Plumbing the
information age

A2A is transitioning to the Agentic AI Foundation as a standalone hosted project, consolidating it with MCP and other open infrastructure. This shift makes it easier for vendors and open-source projects to align on how agents discover each other, delegate tasks, and coordinate work across frameworks without brittle custom integrations. For ops practitioners, this means multi-agent orchestration patterns—where a network agent delegates to a security agent, which coordinates with a remediation agent—can now rely on standardized discovery and capability advertisement instead of hand-coded integrations. Founders and platform teams can now treat A2A plus MCP as a shared backbone for multi-agent ecosystems instead of inventing their own bespoke routing layer. This is significant because agentic NetOps deployments often require cross-domain coordination; standardized protocols reduce the integration surface and governance burden when agents need to hand off work.

The same intelligence can now be accessed via APIs and inside advertising workflows, replacing static target lists with live signals such as predicted buying stages and qualified account status inside the agent workspace. Go-to-market teams can stop manually exporting segments into their agents and instead let sales, marketing, and revenue ops agents act on up-to-date buying signals inside the tools they already use. This makes it more practical to trust agents with outreach, qualification, and campaign tuning, because they are grounded in the same data the analytics team relies on. While focused on go-to-market use cases, the pattern is worth noting for ops teams: MCP servers are becoming the standard way SaaS vendors expose real-time intelligence to agents. For infrastructure operations, this suggests vendors like cloud providers, network management platforms, and monitoring tools will follow similar patterns—publishing MCP servers that expose current state to agents without custom API integrations. This is a leading indicator of how ops agents will consume enterprise system state in production.

The AI for Network Leaders Summit takes place on August 19, 2026, at OASIS by Workville in New York City, kicking off at 8 AM ET. The event is positioned as a practitioner-focused gathering—not a sales pitch or product demo collection—centered on real deployments and open discussions from teams actively using AI to transform network operations. Attendees can participate in-person or via live stream. Sessions cover agentic AI, observability, configuration management, and ITOps topics. Selector positions this as an opportunity for network operations leaders to understand how AI agents (not dashboards) reason against models of network behavior and act on gaps between expected and actual performance. This aligns with broader 2026 industry shift toward agentic NetOps and away from reactive, rule-based automation.

Corma CEO Alon Pluda described a customer scenario where a security executive received a smartwatch notification from a Corma agent: 'I just caught a live attack. I need your permission to block it.' This pattern—autonomous detection with human authorization gates—addresses a core tension in agentic security: defenders need machine-speed execution without losing command and control. Rather than fully autonomous remediation (which carries governance and audit risks), Corma's model surfaces high-confidence detections to humans for approval, reducing response friction while preserving accountability. This approach aligns with how SASE and zero-trust platforms are evolving to incorporate agentic workflows—particularly relevant as security teams struggle with alert fatigue from AI-generated activity. The model also maps well to regulatory and compliance requirements that mandate human review of automated security actions, making it more operationally viable than purely autonomous response in enterprise environments.

AMD's Helios represents the first fully integrated single-vendor alternative to Nvidia's flagship rack-scale systems. The platform combines Epyc 9006-generation CPUs with new Instinct MI455X GPUs and proprietary Pensando networking silicon to create a turnkey solution that claims competitive or superior economics on compute density, memory bandwidth, and cost-per-token metrics. The disclosure came at the Ai4 2026 conference in Las Vegas. Beyond AMD's announcement, the broader industry narrative shifted materially in August: scarcity constraints moved from GPU supply to power-ready datacenter capacity. TSMC's Arizona fab expansion—adding multiple manufacturing lines and packaging capacity—targets not just GPU compute but the networking silicon and custom accelerators that form the AI cluster fabric. This signals that hyperscalers and vendors now view interconnect and fabric infrastructure as equal bottlenecks to raw GPU availability. For network ops teams, the implication is clear: 2026 buildouts require coordinating scale-up fabrics (intra-rack), scale-out fabrics (inter-rack), and cross-datacenter connectivity in parallel with GPU procurement.

