Together AI just committed $240 million to a multiyear deal with IBM Cloud, and the striking part isn’t the price tag. It’s that the company expects the capacity to sell out two to three months before it even goes live.
The deployment centers on roughly 2,000 NVIDIA Blackwell 300 GPUs, running on HGX B300 systems connected through NVIDIA’s Spectrum-X Ethernet networking, based in the US and set to come online in early 2027. It marks IBM Cloud’s first large-scale inference cluster built specifically around HGX B300 hardware, and NVIDIA says the configuration delivers roughly 30 times the AI factory output of previous generations.
Together AI’s chief revenue officer, Kai Mak, told Reuters the company already expects demand to outpace supply well before launch. That tracks with what’s happening across Together AI’s business more broadly. The company says its inference service now processes more than 400 trillion tokens monthly, up from just 30 billion nine months ago, a jump that’s hard to read as anything other than explosive.
This isn’t Together AI’s only move to secure compute, either. Alongside an $800 million Series C round that valued the company at $8.3 billion, it locked in commitments for more than 500 megawatts of additional capacity, which new investors will finance separately. The company also rolled out a Provisioned Throughput service in July, letting customers reserve guaranteed inference capacity measured in tokens per minute rather than managing raw GPU allocation themselves.
CEO Vipul Ved Prakash framed the IBM partnership around a specific gap he sees in the market: enterprises want performance comparable to closed frontier models, but without the pricing that comes attached to them, and that only works if the underlying infrastructure can actually deliver at scale.
The deal builds on an existing IBM-NVIDIA relationship rather than starting a new one. IBM expanded its NVIDIA partnership back in March, when it announced plans to bring Blackwell Ultra GPUs to IBM Cloud, and the two companies have separately been working together on storage infrastructure and sovereign AI deployments for regulated workloads.
Together AI isn’t alone in locking down capacity well ahead of deployment, either. CoreWeave reported a $104.2 billion revenue backlog this quarter, and Nebius disclosed more than $40 billion in customer commitments, suggesting the entire AI infrastructure market is increasingly running on capacity that sells out before anyone even builds it.
