IBM and Together AI sign a $240 million deal for Blackwell Ultra inference on IBM Cloud

IBM and Together AI signed a multi-year agreement worth roughly $240 million, announced August 11, 2026. Under it, IBM will host a large cluster of Nvidia Blackwell Ultra chips on IBM Cloud for Together AI’s open-source model inference.

IBM said it will deploy Nvidia HGX B300 systems with Spectrum-X Ethernet networking, with the infrastructure expected to be available in the first half of 2027. The cluster will let Together AI run inference workloads for enterprise customers using open-source AI models.

The deal matters because inference — running trained models in production — is becoming the larger, more durable slice of AI spending, and open-weight models are the ground both companies are betting on. Together AI Chief Executive Vipul Ved Prakash said enterprises want frontier-model performance without the closed-model price tag, which only works if the underlying infrastructure is fast and reliable at scale.

For IBM, hosting a specialist inference operator is a way to fill its cloud with GPU demand it does not have to generate itself; for Together AI, IBM’s capacity extends its reach without owning data centers.

The dollar figure and 2027 timeline are the companies’ own, and the capacity is more than a year from going live. Whether open-source inference at this scale pays off depends on enterprises choosing open-weight models over the closed frontier APIs — a shift that is underway but far from settled.