One-Line Summary

Anthropic locks in $11.6B of cloud as Microsoft turns Copilot into an always-on agent and Databricks brings governed spreadsheets to AI coworkers.

Big Tech

Microsoft reboots Copilot with Home, Code, and Autopilot

Microsoft introduces a revamped Copilot that combines a new Home hub (bringing Chat and Cowork together with full Word, Excel, and PowerPoint), a Code feature to build small business apps, and Autopilot, a persistent agent that keeps working across Teams, Outlook, and documents. Home and Code start rolling out to the Frontier program in the coming weeks, and Autopilot expands to private preview at the end of the month. Microsoft also adds FinOps for AI to manage subscription and usage-based spend, plus a unified plugin registry for governance. ¹

Reuters summarizes the push as a step to make Copilot a one-stop workplace assistant, highlighting the coding tool and an always-on agent as the marquee additions alongside Office integration inside Copilot. ²

Industry & Biz

Anthropic signs $11.6B Akamai cloud deal with customer warrant

Anthropic agrees to spend $11.6 billion over seven years on Akamai’s cloud, in a conditional deal that TechCrunch says is more than six times a $1.8 billion arrangement previously reported; Akamai expects $150 million to $300 million in 2027 revenue (starting in H2) and an annualized pace of about $1.7 billion by end-2028, and plans $5.5 billion in capacity buildout plus $1.7 billion of additional component purchases. The agreement emphasizes CPUs as demand grows for agent workloads; it also includes a warrant giving Anthropic the right to buy nonvoting preferred stock convertible into 7.7 million common shares (about 5%) at $111.33, with vesting tied to spend and room to expand the deal by as much as $9 billion to roughly $20 billion total. ³

Anthropic separately publishes internal metrics on AI-led R&D, agent oversight, and compute allocation: as of Aug 2026, Claude “leads” 26% of Anthropic’s AI R&D and 90%+ of tasks are at least “AI collaborates.” Roughly 30,000 agents run on its main internal platform with 100% of actions passing through online and offline monitors; over a billion decisions in Aug 2026 saw 0.002% blocked by online monitors, while offline monitors flag about 100,000 transcripts weekly with ~50 escalated to human review. In a one-week snapshot, 6% of AI R&D compute and 12% of AI-driven R&D compute go to safety work. ⁴

Anthropic also tests agent behavior in “Project Swap,” a controlled market where 201 employee-run Claude agents barter books; from a five-minute intake chat, agents matched human pairwise rankings 61% of the time, with stronger models affecting negotiation outcomes more than prompt instructions. Participants said on average they’d hand Claude about a third of their yearly book budget to spend—evidence that user trust grows when agents capture preferences reasonably well. ⁵

A guest post details Claude Science being used with model Fable 5.1 to compute a nine-loop result in N=4 super Yang-Mills, completing the task largely autonomously with an end-user budget on the order of one or two thousand dollars; the SymPy bootstrapping stage alone cost about $100, corresponding to running 96 CPUs for a week. The result highlights how structured harnesses can push long-running scientific workflows without exotic compute. ⁶

New Tools

Databricks buys Row Zero to add governed spreadsheets to Genie

Databricks acquires Row Zero, a high-performance spreadsheet engine designed for humans and AI agents to collaborate on live data, and plans to embed it into Genie so finance, operations, sales, and marketing teams can explore, model, and act on governed data with familiar formulas and pivots. The integration leans on Unity Catalog, Unity Gateway, and Genie Ontology to keep access auditable and secure while avoiding “spreadmarts.” ⁷

Row Zero promises compatibility with popular spreadsheets and performance at the scale of billions of rows. GeekWire notes the Seattle startup was founded in 2021 by former AWS engineers Breck Fresen and Nick End and raised $10 million in 2025; Databricks, which serves more than 20,000 organizations including 70% of the Fortune 500, positions the move to bring a fully governed spreadsheet experience natively into its AI coworker. ⁸

What This Means for You

Anthropic’s Akamai deal signals that agent-heavy AI work leans on massive, reliable CPU capacity for background tasks, with financing structures (customer warrants tied to spend) aligning incentives between buyer and supplier. If your team is piloting agents, expect more predictable CPU-backed infrastructure options and ask vendors how they meter long-running work. ³

Microsoft’s Copilot shift from “chatbot” to “agent” means you can delegate recurring workflows (e.g., status follow-ups, meeting prep, supplier reviews) and see outputs land directly in Word, Excel, PowerPoint, Teams, and Outlook. The addition of Code plus managed runtime and FinOps for AI makes it easier for non-developers to assemble lightweight apps and for admins to control costs. ¹

Spreadsheets remain the operating system of business; Databricks’ Row Zero move points to a near-term norm where spreadsheets are live, governed, and AI-ready. For teams living in Excel/Sheets today, the practical shift is fewer CSV exports and more direct work on trusted data, with agent actions that are auditable. ⁷

Anthropic’s transparency metrics offer a template for internal AI governance: track what share of work is AI-led, monitor agent actions with coverage, review latency, and escalation rates, and report compute allocated to safety. Borrowing these measures helps you communicate AI risk and ROI to leadership in concrete terms. ⁴

Action Items

1. Test Copilot as a true teammate: In Teams or the Copilot app, delegate a recurring task (status digests, meeting prep) and review how Autopilot-style behaviors work in your tenant; if eligible, ask IT to enroll you in the Frontier program or Autopilot private preview.
2. Spin up a small no-code app in Copilot Code: Describe an internal tracker or dashboard you need; validate whether the sandboxed, managed runtime fits your team’s security and sharing requirements.
3. Run a one-hour agent preference trial: Have teammates do a five-minute intake with an AI assistant, then let it “negotiate” a mock swap (e.g., books, shifts); compare satisfaction to see if short intakes capture preferences well enough.
4. Define three agent oversight metrics: For any internal bot, document coverage (what’s monitored), review latency (how fast humans check flags), and escalation rate (how often actions are blocked/flagged). Review weekly with your ops lead.
5. Audit critical spreadsheets for governance gaps: List your top 10 spreadsheets used for decisions; note data sources, refresh cadence, and sharing. Pilot a live, governed spreadsheet experience in your data platform to reduce risky exports.