Up Close
Microsoft turns its own AI chip from backup plan into default infrastructure
Microsoft announced that its in-house Azure Maia AI accelerator family is intended to serve its own products and services, introduced a third-generation Maia 300 system, and set out plans to scale Maia deployments to multi-gigawatt capacity from 2027. Inference is the stage where a trained model answers prompts for users, meaning the hardware that carries the running cost of everyday AI use. This appears to shift Microsoft from buying capacity around external chips to making its own silicon the default cost and supply lever for the workloads it controls. If that holds, the effect reaches beyond Microsoft: customers and rivals alike are being shown that AI infrastructure advantage now depends not only on model access, but on who can control the serving layer beneath it when usage becomes routine and power-hungry.
404 Media reported that Graylark Technologies' Raven compares a single interior photo with listings on sites such as Airbnb and Zillow, and that the Los Angeles Police Department and Miami-Dade Sheriff's Office have bought access to it. This suggests the operational AI problem is no longer only whether a model can generate or classify, but whether institutions can deploy such systems before clear limits are set around what they may infer from ordinary images. The effect is wider than one police tool: once image models are joined to property databases and sold into agencies, governance has to cover real use in the field rather than hypothetical misuse in a lab.
Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full, with the former offered through the Meta Model API and the latter reserved for Muse Code subscribers.