Fireworks and Together AI Raise $2.3 Billion in Inference Bet Over Pre-training

Fireworks AI raised a $1.505 billion funding round, while Together AI completed an $800 million round, signaling a structural shift in directing venture capital dollars in artificial intelligence from pre-training toward inference.

Both companies serve enterprises that need to run open-source and fine-tuned models at scale without falling into the trap of high costs and lock-in to proprietary APIs. Fireworks' annual revenue run rate surpassed $1 billion in an unprecedented milestone for an inference cloud company at this stage, with its round including NVIDIA, Index, TCV, and Lightspeed.

The key analytical question centers on whether cloud inference platforms will maintain their margins or become commoditized infrastructure crowded out by hyperscalers; current evidence points to a middle path: players capable of optimizing models, building hardware partnerships, and offering enterprise SLAs retain a defensible competitive edge.

Analysts point to the next candidates for similar rounds in the second half of 2026, led by Baseten, Anyscale, and Modal; meaning that the battle over open-source model inference is far from decided, and the market is likely to produce one or two winners with a significant margin at the expense of the rest.

What do these terms mean?

Inference: The phase in which the model is used after training to answer queries and generate outputs, in contrast to training which teaches the model from scratch.

Pre-training: The costly phase of building large AI models, consuming thousands of chips for weeks or months to teach the model language patterns.

Hyperscalers: Cloud computing giants like AWS, Azure, and Google Cloud, capable of offering similar services at lower costs due to economies of scale.

Service Level Agreement (SLA): A contract defining the guaranteed minimum service performance, response time, and availability, which is an essential requirement for large enterprises.