Together is a GPU reseller
CoreWeave
This setup means Together wins by packaging scarce compute into an easier product, not by owning the underlying infrastructure. CoreWeave buys or finances large blocks of Nvidia GPUs and rents them through long contracts, while Together takes capacity from CoreWeave, Google Cloud, Lambda, and other fragmented sources, then resells it to startups through APIs, hosted models, and training tools. That lets Together move faster downmarket, but it also leaves its margins and reliability tied to suppliers upstream.
-
CoreWeave and Lambda are closer to landlords. They commit capital up front for GPUs and data centers, then push customers toward reservations that help pay back that fixed cost. Together is closer to a retailer. It rents capacity, marks it up, and bundles it with model serving and fine tuning software for teams that do not want to negotiate infrastructure directly.
-
The customer split follows that business model. CoreWeave is built for enterprises willing to reserve 1,000s of GPUs on yearly contracts. Together is built for startups that want usage based access, open source model support, and a faster path from experimenting with Llama or image models to running them in production.
-
This is why Together could grow quickly without matching CoreWeave's asset base. Together reached an estimated $26M run rate by the end of 2023, $130M by the end of 2024, and $300M by September 2025, while CoreWeave scaled from $229M revenue in 2023 to $1.9B in 2024 and $5.1B in 2025 by owning more of the supply chain.
Going forward, the gap between infrastructure owners and compute resellers should widen. CoreWeave is moving deeper into capital intensive supply control, while Together is moving deeper into developer experience and open source model distribution. The more GPUs become available, the more Together's advantage will depend on software, workflow, and customer acquisition rather than simple access to chips.