Microsoft AI developments (September 2025)
- **Fairwater – the world’s most powerful AI datacenter** – On Sept 18 Microsoft unveiled Fairwater, its newest U.S. AI datacenter. Fairwater houses hundreds of thousands of Nvidia GPUs and connects them with a single “flat” network that delivers 865 k tokens/sec throughput per rack. The datacenter uses NVLink and NVSwitch to pool memory across GPUs and arranges racks vertically to reduce network distances. Over 90 % of the facility’s cooling comes from a closed‑loop liquid system fed by recirculated water and a 250 MW solar projectblogs.microsoft.com. Microsoft plans to build multiple identical Fairwater facilities across the U.S., spending more than $7 billion and creating thousands of jobsblogs.microsoft.com.
- **Microfluidic chip‑cooling breakthrough** – On Sept 23 Microsoft revealed a microfluidic cooling technique, developed with Swiss startup Corintis, that carves hair‑width channels into the back of silicon chips. Coolant flows directly over the chip, and AI designs the channel patterns (inspired by leaf veins) to target hotspots. This method removes heat up to three times better than traditional cold plates and reduces temperature spikes by 65 %, allowing chips to operate at 70 °C without throttling and enabling higher power densities and overclockingnews.microsoft.com. The technology will be integrated into future Maia and Cobalt chipsnews.microsoft.com.
- **New capabilities in Microsoft Fabric** – At FabCon Europe (Sept 16) Microsoft announced major upgrades to its Fabric data‑analytics platform:**Graph in Fabric** – a low/no‑code graph database to discover hidden relationships across data sets.**Maps in Fabric** – geospatial analytics for location‑based data.**OneLake shortcuts & mirroring** – quickly unify data from Oracle, Google BigQuery and Azure Blob; convert JSON or Parquet to Delta tables and add security/diagnostics features.**Fabric Extensibility Toolkit & Model Context Protocol** – tools that bring Visual Studio Code and GitHub workflows to Fabric and allow AI‑assisted code generation and item authoring.Integration with **Azure AI Foundry** to build agentic applicationsblogs.microsoft.com.
- **Graph in Fabric** – a low/no‑code graph database to discover hidden relationships across data sets.
- **Maps in Fabric** – geospatial analytics for location‑based data.
- **OneLake shortcuts & mirroring** – quickly unify data from Oracle, Google BigQuery and Azure Blob; convert JSON or Parquet to Delta tables and add security/diagnostics features.
- **Fabric Extensibility Toolkit & Model Context Protocol** – tools that bring Visual Studio Code and GitHub workflows to Fabric and allow AI‑assisted code generation and item authoring.
- Integration with **Azure AI Foundry** to build agentic applicationsblogs.microsoft.com.
- **Global AI infrastructure expansion** – Besides the Wisconsin datacenter, Microsoft announced a hyperscale AI datacenter in Narvik, Norway through a joint venture with nScale and Aker, and plans to build the UK’s largest supercomputer with nScale in Loughton. These projects continue Microsoft’s push to build purpose‑built AI datacenters worldwideblogs.microsoft.com.
Major AI industry news (September 2025)
- **OpenAI and NVIDIA’s 10‑GW partnership** – On Sept 22 OpenAI and NVIDIA signed a letter of intent to deploy at least 10 GW of NVIDIA systems across U.S. data centers for OpenAI’s next‑generation models. NVIDIA will invest up to $100 billion in OpenAI as each gigawatt of capacity is built, and the first gigawatt will come online in 2H 2026 on the new Vera Rubin platformnvidianews.nvidia.com. NVIDIA CEO Jensen Huang called it the “next leap forward” in deploying massive compute powernvidianews.nvidia.com, while OpenAI’s Sam Altman said compute infrastructure will underpin the future economynvidianews.nvidia.com.
