Together AI $5B Capacity Buildout

This deal pushes Together from a software layer riding on other peoples GPUs into the small group of AI clouds that can shape supply, pricing, and customer mix with their own large clusters. The Humain buildout roughly triples current data center capacity, adds 250 MW and 120,000 chips, and is expected to support about $5B of annual business once fully deployed, versus Together reaching a $1B annualized revenue run rate in February 2026.1 2 3

• The jump matters because Together already makes most of its money from renting Nvidia GPU capacity for training, fine tuning, and serving, while only 30 to 40 percent comes from per token APIs. More owned and controlled capacity means better gross margins than the roughly 45 percent mix generated while relying heavily on CoreWeave and Lambda.1
• Compared with peers, this is the kind of scale usually tied to anchor contracts. CoreWeave grew to $2B of 2024 revenue on the back of a $10B Microsoft contract, and Lambda shifted upmarket around large multi year deployments with Nvidia and Microsoft. The Saudi cluster gives Together a similar path into giant reserved capacity deals.4 5
• Operationally, this is not all live at once. The first 7,000 chips are expected by November 2026, with full allocation in 2027. That means the $5B figure is best read as mature run rate potential after the cluster is filled and sold through, not near term recognized revenue.1 2

The next phase is a shift from startup focused bursty demand toward a more mixed book of sovereign, frontier lab, and enterprise workloads. If Together keeps layering model APIs and developer tools on top of owned infrastructure, it can move closer to CoreWeave and Lambda economics while still selling a more developer friendly product surface.1 4 5