Fireworks AI — AI supply-chain exposure

Fireworks AI is an independent inference provider, founded by engineers who led PyTorch at Meta, that specialises in serving open-weight models fast and cheaply at scale. Where a closed lab both trains and serves its own model, Fireworks runs the growing universe of openly released models for companies that want to build on them without operating their own infrastructure — optimising the serving stack so more tokens move through each chip. In the chain it is a producer of inference, one of the larger independent ones by the volume of tokens it processes.

Its hook is the open-weight thesis: if capable models keep being released openly, value migrates from owning the model to serving it well, and specialised inference shops become the practical on-ramp. That is a real and expanding demand pull, but a competitive one — Fireworks leans on engineering depth in model optimisation rather than on any input the rest of the chain must buy from it. The model places it mid-chain: strong exposure to the inference wave, positioned on speed and cost rather than a structural chokepoint.