Cloud Open Weight Models Are Turning Inference Into A Control Point By Janakiram MSV ,
Forbes contributors publish independent expert analyses and insights. I cover emerging technologies with a focus on infrastructure and AI Follow Author Jul 18, 2026, 10:01pm EDT --:-- / --:-- This voice experience is generated by AI. Learn more . This voice experience is generated by AI. Learn more . Summary AI infrastructure firm Fireworks recently secured a $1.5 billion Series D at a $17.5 billion valuation, leading a wave of significant investments in companies providing managed services for open-weight and custom AI models. Alongside Baseten and Together AI, these firms have collectively raised nearly $4 billion in weeks, aiming to offer more cost-effective inference solutions than proprietary frontier APIs. They specialize in handling the complex operational aspects of deploying AI models, enabling businesses to customize and run models without managing their own GPUs. However, challenges loom, including competition from hyperscalers, the persistent performance gap with leading closed models, and undisclosed gross margins. Enterprise buyers are urged to scrutinize portability and margin structures. The growing competition ultimately empowers buyers in the evolving AI landscape.
Network Panel Pexels Fireworks announced a $1.505 billion Series D at a $17.5 billion valuation on July 15, with the news circulating through July 16. Atreides Management, Index Ventures and TCV led it. Fireworks develops no general-purpose frontier foundation model of its own.
Three companies in that business have raised roughly $3.8 billion between them in under four weeks. All three compete in managed infrastructure for training, adapting or serving open-weight and custom models, with inference at the center of their pitches. Their shared economic proposition is to run a customer’s own model in production for less than a frontier API would charge.
An open-weight model gives users access to trained parameters, and those parameters alone are not a production service. Standing one up means continuous batching, cache management, quantization and autoscaling, plus latency guarantees, observability and a billing system that survives an audit.
That operational work is the product these companies sell. Customers can customize an open-weight model on their own claims data or support tickets. They then serve it through an endpoint without procuring or operating GPUs themselves. Fireworks said more than 95% of the tokens it serves come from models specialized on customer data. That figure covers fine-tuned open weights, adapters, distillations and customer-trained artifacts together. It shows demand for specialization rather than a preference for downloadable third-party models.
Baseten closed roughly $1.5 billion around June 22 at a headline valuation of $13 billion. The Wall Street Journal reported a dual-tier structure, with some investors entering at $11 billion. Sacra, a research firm, estimated its annualized run rate at roughly $600 million by March 2026, up from about $200 million in December 2025. Those are outside estimates, not observed revenue.
Together AI followed on July 1 with an $800 million Series C at an $8.3 billion valuation, led by Aramco Ventures. The company said annual bookings crossed $1.15 billion in the prior quarter and that usage of open models had tripled over the prior year. Bookings are not the same thing as revenue. Fireworks reported an ann…