Kimi and GLM, plus image, audio and embedding models, behind an OpenAI-compatible API. Its performance claims rest on proprietary kernel-level optimisation — the FireAttention attention kernel — together with compilation and batching strategies that raise GPU utilisation and cut per-token latency.^[2]

The company sells three things enterprises would otherwise assemble themselves: throughput for batch workloads at low cost, latency for interactive applications such as chat and agents, and customisation, including low-rank adaptation fine-tuning of open-weight models on private data without training a model from scratch. Pricing is consumption-based — per million tokens — rather than per seat, so cost tracks usage directly.^[1] It competes with Together AI, Baseten, Groq, Cerebras and the hyperscalers' own inference services, and complements aggregators such as OpenRouter, which route traffic across providers.^[4]

Applications and Impact

Typical workloads include customer-support agents, retrieval-augmented generation over company documents, code assistants, content-generation pipelines and embedding and search systems. Because Fireworks serves open-weight models, customers can switch providers or self-host later, which has made hosted open models a credible alternative to proprietary APIs for cost-sensitive production systems. The company's revenue climb — from roughly US$6.5 million in May 2024 to over US$1 billion of run-rate by mid-2026 — is regularly cited as evidence that inference, not training, is where AI spending is concentrating.^[3][4]

>See Also

• Inference
• Model Serving
• Open-Weight Models
• Together AI
• Quantization
• Malaysia Data Centres
• AWS Malaysia Region
• PDPA and AI