☕️ Google Cloud outage hits UK banks

• A power event in Google Cloud's europe-west2-b zone in the London region took down services for UK customers including Lloyds Bank, HSBC, Royal Mail, Revolut, and Cloudflare, with Google pledging full recovery within the hour.
• Downtime began at 0608 UTC when the availability zone faltered, knocking out BigQuery Data Transfer Service, Cloud Filestore, Cloud Run, App Engine, GKE, and Managed Service for Apache Kafka, an estimated 38 services in total.
• Google mitigated the power event and prioritized recovering existing workloads, though customers may still be unable to create new ones; one academic noted true resilience needs multi-region deployments and tested disaster recovery, not just moving to the cloud.

• Mac.apk, now in alpha, runs Android apps natively on Apple Silicon Macs with no emulator, virtual machine, or Linux kernel underneath, reimplementing the Android application environment directly on top of macOS.
• Apps launch as ordinary macOS processes, draw through the Mac's GPU, and run on the CPU at native speed, appearing in the Dock with resizable windows, a back button by the traffic lights, and light/dark mode that follows the system.
• The catch: there's no phone-style sandbox, so an APK runs with your full macOS user permissions and can read and write your files like any app, meaning you should only install APKs from sources you trust.

• Inception's Mercury 3, a diffusion-based language model, reportedly outperforms speed-optimized rivals GPT-6 Luna and Claude Haiku 5.5 on quality benchmarks while running four to eight times faster at over 1,000 tokens per second.
• Unlike sequential LLMs, the model refines multiple tokens in parallel and ships roughly 40x larger than Mercury 1 without losing speed, extending the context window to 1 million tokens, nearly quadrupling Mercury 2.5's capacity.
• Mercury 3 is compatible with the OpenAI API format for drop-in integration and is available via the Inception API, Baseten, OpenRouter, and Vercel, with Inception targeting agents, live translation, and tutoring where cumulative latency hurts usability.

• DeepSeek has open-sourced DeepInfer, the inference engine it runs in production, under Apache 2.0, letting developers serve its V4.1 Flash model on their own hardware with the same optimizations behind its API and support for future models planned.
• The framework bundles DeepSeek's FlashMLA, DeepGEMM and DeepEP kernels, multi-chip distribution, speculative decoding, async scheduling and structured outputs, and ships alongside deepseek-recipe, a Rust and Python library that reproduces its official API's request and response handling.
• DeepInfer supports Nvidia Blackwell and Huawei's Ascend 950 NPUs from day one but omits Hopper, currently handles offline inference only, and leaves KV cache management, load balancing and stage separation for developers to supply themselves.

• Google has launched TorchTPU, a specialized build of PyTorch that lets developers create and train AI models on its TPUs rather than being locked into Nvidia's GPUs, which run the majority of AI workloads today.
• The tool lets developers run existing PyTorch apps on TPUs with a few config tweaks and no code rewrites, and Google worked with Meta, Hugging Face, and the teams behind vLLM and SGLang to keep those libraries compatible.
• Google also shipped a TPU agent suite that migrates AI apps to Google Cloud and auto-fixes migration errors, and updated LiteRT so PyTorch-based apps can run on-device; AWS released a rival build, TorchNeuron, for its Trainium chips earlier this week.

• Cilium 1.20 shipped with an extensible eBPF datapath that lets cloud providers plug in custom networking without forking the project or merging changes upstream, alongside broader Gateway API support and progress on sidecarless mutual TLS.
• Plugins now run as separate processes, independently built and versioned, talking to Cilium over a custom resource and gRPC; the release also jumps Gateway API from 1.4 to 1.6, adding ExternalAuth, ListenerSets, CORS filtering, TCPRoute and UDPRoute.
• The plugin mechanism is beta and aimed at operators who know Cilium's internals, with docs warning a plugin can break Cilium or weaken security and may need updates as the architecture shifts; ordinary cluster users aren't the target.

📚 Tutorials & Advice

🔗 Other News & Articles

• Microsoft previews dotnetup: Toolchain manager for .NET SDKs and runtimes LINK

• TornadoVM 7 enables GPU-native workloads from Java applications LINK

• New in DevTools - November 2026 | Blog LINK

• Letting Rust support models bigger than 4GB in the browser LINK

• Xcode 27 replaced Simulator.​app. Expo CLI and expo-device-hub already know. LINK

• Your coding agent loses its cache while it waits for you LINK

• Apple stops pretending macOS is UNIX LINK

• Popular screenshot tool gets its biggest update ever, adds native screen recording LINK

📄 Papers & Resources

🧰 Latest Tools

💬 How did you find today's edition?

We read every reply — just reply to this email and let us know how we can improve!