Microsoft's outage week ends with a 15.7% price hike in India
6 min read · 12 sources
- Microsoft's global outage hit Exchange Online, Teams, and Copilot from Thursday into Saturday, with servers failing automated health checks and initial recovery attempts failing.
- Microsoft will raise commercial cloud prices 15.7% in India and 10% in New Zealand on January 1, 2027, citing currency adjustments, while cutting prices in China, Australia, and Norway.
- Anthropic launched Claude Meetings in beta for Team and Enterprise, letting Claude join Google Meet calls to answer questions, build docs, and capture action items.
- AWS Secrets Manager added cross-account replication, enabling centralized secret ownership and rotation without shared KMS keys or cross-account policies.
- LMSYS formed the SGLang Consortium with Nvidia, AMD, Intel, Google, Meta, Cloudflare, Modal, and RadixArk, as inference is projected to be two-thirds of AI compute in 2026.
Microsoft spent Thursday through Saturday telling customers its services were coming back, then turned around and told Indian and New Zealand customers their bills are going up. The outage that took down Exchange Online, Teams, and Copilot was driven in part by servers failing automated health checks, and initial recovery attempts didn’t provide the expected relief. For anyone who consolidated their entire collaboration stack on one vendor, this week was the bill coming due.
The rest of the day is about what happens when you can’t undo what an AI agent did, Claude getting a seat at your meeting table, and a serious attempt to coordinate the open-source inference stack. Here’s what you need to know.
Inference is expected to account for roughly two-thirds of all AI compute in 2026, up from about half in 2025.
Microsoft's outage was a crash course in single-vendor risk
Microsoft says service is restored to all customers after the prolonged global outage that disrupted Exchange Online, Teams, Copilot, and connected services from Thursday into Saturday. Servers failing automated health checks contributed, and early recovery attempts didn’t help. The company apologized.
The technical root cause is thin in the reporting, but the operational lesson is not. If your email, chat, and AI assistant all share one control plane, one bad deployment takes out your entire communication layer. The analysis of the incident makes the point bluntly: consolidating email, collaboration, and productivity under one provider amplifies disruption, and it leaves you unable to switch providers when things fail. Multi-cloud resilience is no longer a nice-to-have for critical communication paths.
Microsoft raises cloud prices in India and New Zealand
Effective January 1, 2027, Microsoft will increase commercial cloud prices by 15.7% in India and 10% in New Zealand, citing currency adjustments. Prices will fall in China, Australia, and Norway, while some on-premises products in India face increases of up to 20%.
For engineers planning budgets, this is a regional repricing, not a global one. If you run workloads in India or New Zealand, lock in commitments before the new year. The China price cut is notable given the geopolitical context, but the practical takeaway is that cloud pricing is now a per-region negotiation, not a list price.
AI recovery is not an undo button
Commvault’s SHIFT 2026 conference pushed “AI recovery” hard, with CEO Sanjay Mirchandani framing it as “resilience operations” or ResOps. The key announcements include Recovery Studio for orchestrating full-stack recovery with built-in threat scanning, and AgentRecover now supporting Microsoft Foundry and Amazon Bedrock AgentCore. Identity recovery ARR nearly doubled year-over-year, and the company highlighted change auditing across Active Directory, Entra ID, and Okta.
The framing matters. Restoring an AI agent, reversing its changes to internal systems, and undoing its business consequences are three different problems. Commvault’s tools address parts of the first two, but actions such as sending customer emails or supplier orders still require controls before execution. Treat recovery as an assurance mechanism for AI agents, not just a backup tool.
Claude joins your meetings, and it's a visible participant
Anthropic launched Claude Meetings in beta for Team and Enterprise plans. Claude joins Google Meet calls as a visible participant, answers questions in writing, builds docs and whiteboards live, and writes notes and action items. All artifacts save to a meeting page on claude.ai. Integrations with Gong, Attention, Granola, Fireflies.ai, Otter, and Plaud are planned.
By default, Claude stays out of calls with outside guests unless an admin allows it, and it posts a notice in the chat but doesn’t reply there. If you’re running meetings with external clients, check your admin settings before this surprises anyone. The integration list is the real signal: Anthropic is positioning Claude as the note-taker and action-item generator that plugs into your existing transcript tools, not a replacement for them.
