Fireworks AI is making a direct bet that the future of enterprise AI runs on custom-tuned open models, not rented access to closed ones. The company closed a $1.5 billion Series D at a $17.5 billion valuation, with NVIDIA and a cohort of major growth funds participating. Revenue has hit $1 billion annualized, up fivefold year-over-year. Daily token volume jumped from 15 trillion to over 40 trillion. Customers include Uber, Shopify, and GitLab, plus AI-native firms like Harvey and Cursor. A Fireworks AI Series D<\/a> of this scale signals that the open-model serving layer is hardening into real infrastructure, not an experiment.<\/p>\n The number that should reframe your vendor map is 95%. Fireworks reports that 95% of tokens processed on its platform now come from custom-tuned models rather than standard foundation models, meaning the market has already moved. If your AI stack is still primarily API calls to a closed frontier model for production workloads, you’re running an architecture that the fastest-moving enterprises have already iterated past. The question isn’t whether to evaluate open-model fine-tuning, it’s whether you’re late enough that a competitor has already built the proprietary data advantage you’re still planning.<\/p>\n The Microsoft distribution deal matters more than it first appears. Fireworks models now surface inside the Microsoft Azure ecosystem, backed by compute from over 20 infrastructure providers. That’s not a niche play: it’s a direct insertion into the procurement path that most large enterprises already use. CTOs who assumed open-model fine-tuning required standing up bespoke infrastructure are looking at a different calculus now. The capability is becoming a line item in existing cloud agreements, which removes the activation energy that kept many teams on managed closed-model APIs.<\/p>\n The Fireworks growth story, compelling as it is, is narrated partly by investors who funded the round and by a company describing its own metrics, so the 5x revenue figure and the token volume jump deserve independent validation before they anchor a build-versus-buy decision. What’s harder to dismiss is the customer list and NVIDIA’s participation, which reflects supply-side conviction about where inference workloads are heading. If you have a model-serving contract renewal coming up, the comparison point has shifted: cost-per-token and data-residency control are now legitimate negotiating levers, not aspirational ones.<\/p>\nWhat this means for your business<\/h2>\n