Arista's Q2 FY26 earnings confirm the company's dominant position in Ethernet-based AI fabric adoption. The EtherLink platform—now deployed across 100+ hyperscaler customers—scaled from four to five early customers in 2024 to near-ubiquitous adoption in 2026 as Ethernet displaced proprietary interconnects for scale-out. The 7060XE7 switch generation, delivering 100 Tbps of capacity with 1.6 Tbps throughput per port and liquid cooling, represents the technical centerline of 2026 datacenter switching. Multi-year purchase commitments tripled in one year ($3.6B to $9.7B), a strong signal of sustained hyperscaler capex momentum and customer conviction in Ethernet-based fabric standardization. For network operators, this shift matters: Ethernet fabrics require different congestion control (QCN, FECN) and scheduling layers compared to InfiniBand, and operator skillsets are now moving toward technologies like DriveNets' scheduled fabric or Nvidia's MRC to handle the lossiness endemic to Ethernet at scale. Analyst sentiment is uniformly bullish (97% buy/strong buy), though execution risk remains concentrated in two or three mega-customers.

Attribute, founded in 2023 by Izhak Zimmermann and Liad Tropp, previously raised $13.5 million in seed funding. The startup's technology bridges a major operational blindspot: most organizations moving from AI pilots to production have limited visibility into actual AI spending. Traditional FinOps was built around VMs, storage, and networking—not token consumption. DoiT's acquisition is the fifth in 18 months, reflecting aggressive expansion into AI cost management. The timing aligns with the launch of the Tokenomics Foundation (jointly by Linux Foundation and FinOps Foundation), which aims to set operational standards for AI token attribution. Attribute's kernel-level instrumentation provides zero-instrumentation cost attribution, enabling DoiT to track spending at the level of teams, products, features, AI agents, and individual customers. DoiT data shows monthly AI spending is expected to triple over the next 12 months, making fine-grained cost visibility a critical operational requirement for managing AI infrastructure economics at scale.

Shieldstral represents a practical shift in how production safety workflows are built. Instead of fixed taxonomies and binary safe/unsafe labels, it accepts natural-language safety policies at inference time, enabling policy changes without model retraining. This is operationally significant: teams can update moderation rules via prompt updates rather than redeploying models. The model was trained on heterogeneous public safety datasets that traditionally disagree on taxonomies and annotation conventions—Mistral unified them into a single framework. A single 16GB GPU is sufficient for inference, avoiding the infrastructure scaling required by larger guardrail models. For AIOps and SRE teams deploying content moderation at scale, the efficiency and policy flexibility matter more than marginal accuracy gains over larger models. The release uses calibrated confidence scoring (yes/no probability), allowing practitioners to threshold by confidence level rather than accepting discrete labels, which is essential for production risk management.

OpenAI is heading for a stock market debut worth over $1 trillion even though it loses about $14 billion a year, while rival Anthropic just turned its first profit. This divergence in unit economics reveals fundamental strategic differences: OpenAI is betting on scale and market dominance through losses-funded growth; Anthropic is focused on profitability and efficient revenue per compute dollar. For enterprises evaluating long-term partnerships with foundation model providers, this matters. OpenAI's path depends on sustained capital markets access; Anthropic's profitability signals that Claude's inference efficiency and enterprise pricing are converging. The timing—IPO targeting September 2026—creates a potential inflection point for how much capital the frontier labs can raise, and whether public markets will reward or penalize massive infrastructure losses. Both approaches reflect different bets on whether the moat is compute scale or model efficiency.

The State Department is consolidating an AI technology bloc by explicitly urging countries to choose between US-led and Chinese-led AI ecosystems through the Pax Silica Declaration. This signals a shift from voluntary AI cooperation frameworks to explicit geopolitical alignment. For enterprises with global operations, the implications are significant: the US is consolidating an AI technology bloc similar to Cold War-era arrangements. Chinese AI models are simultaneously penetrating global developer ecosystems even as Washington attempts to restrict Chinese technology, creating operational complexity for multinational teams. Developer ecosystems—especially AI model routing platforms—are increasingly global, but government policy is pushing toward technology sovereignty. Enterprises should map which geopolitical bloc their vendor stack aligns with and plan for potential supply-chain constraints and export restrictions.