Anthropic deal taps IREN, while Microsoft eyes broader AI inference hardware with Maia 300

Microsoft is positioning its AI roadmap around inference, the stage of AI work where already-trained models generate outputs for users, not the training phase that builds those models. The emphasis is consistent with a broader industry shift toward optimizing the cost and latency of running AI systems in production. In parallel, a separate report indicates Anthropic has signed an undisclosed deal with IREN, another move that suggests new supply relationships are forming to support deployment of AI models.

The latest reporting, circulated through Yahoo Finance and attributed to The Information, says Anthropic inked an undisclosed deal with IREN. The companies did not make the deal terms public in the material provided, including what specific infrastructure, compute capacity, or deployment timeline is involved. Without those details, it remains unclear whether the agreement centers on a particular data center footprint, a specific customer workload, or a broader commercial arrangement.

Microsoft, for its part, is looking to extend its AI chip strategy with Maia 300, a newer generation in its line of inference chips designed to speed and reduce the cost of running large-scale AI workloads. Inference chips matter because many enterprises do not run their own training from scratch, instead relying on model providers or pre-trained systems. That makes per-request efficiency, energy consumption, and throughput especially important as usage expands.

Maia 300 is described in the circulated coverage as part of Microsoft’s plan to broaden its AI inference chip footprint. The core idea is to support more deployment scenarios, from Azure cloud workloads to large-scale production environments that need predictable performance. For Microsoft, which sells both cloud capacity and AI tooling, shifting more compute to specialized inference hardware can improve unit economics over time, particularly when inference demand grows faster than new training runs.

The coverage also points to what it calls an “inference compute market” being shaped around “liquid” models of compute delivery, meaning a more dynamic approach to obtaining and allocating AI processing resources rather than a strictly static, long-term provisioning model. While the specific mechanism behind “liquid inference compute” is not spelled out in the provided material, the phrase aligns with broader efforts in the industry to match AI resource availability to fluctuating demand, potentially using more modular capacity and clearer performance isolation between workloads.

Still, several key points are missing from the information available here. The exact structure of Anthropic’s arrangement with IREN, including what IREN provides and how long the agreement runs, was not disclosed in the material provided. Likewise, the details of Microsoft’s Maia 300 rollout, such as which Azure regions or customer segments receive the chips first, were not included in the provided packet. Without those specifics, the near-term impact on customers and timelines can only be inferred rather than confirmed.

Going forward, market watchers will likely look for confirmation from Microsoft on Maia 300 availability and performance claims, including any public documentation on supported AI workloads and deployment constraints. For Anthropic and IREN, the key follow-up is whether either side later clarifies the scope of the undisclosed deal, such as the compute capacity involved or the intended use case. Together, these developments underscore a central industry question in 2026: not just which models win, but which compute ecosystems can deliver them reliably and efficiently.