Together AI Shifts to Owned Data Centers

This shift shows Together is trying to stop being just a smart reseller of other people’s GPUs and become an owner operator of the capacity that sets its margins, reliability, and scale ceiling. Early on, Together won customers by wrapping rented CoreWeave and Lambda infrastructure in easier token based APIs and startup friendly rentals. Owning data center capacity changes the business. It lets Together keep more of each training and inference dollar, guarantee privacy for dedicated workloads, and package reserved clusters and serverless APIs on top of hardware it controls.1 2 3 4 5

• The original model sat one layer above the neoclouds. Together sold per token API access and bundled GPU jobs, while CoreWeave and Lambda focused more on hourly or reserved infrastructure contracts. That made Together flexible for startups, but left margin and supply control with upstream providers.1 2
• The infrastructure strategy is now visibly different. Together says it has its own data centers in Maryland, Memphis, and Sweden, over 200 MW of capacity, and secure hosting in its own facilities for privacy sensitive model deployments. It has also launched reserved cluster products and larger dedicated capacity offerings.3 4 5 6
• This is the same direction the category leaders took. CoreWeave used long term contracts to finance rapid data center expansion, and Lambda moved away from self serve inference toward large reserved training deals. Together is following that playbook while keeping its open model API layer on top.2 7 8

The next phase is a convergence of software layer and infrastructure layer. As Together adds owned capacity and large power commitments, including 500 MW of compute commitments around its Series C, it can move up from brokering GPU supply to shaping market prices and capacity access for open model developers. That should make the API and rental businesses reinforce each other more tightly over time.5 6