6 min read · 11 sources

DeepSeek has closed an estimated $15 billion round from domestic industrial giants CATL and Tencent. Coming just days after the US barred civilian federal agencies from touching its systems, the cash injection puts real balance-sheet mass behind the Chinese lab’s open-weight model releases. It also sets up an aggressive pre-IPO infrastructure buildout aimed squarely at the American hyperscale moat.

The timing intersects with a distinct platform reshuffle in the West. As OpenAI breaks exclusive cloud distributor patterns to land inside Google Cloud, software vendors and infrastructure teams are actively re-architecting their underlying storage and runtimes to support always-on, high-concurrency autonomous agents.

AI infrastructure capex forecasts now span $1.21 trillion to $3.8 trillion for 2030, a split that will force planners into wildly different hardware strategies.

DeepSeek secures $15B domestic war chest to scale open-weight clusters

Source: techcrunch.com ↗

DeepSeek reportedly raised approximately $15 billion - around 100 billion yuan - in a syndicate round featuring battery manufacturer CATL and Tencent. The round follows its $7 billion raise in June 2025 at a $50 billion valuation and positions the company for an eventual domestic public listing.

The raise gives DeepSeek the capital runway required to absorb hardware export restrictions by purchasing domestic hardware clusters and stockpiling compute. For engineering teams evaluating model options, this capital infusion ensures that DeepSeek’s aggressive, low-cost API pricing and open-weight model cadence will continue challenging proprietary hosted models from Western labs.

OpenAI models and Codex agents land natively in Google Cloud

Source: openai.com ↗

In a major shift away from pure Microsoft Azure exclusivity, OpenAI announced a partnership with Google Cloud to bring its frontier models and Codex agentic workflows into Google Cloud’s Model Garden. The integration allows enterprises to procure OpenAI endpoints directly through the Google Cloud Marketplace, letting teams draw down existing GCP committed-use discounts and minimum-spend agreements.

Microsoft remains OpenAI’s primary infrastructure and supercomputing partner, but the commercial distribution barrier has broken. General availability is scheduled in the coming weeks, meaning enterprise architects can soon run cross-cloud stacks without configuring custom routing proxies or managing fragmented billing engines across Azure and GCP.

Google Cloud plans 960,000 NVIDIA Vera Rubin GPUs for agentic inference

Source: blogs.nvidia.com ↗

Targeting massive agentic inference pipelines and superlearner training, Google Cloud and NVIDIA announced the upcoming general availability of A5X instances. The instances run on NVIDIA Vera Rubin NVL72 rack-scale systems coupled with ConnectX-9 SuperNICs, Google Cloud’s Virgo Network fabric, and direct-to-chip liquid cooling. Google plans to deploy up to 960,000 Rubin GPUs across multiple facilities.

The Rubin architecture tightly couples Vera CPUs and Rubin GPUs across high-bandwidth NVLink interconnects, specifically designed to shorten the memory-bound bottlenecks typical in experience-driven reinforcement learning (RL) loops. Google Cloud also announced bare-metal rack-scale confidential computing for the hardware, allowing security teams to run isolated tenant workloads on shared hardware without paying standard virtualization penalties.

GitHub rebuilds Git storage using per-reference quorums for AI agents

Source: github.blog ↗

Autonomous coding agents generating automated refactors and pull requests are crushing legacy Git storage architectures. To survive the volume, GitHub detailed how it migrated 99.99999% of its Git repositories across on-premises data centers and Azure. Only the most active repositories remain temporarily pinned on-prem by a 25ms round-trip latency floor.

GitHub’s architectural fix decouples monolithic Git storage from compute by ditching repository-level locking. Instead, GitHub routes traffic regionally via its Spokes routing tier across five storage replicas, coordinating writes through per-reference quorums. By acquiring write locks only on specific ref pointers (like branch heads) rather than locking the entire repository namespace, internal benchmarks show a 3x to 4x throughput improvement under heavy agent write operations.

