Fireworks AI platform scale
Fireworks AI completed a $1.5 billion Series D round in July 2026, which set the company valuation at $175 billion. This is a significant increase from its $552 million valuation in July 2024 and its $4 billion valuation in October 2025. The company processes 40 trillion tokens daily for customers such as Shopify, Uber, Notion, MongoDB, GitLab, and Cursor. You can deploy the Mixtral 8x22B model on their serverless API. This model from Mistral AI uses a sparse Mixture-of-Experts architecture with 141 billion parameters. The architecture activates only two experts from its eight expert groups for every token. It handles 64,000 tokens of context and works for multilingual tasks. The company was founded in 2022 by former Meta engineers, including Lin Qiao. Lin Qiao holds a bachelor’s and master’s degree in Computer Science and previously served as a leader in the platform architecture of Meta AI. The company has about 200 employees and intends to reach 600 by the end of 2026. The company also acquired the game server hosting company Hathora in March 2026. The platform provides a Serverless API where developers invoke models with a single line of code. Users pay per token for these services. The company reached a $552 million valuation in July 2024 after a $52 million Series B round. This round was co-led by Lightspeed Venture Partners, Index Ventures, and Evantic. Prior to this, the company secured $25.29 million in a Series A round in March 2024. The company’s annualized revenue in February 2026 was $315 million, and it has since exceeded $1 billion. The company also has a strategic partnership with Microsoft to integrate its inference services into the Model Foundry platform.
Serverless deployment vs dedicated hardware
The platform provides on-demand GPU compute including H100, H200, B200, B300, and GB300 models. The platform provides on-demand access to various GPU configurations, including the H100, H200, B200, B300, and GB300, which allows developers to select the specific compute power required for their model inference workloads. For developers running Mixtral 8x22B in fp16 resolution, the model requires 4x 80 GB GPUs or 8x 48 GB GPUs. The 8x22B model uses 39 billion active parameters per token. If you use the smaller Mixtral 8x7B version instead, you only need 2x 80 GB GPUs or 4x 48 GB GPUs. The company uses a self-developed FireAttention inference engine to increase model efficiency. If you need to fine-tune models like Mixtral, the company provides a managed reinforcement learning service called Fireworks Reinforcement Fine-Tuning. The Mixtral 8x22B model has 141 billion total parameters.
| Model | Input Price | Output Price | Latency |
|---|---|---|---|
| Mixtral 8x7B Instruct | $0.50 | $0.50 | 250ms |
| Llama-v3p1-8b-instruct | $0.10 | $0.10 | 180ms |
| Qwen3-235b-a22b | $0.22 | $0.88 | 500ms |
| Llama-v3p3-70b-instruct | $0.90 | $0.90 | 200ms |
The token factory layer turns raw GPU compute into a metered, OpenAI-compatible API. Fireworks AI is a Tier 1 neutral open-model host. It provides serverless per-token pricing with no idle charge. The company also provides dedicated single-tenant endpoints for higher volume or latency-sensitive loads.
Deployment trade-offs for Mixtral 8x22B
Serverless GPU reliability depends on how fast instances wake and how the platform manages shared capacity. Cold starts occur when the platform must load multi-gigabyte weights into GPU memory, which can take several seconds for large models. You should test the idle wakeup time and latency stability through a burst ramp before you commit to a workload. If you are building a latency-sensitive interactive product, you will find that serverless scales-to-zero behavior creates significant issues. I find that serverless pricing is a bad choice for steady, high-utilization workloads because per-request pricing exceeds the cost of dedicated capacity. Latency differences between providers come from model loading time, request queuing, GPU compute time, and response serialization.
The build-vs-buy gap exists because the inference platform between the weights and the GPUs requires months of engineering. A single team running its own fleet of GPUs must manage serving-stack tuning, autoscaling, and GPU-failure handling. A dedicated 8x H100 node at $20 to $25 per hour costs roughly $15,000 to $18,000 per month. Serverless at $0.16 per million tokens reaches that monthly spend at roughly 90 to 110 million tokens. The Mixtral 8x22B model is under an Apache 2.0 license. This license allows users to access, modify, and distribute the model. The model works for code generation and multilingual text tasks. Will the increasing popularity of custom inference silicon eventually erode the market share of GPU-based token factories?