AWS Secrets Manager replicates secrets across accounts
AWS Secrets Manager now supports cross-account replication. You can maintain a synchronized copy of a primary secret in a secondary AWS account for access, backup, and disaster recovery. The replica is a full, independent resource encrypted with a customer-managed KMS key owned by the replica account, and it updates automatically on primary secret changes, including rotations. It’s available in all Regions where Secrets Manager exists.
This is a meaningful operational improvement. Central secret ownership and rotation no longer require shared KMS keys or cross-account policies. You can also promote a replica to a standalone secret. For multi-account architectures, this removes a whole class of IAM and KMS complexity.
SGLang gets a consortium with NVIDIA, AMD, Google, and Meta
LMSYS Org launched the SGLang Consortium with founding members NVIDIA, AMD, Intel, Google, Meta, Cloudflare, Modal, and RadixArk. SGLang is a full-stack inference engine in the PyTorch ecosystem that runs on hundreds of thousands of GPUs and powers production deployments for Meta, Microsoft, and Google. Inference is expected to account for roughly two-thirds of all AI compute in 2026, up from about half in 2025.
Members designate technical representatives and join two roadmap meetings a year, while maintainers keep control of technical decisions. This is a coordination body, not a governance takeover. If your inference stack depends on SGLang, this formalizes the roadmap process and gives you a channel to influence it.
Building always-on agents
Anthropic’s guide to always-on agents lays out the architecture for stateful, proactive agents that run indefinitely across sessions. The core primitives are an event inbox, timers, memory, permissions, and an outbox. The agent spends most of its time asleep and wakes for “runs” triggered by events.
The economics are shifting. Opus 5.5 costs 78% less per solved task than Opus 5 on a long-horizon coding benchmark. That’s the number that makes always-on agents practical: if waking up and acting is cheap enough, you can afford agents that monitor, decide, and act without direct supervision. The permission system is the part to get right before you let one loose on production systems.
Marvell attacks AI's memory and networking bottlenecks
Marvell is showing off optical interconnects, CXL memory expansion, and networking technology at Supercomputing 2026. The demos focus on scaling memory and compute across racks to address bandwidth constraints and infrastructure underutilization. If you’re building AI infrastructure, the bottleneck is no longer just the GPU - it’s getting data in and out of the box.
Cloudflare makes Private Origins generally available
Cloudflare’s Application Services for Private Origins is now GA. Public traffic can reach private-network applications without exposing IP addresses or opening inbound firewall ports. Private IP addresses are auto-detected in proxied DNS records, and existing WAF, bot management, rate limiting, caching, and Workers apply. It requires a Private Origins entitlement, Cloudflare One connectivity, and Unified Routing unless you use Cloudflare Tunnel, plus return routes for Cloudflare’s IPv4 (100.64.0.0/12) and IPv6 (2606:4700:cf1:5000::/64) ranges.
This changes the assumption that a public DNS name implies public reachability. You can keep applications on private IPs and still serve them through the edge with full security stack applied.
GitHub Copilot app gains Enterprise Server support
The GitHub Copilot app now supports GitHub Enterprise Server, generally available after an opt-in preview. Users can add a GHES account to work with remote repositories, issues, and pull requests in Copilot app 1.2.0 and later. Cloud sessions are not available for GHES, and you must sign in with a GitHub.com or GHE.com account first.
GitHub schedules Copilot model retirements for December
On December 14, 2026, GitHub will retire six Copilot models: Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6, and Kimi K3. Suggested alternatives are auto-enabled for Enterprise and Business customers unless an admin disables them. Claude Haiku 4.5 remains available to individual annual-plan subscribers. Update your workflows and integrations before the date, or you’ll find your pinned model suddenly 404ing.
You May Also Like
Microsoft taps Claude for Copilot in $2B deal as Anthropic agent hours outpace staff 6:1
Microsoft is committing over $2 billion annually to run Anthropic's Claude inside Copilot, pairing it with an internal dynamic router to trade off between …
Anthropic drops Haiku 5.5, Microsoft splits Copilot inference, and SAP adds USDC settlement
Anthropic launched Claude Haiku 5.5, cutting operating costs by 75% compared to Haiku 4.5 while adding dynamic effort controls and beating older mid-tier models …
Anthropic turns Claude into a coworker, and Yandex loses 2.84% of customer data for good
Anthropic is pushing Claude from code assistant into full workplace collaboration: new artifact types (Sheets, Whiteboards, Data) plus Claude Meetings, which …




