Microsoft Targets AI Inference Shift to Maia 300 Chips in 2027

01Microsoft Unveils Maia 300, Plans to Make In-House Chips Its Primary AI Inference System From 2027

Microsoft has unveiled Maia 300, the third generation of its in-house AI accelerator, and plans to begin deploying it to production workloads in 2027. The chip is designed for inference—the process of running trained AI models—and is intended to reduce Microsoft’s dependence on accelerators supplied by Nvidia and AMD.

The company already uses the previous Maia 200 chip for Microsoft Copilot, GitHub Copilot, Microsoft Foundry and its own AI models. Microsoft says Maia 300 is built for newer workloads, including reasoning over long contexts, conversations requiring low latency and sustained, multi-step tasks performed by AI agents.

Microsoft claims Maia 300 will produce 40% more tokens per megawatt than leading GPU-based systems. Tokens are the units of text that AI models process and generate, so the measure links output directly to the electricity needed to run the infrastructure. The company has not provided results from actual production workloads for Maia 300, however, and did not disclose its procurement costs.

At rack scale, Microsoft says a complete configuration of 96 accelerators can deliver as much as seven times the token throughput of Maia 200. That increase would allow each rack to handle substantially more inference work, while the claimed efficiency gain would raise the amount of output available from a fixed power supply.

The underlying design uses a new chiplet architecture, which combines specialized pieces of silicon rather than placing every function in one uniform design. Built using TSMC’s N2P manufacturing process, Maia 300 separates compute, memory-controller and input-output functions. Microsoft says this leaves more silicon area for computation and allows each function to use technology suited to its particular role.

Those improvements underpin Microsoft’s planned expansion of the Azure Maia family to multi-gigawatt deployments beginning in 2027. The company says Maia will then become the primary inference system across its products and services, extending the deployment path established by Maia 200.

That plan would shift a larger share of Microsoft’s AI inference onto hardware it owns and co-designs with its models, reducing its reliance on Nvidia and AMD systems. It does not amount to a completed replacement: deployment has not begun, and Microsoft has not specified an exact 2027 launch date or published production test results showing that Maia 300 meets its performance claims.

02QGIS merges new rule banning all AI-assisted code submissions to its core repository

QGIS is a software project whose core code is maintained in the qgis/qgis repository. It has replaced its disclosure-based policy for AI-assisted contributions with a blanket ban under QEP428, proposed by core contributor Nyall Dawson after the project concluded that its previous safeguards were not being followed in practice.

Under the earlier QEP-408 policy, contributors could use large language models if they reviewed all generated output and disclosed the assistance through an Assisted-by marker. It also prohibited agents from posting comments that had not been reviewed by a person. QEP428 removes that permission entirely: no code produced with help from an LLM, other AI system, or agent-based workflow may be contributed to the repository.

The change followed concerns that human review and disclosure had failed as an enforcement model. In discussion of the proposal, Dawson said the required process “clearly” had not happened in reality. He also said QGIS had already lost at least one extremely high-value contributor as a result, though the proposal did not provide further details about that departure.

The prohibition carries consequences for both the submitted work and the person responsible. If QGIS later discovers that prohibited tools were used, the affected code may be reverted and the contributor may face sanctions against their contribution rights. The policy does not explain how maintainers will identify AI-assisted code or specify a formal process for determining violations.

QEP428’s discussion period ended on October 5, after which the proposal moved to a vote. It was merged into the main branch of the QGIS Enhancement Proposals repository on October 7.

The ban may not be permanent in every form. Dawson said QGIS wanted to begin with the door closed, then consider possible uses individually. Potential future exceptions could include AI used as an autocomplete replacement or for translations, but those examples are not permissions under the current rule. No approval procedure or timetable for reopening those uses has been announced.

03WindBorne Wins U.S. Air Force Contract Worth Up to $49 Million for Balloon Data From Specified Regions

The U.S. Air Force has awarded WindBorne Systems a contract to buy atmospheric observations from the startup’s autonomous weather balloons. WindBorne, founded in 2015 and based in Redwood City, California, operates about 700 balloons worldwide and has previously worked with the Air Force, Navy and National Weather Service.

The new contract has a ceiling of $49 million, but that is not money already committed to WindBorne. The Air Force will pay only when it needs data, placing orders for coverage of particular regions during particular periods. It has ordered $1.7 million of data so far; the source does not specify the size or timing of future orders, or whether spending will eventually reach the contract limit.

WindBorne’s balloons gather atmospheric sensor readings as they travel around the globe. The company sells those observations to customers and feeds them into its own artificial-intelligence forecasting model, WeatherMesh 6. WindBorne says the model produces more accurate forecasts than the traditional physics-based system operated by the European Centre for Medium-Range Weather Forecasts, but that comparison is a company claim and was not established by the Air Force contract.

Although the balloons have no engines, WindBorne can steer them indirectly by changing their altitude. Operators can release air or drop sand carried as ballast, moving a balloon into a wind current that can carry it toward a required area. The approach depends on forecasts that identify useful winds and on having enough balloons positioned around the world.

WindBorne co-founder and chief technology officer Andrey Sushko described the agreement as the largest single contract in which the company has participated. He said it would provide a way to support continued growth of the balloon constellation, alongside research and development intended to improve balloon hardware and performance.

The contract was awarded without competitive bidding. Sushko said the government determined that WindBorne was the only company able to provide this global data-collection capability in a way that met its operational needs. The Air Force’s specific uses for the observations have not been disclosed, and only $1.7 million of the possible $49 million has been ordered.