IBM, Together AI, and Equinix: the money moves to inference clusters for open models

On August 12 IBM and Together AI signed a $240M multiyear deal for an Nvidia inference cluster on IBM Cloud: roughly 2,000 Blackwell chips in HGX B300 form, tied together with Spectrum-X networking, and explicitly aimed at inference of open models at scale, not training. Reuters had it first. Three weeks later, on September 4, the Equinix Inference Exchange surfaced: Nvidia plus Together AI again, 200-plus open models, distributed inference close to where the data sits, available Q1 2027. The money is not going to bigger training runs. It is going to serving open weights near enterprise data.

Why inference, and why open models

Training capacity is a bet on a model that may not win. Inference capacity for open models is a bet on demand that already exists: every enterprise that has decided it wants an option outside the proprietary platforms and has a DeepSeek, Qwen, GLM or Kimi in its eval suite. Together AI's whole business is serving those weights, IBM's is selling to enterprises that will not put data on a consumer API, and Equinix's is being where the data already lives. The deal shape tells you what each party thinks the bottleneck is: not model quality, but the cost per token of running a model you do not own on hardware you do not want to buy.

B300 is the current Blackwell serving part, and the Hopper vs Blackwell comparison explains why the generation matters for inference specifically: memory and interconnect, not peak FLOPs, set the batch size and latency you can hit at a given cost.

The open-weight economy does not run on training clusters; it runs on somebody else's inference cluster sitting next to your data.

Three ways to pay for an open model

For a team choosing how to serve an open model, the IBM deal is a data point for the managed tier, and it sharpens the comparison.

• Managed (Together, IBM Cloud, Equinix from 2027): pay per token, zero ops, someone else's batching. Best when traffic is bursty or you cannot staff serving. Watch promo cliffs and provenance questions on Chinese models.
• Self-hosted on rented or owned GPUs: the H100 vLLM team serving math applies; you win on cost only once utilisation is high and steady, and you own the pager.
• Local on unified memory: zero marginal cost, bounded throughput, strong data story. Right for dev boxes and small teams, wrong for a customer-facing API.

The crossover between managed and self-hosted is a utilisation question. A cluster like IBM's is built to keep 2,000 chips busy across many tenants; a single team's traffic almost never does that, which is why managed wins at low volume and self-host wins only at sustained high volume. Measure your tokens per day before you price a rack.

The Equinix angle: inference near data

Distributed inference is the part I would watch. Q1 2027 is a long way off, but the design intent, 200-plus open models served inside the same facility where a company's databases live, addresses the objection that actually blocks enterprise adoption: data leaving the building. If it lands as described, a lot of the argument for on-prem inference goes away for regulated customers who were only on-prem because of egress, not because of cost.

What to do now

• Get a tokens-per-day and peak-concurrency number for each open-model workload you run.
• Price it three ways: managed per token, self-host at realistic utilisation, local where it fits.
• Route through one abstraction so the managed provider is swappable; OpenRouter-style routing makes the price comparison a config change.
• Note the September 4 context: Nvidia confirmed its $12.93B acquisition of Hugging Face the same day, putting the model marketplace and the compute under one owner, with a stated hardware-neutral commitment to watch.

The honest limitation

$240M and 2,000 chips are the announced figures; nobody has published what IBM will charge per token or what utilisation the cluster will actually run at. Equinix is a 2027 promise with no pricing attached. The direction is clear. The cost per token, which is the whole point of the exercise, is not public yet, and until it is, the managed column in your comparison is an estimate.