Capital rotated from training to inference — Baseten and Fireworks raised $1.5B each
Baseten closed a $1.5B Series F and Fireworks AI raised $1.5B within weeks of each other, both for enterprise inference serving. The money has moved from building models to running them, and it happened without a narrative moment to mark it.
Two nine-figure rounds into the same layer inside a few weeks is a thesis, not a coincidence. The implicit bet is that model quality is converging while serving economics are not — that the durable margin sits in latency, throughput, utilisation and cost per token rather than in the weights themselves.
It is a reasonable bet on the evidence. Capability differences between frontier models have compressed to the point where most enterprise workloads are genuinely indifferent between two or three options, while inference cost differences remain large enough to determine whether a deployment is viable. When the product is undifferentiated and the operating cost is not, value migrates to whoever operates it best.
The uncomfortable corollary for model developers is that the layer capturing this capital is the layer that treats their models as interchangeable, and is commercially motivated to keep them that way.
Business Wire — Baseten raises $1.5 billion to power the next era of AI inference → · Pulse 2 — Baseten raises $1.5 billion Series F to power AI inference → · CNBC — Nvidia-backed Fireworks hits $17.5 billion valuation as companies pursue cheaper AI models →