IBM and Together AI Set Out Enterprise Inference Capacity Partnership

IBM and Together AI have described an arrangement to provide enterprise customers with dedicated capacity for running artificial-intelligence models, with workloads served from managed accelerator clusters rather than shared public endpoints.

The companies said the arrangement is aimed at organisations in regulated sectors that require predictable latency, defined data residency and contractual guarantees about where inference occurs. Customers would run open-weight models on reserved hardware rather than sending requests to a shared multi-tenant service.

Demand for inference capacity has grown faster than demand for training capacity over the past year, as organisations move from evaluating models to embedding them in workflows that run continuously.

Berean News Technology Analysis

The shift from training to inference changes what infrastructure needs to be good at. Training rewards raw throughput in large concentrated bursts; inference rewards consistent low latency, high utilisation and predictable cost per request. Those are different engineering problems and different economics.

Editorial Verdict

This is the least glamorous part of the AI build-out and probably the most durable. Training budgets are discretionary and can be cut in a downturn; inference spending tracks usage of systems already embedded in operations. Sustained growth here would be a better indicator of real adoption than any model launch.