There is a moment in every infrastructure cycle where the economics tip. I have watched it happen with bandwidth in the mid-nineties, with Linux displacing commercial Unix in the early 2000s, and with object storage making SAN look like a mortgage payment you took out on a house you did not need. The moment usually does not announce itself with a press release that says "the old way is now too expensive." It announces itself with a deal that seems boring on the surface and turns out to be a structural fault line.
The deal IBM and Together AI signed on August 11, 2026 is that kind of fault line.
IBM is investing $240 million in a multi-year agreement to build a large-scale AI inference cluster on IBM Cloud, powered by Nvidia's HGX B300 systems — the Blackwell 300-generation hardware — connected over Nvidia's Spectrum-X Ethernet fabric. The initial cluster will house roughly 2,000 Blackwell chips and is targeted to go live in Q1 2027. Together AI, now valued at $8.3 billion as of July 2026, will run its platform on top of that hardware. Their platform serves open-source models — DeepSeek, MiniMax, Kimi, and whatever ships next — at prices positioned well below what AWS, Azure, and Google charge for equivalent workloads on their proprietary APIs.
Together AI's chief revenue officer noted the cluster will likely be sold out two to three months before it goes live. That is not a capacity forecast. That is a statement about where enterprise demand is moving.
What "Open Inference" Actually Means
People sometimes use "open source AI" as a fuzzy umbrella for anything that is not OpenAI. That framing misses the point. What Together AI offers is something more specific: a hosted inference platform where the underlying model weights are public, where you are not locked into a single vendor's fine-tuning pipeline, and where the pricing is driven by raw compute cost rather than by what the market will bear for a proprietary capability.
The distinction matters operationally. When you call a proprietary API, you are renting access to a capability you cannot audit, cannot run on-premises, and cannot reproduce if the provider changes pricing, changes the model, or discontinues a version. For most consumer applications, that tradeoff is fine. For a bank's document processing pipeline, a hospital's clinical summarization system, or any regulated environment with data residency requirements, it is a recurring risk that sits on someone's audit checklist.
Together AI is currently serving approximately 400 trillion tokens per month across its platform. That number did not come from hobbyists. That volume is enterprise workloads: document ingestion, code generation, internal search, support automation. The customers paying for that inference have already decided that open-weight models are good enough for their use case. What they are negotiating now is the price and the SLA.
IBM's Reading of the Room
I have spent enough time watching IBM strategic pivots to be reflexively skeptical of them. Watson never lived up to its keynote. The Red Hat acquisition took years to find its footing in IBM's sales culture. So when I see IBM put $240 million behind a bet, I look for whether the underlying logic is sound — not whether IBM is the right company to execute it.
In this case, the logic is sound. The Next Web's coverage of the deal captures IBM's framing directly: "the money in artificial intelligence is no longer only in building the cleverest model, but in running it cheaply." That is a correct reading. The race to produce frontier models is a game for a handful of companies with hundred-billion-dollar compute budgets. The race to run those models efficiently at scale is a game that every cloud provider and enterprise infrastructure vendor can participate in.
IBM is not trying to out-model Anthropic or Google. They are trying to be the platform that runs whatever model you choose, cheaper than the hyperscalers charge you to run it on theirs. That is a defensible position if you have the hardware and the pricing discipline. The HGX B300 cluster gives them the hardware. Partnering with Together AI gives them a platform that has already figured out the inference optimization stack.
The Hyperscaler Tax
Let me be direct about what this deal is taking aim at, because the press coverage tends to soften it. AWS Bedrock, Azure OpenAI Service, and Google Vertex AI charge a markup over raw compute cost in exchange for the managed API, the SLA, the integration with the rest of their ecosystem, and the convenience of not having to think about infrastructure. For organizations that are deeply embedded in one cloud ecosystem, that markup is often worth it.
But "often worth it" is doing a lot of work in that sentence. AI inference tokens are not a commodity you buy once. They are a recurring operational cost that compounds with every new workflow you automate. A team running a modest customer support summarization pipeline might process 50 million tokens a day. Scale that across a mid-size enterprise with a dozen AI-augmented workflows and you are looking at serious spend — spend that is sensitive to per-token pricing at a level that commodity cloud compute was not.
The customers Together AI has already won — banks, hospitals, enterprises with data security concerns — are exactly the customers who do the math on that spend and find proprietary APIs expensive. They are also the customers who are most motivated by the ability to keep sensitive data away from a model that might be used to train a future version of a competitor's product.
Open-weight models eliminate that concern structurally. If you run inference on DeepSeek's weights via Together AI's platform on IBM Cloud in a region that meets your compliance requirements, your data does not leave your control perimeter in the same way it does when you send it through a third-party model provider's training pipeline.
The Hardware Layer Has Chosen Sides
One detail in this deal that I think is under-discussed: it is built on Nvidia HGX B300 with Spectrum-X Ethernet, not InfiniBand. Spectrum-X is Nvidia's Ethernet-based networking for AI clusters, and its adoption alongside Blackwell is a signal that inference workloads — as opposed to training workloads — can be served efficiently over standard Ethernet fabric at this scale.
This matters for the infrastructure economics. InfiniBand clusters are expensive to build and operate and require specialist knowledge to tune. Ethernet-based AI clusters are operationally closer to what most infrastructure teams already manage. As inference becomes the dominant GPU workload — which the volume numbers suggest it already is, or soon will be — the ability to run it on Ethernet-native infrastructure rather than InfiniBand-native infrastructure changes the total cost of ownership calculus significantly.
Together AI's platform is serving 400 trillion tokens monthly right now, before this cluster is live. When 2,000 Blackwell 300 chips come online in Q1 2027 — already sold out months ahead of schedule — that number will climb. The demand is not theoretical. It is contracted.
What This Means If You Run Systems
I want to be concrete about the practical implications, because this is not just a business story about IBM's pivot strategy.
If you are evaluating AI infrastructure for 2027 budget cycles, this deal sets a reference point: dedicated open-inference clusters built on current Blackwell hardware, operated by a platform tuned for open models, positioned as cheaper than hyperscaler proprietary APIs. That is a real alternative, with real production volume behind it, not a startup claim.
If you are already running workloads on proprietary APIs and you have not done a cost comparison against Together AI's current pricing for equivalent open models, you should do that math now. The gap is not always decisive — sometimes the proprietary model is genuinely better for your use case, sometimes the ecosystem integration is worth the premium — but you should know what the gap is before your next renewal negotiation.
And if you are building internal platforms for AI access within an enterprise, the architecture this deal represents — dedicated inference cluster, open models, standard Ethernet fabric, IBM Cloud control plane — is a blueprint worth understanding. The days of "we use OpenAI because it is the obvious choice" are ending. The infrastructure for running the alternative at scale is being built out, and the economics are beginning to force the conversation.
The Linux Parallel
I have been in this industry long enough to remember when Linux was what you ran if you could not afford Solaris or AIX. Then it was what you ran if you wanted to save money but did not mind some operational complexity. Then it was what you ran if you were serious about operations, because the ecosystem was deeper, the skills were more available, and the economics were simply better.
We are at the "you run it if you want to save money but do not mind some operational complexity" phase for open-weight AI models. IBM and Together AI just made a $240 million bet that the next phase — "you run it if you are serious about AI operations" — arrives faster than anyone is planning for. Based on the token volumes and the sold-out demand, I would not bet against them.
The moment the economics tip does not always feel dramatic. Sometimes it is a deal press release on a Tuesday in August. Watch this one.