AlphabetActive development
Native PyTorch backend for Google TPUs. The PyTorch project post says TorchTPU is in active development and is used in production for inference and training for some of the largest Google Cloud customers. On 22 October 2026 the repository at github.com/google-pytorch/torch_tpu was public. Its LICENSE file begins with the Apache License, Version 2.0, and also includes an MIT notice for examples/mingpt. The post does not call the stack generally available.
Latest update Oct 22, 2026
01
Open source and in active development as of the PyTorch post read 22 October 2026. Not described as general availability. Production use is stated for some unnamed large Google Cloud customers.
02
Google said TorchTPU, its native PyTorch stack for TPUs, is now open source. The company said existing PyTorch workloads can run on TPUs with small code changes, including through vLLM, SGLang, Hugging Face, and TorchTitan, and that it built the stack with Meta's PyTorch team. The PyTorch project said TorchTPU is in active development and is already used in production for some Google Cloud customers.
03
Google Search is the web search service inside Google Services. The FY2025 10-K says it helps people find information, with trillions of searches a year. In the quarter ended June 30, 2026, Google Search and other revenue was $63,271 million, up from $54,190 million, and the earnings release says that line grew 17%. Google Services as a whole, which also includes YouTube, subscriptions, platforms, and devices, was $94,540 million.
Google Cloud sells infrastructure and platform services, applications, and other enterprise products. The Q2 2026 release says service revenue comes from Google Cloud Platform, Google Workspace, and other enterprise services, and product revenue comes primarily from TPU systems. Cloud revenue was $24,768 million, up 82% from $13,624 million, and Cloud operating income was $8,814 million.
Google Cloud workplace agent announced Oct. 8, 2026 that takes objectives, plans work, and returns finished output across Workspace, Microsoft 365, Slack, CLI, and other channels. Google said it orchestrates Gemini and Anthropic Claude models, can create coworker agents with their own identities, and starts with businesses. 9to5Google reported private preview.
Google fitness tracker with a display, built-in GPS and up to seven days of battery life. Pre-orders opened 12 October 2026 at $179.99, with retail availability set for 27 October. A Stephen Curry special edition is $219.99. The same post describes Health Guardian as background tracking of blood pressure trends, insulin resistance trends, and high or low heart rate alerts, and says availability varies by country.
04
Microsoft
Azure is Microsoft's cloud platform. In the July 29, 2026 release, Satya Nadella said Azure revenue surpassed $100 billion for the first time and was up 41%. The financial statements do not print a separate Azure revenue line. Intelligent Cloud, a wider segment, had quarterly revenue of $39.3 billion, up 32%.
Microsoft
Microsoft-Decision-1 is a decision-scoring model Microsoft opened in Foundry and on OpenRouter on 9 October 2026. It is post-trained from Qwen3.5-9B and returns calibrated probabilities for fixed yes/no, multiple-choice, and rating options. Input tokens are $0.042 per million; output tokens are free. Microsoft says it was about 35 times faster than GPT-6 Sol and the most accurate model in a 36-benchmark comparison kept blind from training. Those latency and accuracy figures were not reproduced on a public suite in this pass, and the post does not show that the Foundry and OpenRouter endpoints serve identical weights.
Microsoft
Microsoft's third-generation inference accelerator, announced 12 October 2026. The blog says production deployments begin in 2027 and that the Maia family will scale to multi-gigawatt deployments starting in 2027 as the primary inference system for Microsoft products and services. A rack holds 96 accelerators in a 167 kW envelope, with eight HBM4 stacks at 18.0 TB/s inside a 950 W TDP, and compute chiplets on TSMC N2P. Microsoft expects up to 7x the rack token throughput of Maia 200 and 40 percent more tokens per megawatt than leading GPU systems. Alistair Speirs told CNBC he would not set a completion date beyond that start. CNBC reported the accelerators cannot be rented by the hour.
Microsoft
Microsoft AI said on 15 October 2026 that pre-training of MAI-Base-2 is complete. It is a mixture-of-experts model with 80 billion active parameters and 2.8 trillion total parameters, trained on 16,000 Nvidia GB200 GPUs for 38 days in August and September 2026. Microsoft says it beat every open-weight model it compared on held-out tests and estimates 4.2 to 12.3 times the pre-training FLOPs efficiency of DeepSeek-V4-Pro-Base. The held-out sets were not published, and no independent FLOPs check was found. The model is in post-training. Microsoft says the pipeline will enable licensing weights to trusted partners through Foundry, with no date.