Plumbing the
information age

NVIDIA released the Multipath Routing and Congestion (MRC) protocol as an open specification through the Open Compute Project, proven first in production on Spectrum-X Ethernet hardware and now available for ecosystem adoption. MRC achieves high GPU utilization by load-balancing traffic across all available paths to ensure every GPU gets the bandwidth it needs during training runs. The protocol sustains high bandwidth even when network congestion occurs by dynamically avoiding overloaded paths in real time, and uses intelligent retransmission to recover from packet loss while minimizing long-running job interruptions. This open-source availability is significant for practitioners evaluating Spectrum-X deployments at scale, as MRC is a key component that enables the claimed 1.6x performance advantage over standard Ethernet. The open specification pathway allows vendors and operators to implement MRC independently while maintaining vendor-agnostic ecosystem compatibility—a departure from proprietary InfiniBand approaches.

Microsoft's move to add agent management to the Microsoft 365 admin center signals that enterprises are treating AI agents as first-class managed objects requiring the same governance, licensing, and lifecycle controls as traditional IT assets. Admins can now inventory agents deployed across multiple tenants they manage, reducing blind spots in agent sprawl. The consolidation in the admin center (rather than a separate portal) reflects a principle that agent governance should integrate into existing IT operational workflows rather than force admins to juggle another control plane. This matters for large enterprises and managed service providers because ungovernanced agents become compliance and security risks—agents accessing email, calendars, or confidential databases need the same audit trails and access controls as service accounts. Teams operating Microsoft 365 tenants should prepare for this tooling now; it reduces the friction for deploying managed agents into production and likely makes agent governance a new procurement and security requirement. The timeline aligns with broader 2026 trend of treating autonomous systems as requiring identity, inventory, and lifecycle management.

Verizon and Google Cloud formalized a strategic partnership leveraging Google Cloud's full-stack AI—Vertex AI, Gemini LLMs, and Gemini Enterprise—across three core domains: customer experience, network operations, and enterprise functionality. The contact center deployment, which began quietly in July 2024 and reached full scale by January 2025, now covers all 28,000 customer care representatives. Verizon reports a 95% 'comprehensive answerability' rate for customer inquiries and a near 40% increase in sales through service teams. Beyond contact centers, Verizon is deploying Google Cloud's data and AI solutions to build autonomous network intelligence—using Google Cloud as a data platform partner to predict and resolve network anomalies before they affect customers. The partnership also modernizes marketing automation through AI-driven content creation and campaign orchestration. From a network operations angle, this represents a significant shift toward closed-loop network intelligence, combining Verizon's nationwide connectivity infrastructure with Google Cloud's data pipeline and Gemini models. The deal includes advanced threat detection and proactive risk governance for security posture. While the announcement lacks committed spend figures, it demonstrates production-scale deployment of LLM-based agents across a major operator's OSS/BSS and customer-facing workflows—a tangible shift from pilots to operational reality.

For standard short-context use, GPT-5.6 Sol now costs $4 per 1 million input tokens and $20 per 1 million output tokens, down from $5 and $30 respectively. The lower API rates are live since August 21, 2026, with prices guaranteed through November 21. This reflects OpenAI facing growing pressure to cater to cost-sensitive customers and fend off competition from Chinese startups and other tech giants. The move signals a fundamental shift in how frontier model providers compete—moving from capability differentiation to infrastructure cost optimization. For enterprises running GPT-5.6 Sol in production, this pricing change directly impacts API spend and workload ROI calculations.

Alibaba sold 710 million shares at HK$112.70 each at a 3.6% discount to prior close, with institutional investor demand for nearly three times the offering amount. This represents Alibaba's first capital raise since its 2019 IPO, signaling aggressive commitment to AI infrastructure competition. The transaction reflects intensifying global race for compute capacity: the four major U.S. hyperscalers are expected to spend roughly $725 billion in capital expenditures in 2026, much of it on AI data centers, chips and infrastructure. For enterprises and investors tracking geopolitical AI dynamics, this signals China's determination to build sovereign AI infrastructure independent of U.S.-controlled supply chains.

Claude Fable 5 reached 2,726 optimization steps, Claude Opus 5 reached 2,920, and Moonshot's Kimi K3 reached 2,930—all compared to a 2,600-step human record. This first public large-scale evaluation of autonomous research capability found substantial differences between models in experiment selection, execution, and interpretation of noisy results. Unlike Anthropic's internal CPU-based evals or OpenAI's single-H100 nanoGPT runs, Prime Intellect's setup represents genuine frontier-model research tasks with compute and time constraints matching real scenarios. The study provides empirical evidence for which models excel at autonomous research—increasingly central to competitive AI development cycles and capability assessment.

Article 50 transparency duties apply to any organization applying its name and trade mark to AI systems that create content, including in-house chatbots. Violations carry maximum penalties of €15 million or 3% of global annual turnover. Enforcement sits with national market surveillance authorities rather than centrally with the EU AI Office. This represents a hard regulatory deadline for enterprises operating in or serving the EU: chatbot disclosure, AI-content marking, and deepfake labeling are now enforceable with immediate liability exposure. Companies without documented compliance programs and audit trails face risk. Digital Omnibus amendments delayed high-risk AI system obligations to December 2027 and August 2028, but Article 50 transparency enforcement was not postponed.