- **NVIDIA Rubin CPX GPU** – At the AI Infra Summit (Sept 9) NVIDIA unveiled the **Rubin CPX** GPU, built for massive‑context inference. The NVL144 CPX platform packs 8 exaflops of AI compute and 100 TB of memory in a single rack—7.5× more AI performance than previous systems. Each CPX chip delivers 30 petaflops of compute and 3× faster attention for long‑context tasks, and uses cost‑efficient GDDR7 memorynvidianews.nvidia.com. Early adopters like Cursor and Runway plan to use CPX for million‑token coding and generative videonvidianews.nvidia.com.
- **Google integrates Gemini into Chrome** – On Sept 19 Google announced that its Gemini AI assistant is being built directly into the Chrome browser for U.S. users. The company called it “the biggest upgrade to Chrome in its history.” Google said agentic features that can book appointments or make purchases will follow, and users will soon be able to enable AI‑mode search from the address barthecurrent.com. The announcement came two weeks after a judge ruled that Google must share some search data with rivals but does not have to divest Chromethecurrent.com.
- **Meta’s Ray‑Ban Display smart glasses** – At Meta Connect on Sept 17, Meta debuted the **Meta Ray‑Ban Display**, its first consumer‑ready smart glasses with a built‑in display. The $800 glasses come with a wristband controller; CEO Mark Zuckerberg said they aim to usher consumers into the AI erareuters.com. Analysts noted Meta is ahead of rivals in easy‑to‑wear smart glasses but the high price could limit adoptionreuters.com.
- **Meta’s Llama AI approved for government use** – On Sept 22, the U.S. General Services Administration added Meta’s Llama AI to its list of approved tools for federal agencies. This allows government agencies to use Llama, signalling broader integration of commercial AI systems into U.S. government workflowsreuters.com.
- **Scale AI launches Seal Showdown benchmarking platform** – On Sept 23, Scale AI introduced **Seal Showdown**, a platform where users test generative‑AI models head‑to‑head and vote on which performs better. The tool captures preferences across demographics, languages and use cases, giving model rankings based on real human feedbackradicaldatascience.wordpress.com.
- **Qwen3‑Omni – a unified omni‑modal model** – On Sept 22, the Qwen team announced **Qwen3‑Omni**, an end‑to‑end model that combines text, images, audio and video. It achieved state‑of‑the‑art results on 22/36 audio and AV benchmarks, handles 119 language tokens or 19 seconds of speech input and 10 seconds of speech output, supports 30‑minute audio understanding and built‑in tool calling, and includes an open‑source captionerradicaldatascience.wordpress.com.
- **CoreWeave–NVIDIA $6.3 B cloud‑capacity order** – On Sept 15, AI cloud provider CoreWeave signed a $6.3 billion initial order with NVIDIA under which NVIDIA will purchase any unused cloud capacity. The deal underscores efforts to secure long‑term access to AI infrastructureradicaldatascience.wordpress.com.
- **Chinese “SpikingBrain 1.0” runs 100× faster** – Chinese researchers unveiled **SpikingBrain 1.0**, a spiking‑neural‑network model that fires neurons selectively like the human brain. The 7B and 76B‑parameter versions matched the performance of traditional models while using <2 % of the training data, and a 4M‑token version processed prompts over 100× faster than standard systems. The model ran entirely on China’s MetaX chips rather than NVIDIA hardwareradicaldatascience.wordpress.com, highlighting progress in hardware‑efficient AI.
- **Other model and tool announcements** – OpenAI released **GPT‑5‑Codex**, a coding agent that dynamically allocates compute time per task, improving performance on complex coding benchmarksradicaldatascience.wordpress.com. Weights & Biases added **GLM‑4.5**, a 300‑billion‑parameter mixture‑of‑experts model optimized for reasoning, coding and agentic tasksradicaldatascience.wordpress.com. Augment Code launched a context‑aware coding agent for navigating millions of lines of production code and automating development processesradicaldatascience.wordpress.com. NVIDIA’s acquisition of Enfabrica’s networking technology (≈$900 M) will allow integrated systems connecting hundreds of thousands of GPUsradicaldatascience.wordpress.com.