Together AI, an inference platform provider that rents compute from cloud and neocloud operators, has struck a $240 million deal with IBM Cloud to run a large cluster of Nvidia GPUs. The arrangement highlights how scarce AI infrastructure capacity has become, pushing even competing service providers to work together when the hardware is available.
According to IBM’s announcement, Together AI chose IBM and Nvidia for their product roadmaps and their ability to deliver GPU capacity quickly enough to keep pace with AI scaling demands while keeping token costs low. In practice, that means IBM had the capacity Together AI needed, when it needed it.
What’s being deployed
The deal will bring Nvidia’s HGX B300 platform online, with compute capacity expected in the first quarter of 2027. IBM says this will be its first large-scale use of B300 systems specifically for inference workloads.
The HGX B300, announced in early 2025, is not Nvidia’s flagship rack-scale offering. Unlike the 72-GPU systems Nvidia typically showcases, the HGX B300 is a more conventional design that can run in standard air-cooled datacenters. Each box packs eight B300 GPUs, drawing 14 to 15 kW, and connects them via NVLink internally and Nvidia’s Spectrum-X Ethernet fabric between units.
A pattern of cross-cloud deals
Together AI’s business model centers on renting GPU compute from various providers and offering an OpenAI-compatible API for inference, fine-tuning, and training. The company is largely hardware-agnostic, prioritizing price and performance over any single vendor relationship.
Beyond the IBM deal, Together AI has been expanding its own datacenter footprint in Maryland, Memphis, and Sweden. It is also deploying services on SambaNova’s heterogeneous compute platform, built with Intel and using Nvidia GPUs for prefill processing, which recently went live in a new AI-focused datacenter operated by Vector Core Compute.
Why it matters for the hosting industry
The deal illustrates a broader trend: with power, datacenter capacity, and GPU supply chains all under strain, AI service providers are increasingly willing to source compute from whoever has it available, including direct competitors. For hosting and infrastructure operators, it signals continued strong demand for GPU capacity well into 2027, even for hardware generations that are not Nvidia’s newest.