AI infrastructure spending projections hit a $3.8T wall of disagreement

Source: softwarestrategiesblog.com ↗

Financial and technology analysts are sharply dividing over long-term data center capacity forecasts. A comparative analysis shows that major analyst firms have upgraded their 2029 and 2030 capex projections over the course of 2026, but the absolute numbers diverge wildly. IDC sits at the conservative end with a $1.21 trillion projection for 2030, while Citi forecasts $3.8 trillion. Dell’Oro bumped its 2030 capex target past $3 trillion, and Gartner lifted its data center systems forecast by 58% to $2.16 trillion.

The variance reflects fundamentally different operating models. Conservative forecasts baseline their models strictly against known electric utility interconnect queues, whereas aggressive projections assume hyperscalers will deploy off-grid generation, small modular nuclear reactors, and microgrids. Systems planners sizing multi-year infrastructure commitments must treat these top-line forecasts cautiously, factoring in supply-chain pinches on transformers, chillers, and substation allocations rather than raw silicon availability.

Proactive agent architectures emerge: Claude Duet and ServiceNow Arc

Source: claude.com ↗

Agent runtimes are shifting from reactive text input boxes toward continuous, background execution environments with isolated operating privileges.

Anthropic launched a research preview of Claude Duet, a proactive autonomous agent app powered by Claude Opus 5.5 for Pro and Max users. Duet runs continuously inside an isolated cloud virtual machine provisioned with its own headless browser. The system monitors connected workspaces, surfaces morning briefs, and executes background work like dashboard prototyping. Anthropic isolates execution security by combining credential sandboxing from its Managed Agents stack with an automated classifier that prompts for explicit user approval when crossing permission thresholds.

Concurrently, ServiceNow and NVIDIA released the general availability of Arc, an enterprise desktop agent designed for macOS systems powered by NVIDIA RTX Spark local hardware. The architecture pairs ServiceNow’s Super Intelligence Control Tower for centralized policy definitions with the NVIDIA OpenShell runtime. Instead of handing cloud APIs broad system permissions, OpenShell enforces hardware-level sandboxing directly on the client machine to limit agent command execution, file manipulation, and enterprise application access.

Microsoft sets retirement date for native SMS and voice MFA in Entra ID

Source: neowin.net ↗

Telecom-based multi-factor authentication is reaching end-of-life on Microsoft identity stacks. Microsoft announced the retirement of native SMS and voice authentication for Entra ID, effective February 1, 2027, for general enterprise users. Global Administrators and external B2B guests get a short reprieve until July 1, 2027.

Enterprises determined to keep telephony fallbacks will have to contract and integrate a third-party telephony gateway via the Microsoft Security Store starting October 30, 2026. Microsoft’s guidance is straightforward: bypass the carrier aggregators entirely and migrate the tenant directory to FIDO2, WebAuthn, and platform passkeys to eliminate SIM-swap and SS7 interception vulnerabilities.

Infrastructure shifts: Crusoe in Paraguay, PHP2Shell honeypots, and Copilot defaults

Source: crusoe.ai ↗

Compute developers continue chasing generation-adjacent power assets outside traditional data center alleys. Crusoe partnered with Paraguayan energy firm Araico to build a computing facility pulling baseline power directly from the Itaipú Hydroelectric Dam. The site targets a mid-2027 deployment with 106 MW of IT capacity, scaling up to roughly 280 MW. Moving compute directly to South American hydro generation bypasses North American interconnect backlogs while serving both exportable cloud workloads and emerging regional demand.

In security monitoring, Amazon Threat Intelligence reported that PHP2Shell was exploited in the wild inside its MadPot honeypot fleet on October 17, 2026 - four full days before the coordinated public disclosure on October 21. Attackers leveraged automated diffing against embargoed patches, delivering functional exploits hours before public documentation dropped. Security teams must assume that embargoed vendor coordination windows no longer provide reliable lead time, requiring runtime detections rather than waiting for static CVE patch schedules.

Finally, platform administrators have a configuration audit to run. GitHub is enforcing a global default policy rollout for generally available Copilot features and Model Context Protocol (MCP) servers on Business and Enterprise tiers through October 26. This administrative update converts unconfigured feature toggles directly to the parent organization’s global policy state. Explicitly set policies and preview toggles remain unchanged, but platform engineers should review their AI Controls dashboard immediately to avoid unintended defaults propagating to developer seats.