Plumbing the
information age

Temporal's comprehensive survey (conducted April 29–May 25, 2026) documents a seismic shift in AI agent adoption within engineering organizations. The 70.8% relative increase in daily agent usage from 47.3% to 80.8% signals that agents have transitioned from experimental to production workflows. However, the report surfaces persistent pain points: state management, debugging visibility, and reliable execution remain the top blockers for scaling agent deployments. Critically, the data confirms that AI agents have moved from side projects to core engineering tools, raising expectations around reliability, observability, and governance. Teams treating agents as optional risk falling behind competitors systematically redesigning workflows around agentic execution. Temporal's latest releases directly address this gap: Serverless Workers for Lambda (public preview) eliminates infrastructure management while providing built-in observability; the Google GenAI integration (public preview) wraps Gemini model and tool calls in Activities without code changes; Deep Agents integration (pre-release) enables LangChain agents to run inside deterministic, replayable Workflows. For SREs and platform engineers, this report establishes that observability infrastructure must evolve from APM-centric (designed for deterministic services) to agent-aware telemetry that captures reasoning chains, tool calls, state transitions, and memory operations.

MCP recently surpassed 400M monthly SDK downloads, a 4x increase this year, and has become the industry standard for connecting AI agents to applications. The stateless architecture represents a fundamental shift for production deployments. Maintainers are now talking about MCP not as a tool-calling API, but as a full application protocol for agents. Standardized extensions include Tasks officially graduating from experimental status. The previous MCP spec left too much ambiguity in how auth was supposed to work; the 2026-07-28 spec closes those gaps, making it easier to get enterprise IT approval for Claude Agent SDK deployments that touch internal systems. Private network tunnels are another enterprise-facing addition that let MCP servers communicate with Claude agents without exposing internal services to the public internet, removing a genuine blocker for organizations with strict network security requirements.

Data Center Knowledge's analysis reveals that modal rack density in 2026 reached 11 kW (up from 9 kW in 2025), but AI training infrastructure operates in a completely different power class. Nvidia's GB300 NVL72 platform requires up to 142 kW per rack according to the NVL72 AI Factory reference architecture, with the newer Vera Rubin NVL72 entering full production in June 2026. Liquid cooling via cold plates on chips now dominates at 55% market share as of 2026, handling 100-150 kW per rack. Two-phase immersion cooling faced setbacks due to PFAS restrictions on coolant supply, though a replacement fluid was qualified in early 2026. Microfluidic cooling remains roadmap-level technology. The practical constraint is no longer compute density but electrical distribution design, cooling infrastructure, and single-point-of-failure risk. As hyperscalers push toward 200 kW per rack, transformer lead times have stretched from 140 weeks in 2023 to over 160 weeks in 2026, creating a bottleneck that rivals chip availability for infrastructure deployment speed.

Stripe's payment platform processes traffic that has grown to represent 1.6% of global GDP, creating unprecedented strain on the service mesh layer. The company's infrastructure team evaluated and ultimately replaced Envoy with an alternative architecture designed specifically for their scale and reliability requirements. This post walks through the technical reasoning behind the architectural shift, the constraints that drove the decision, and how the new proxy handles the company's traffic patterns at scale. For practitioners managing high-throughput payment or financial infrastructure, this represents a rare detailed account of when and why to move beyond popular off-the-shelf solutions like Envoy. The shift reflects practical lessons about service mesh limitations at scale that apply to any organization processing mission-critical transactions at global scale where every millisecond and every failed payment has direct business impact.

OpenTelemetry's success as a unified instrumentation standard has created an unintended consequence: teams can now collect telemetry at scale but face a steep cost wall when trying to store, retain, and query it cost-effectively. The article examines how this storage gap creates blind spots in observability despite OTel's collection capabilities. As AI systems generate exponentially more logs, traces, and metrics than traditional applications, this gap becomes critical. For AIOps and SRE teams, the implication is stark: collecting more data doesn't improve MTTR if you can't afford to retain and search it. The piece argues that the next battle in observability will be fought at the data layer—not the collection layer—requiring teams to rethink retention policies, sampling strategies, and storage architecture alongside their instrumentation choices.

Stripe's acquisition of OpenRouter for over $7 billion represents a significant consolidation play in the model-routing infrastructure layer. OpenRouter operates an API gateway that abstracts requests across approximately 400 AI models, having raised at a $1.3B valuation only months prior. For practitioners, this signals that the model-access layer itself is being consolidated—reducing fragmentation but increasing dependency on a single routing layer now owned by a payments infrastructure company.

August 2026 became the busiest month on record for model releases, with 14 new versions shipped across multiple providers. Google released Gemini 3.7 Flash on August 13, and Z.ai shipped GLM-5.3 Flash on August 26. Simultaneously, deprecations accelerated: Anthropic retired Claude Opus 4.1 on August 5, and OpenAI's Assistants API removal on August 26 forces teams still on that API to migrate to Responses API or rebuild equivalent functionality. Model lifecycles are now compressing to 6–12 months; teams must track deprecation calendars across OpenAI, Anthropic, and Google separately to avoid production breakage.

Verizon plans to integrate scattered customer, network, and operations data into a unified Agentic Data Cloud platform, with Gemini systems already handling the majority of inbound consumer calls and chats in certain channels. For network operations, the partnership focuses on autonomous agents that gain programmatic access to network systems via APIs, enabling automated responses without manual intervention—addressing the core operational challenge of scaling autonomous network functions safely. Verizon will also use AI agents as personal research assistants for human agents, surfacing relevant information during customer interactions, while internally Gemini-powered agents coordinate workflows and analyze network incidents. For practitioners, this represents a shift from reactive, rule-based network operations toward intent-based systems where agents detect and remediate anomalies proactively. The practical value lies in Verizon's unified data foundation—breaking organizational silos that typically fragment customer, network, and operational visibility.

Anthropic's lawsuit in federal court in California alleged that Defense Secretary Pete Hegseth overstepped his authority when he designated Anthropic a national security supply chain risk. The company was labeled a national security risk after it refused to allow its Claude chatbot to be used for domestic surveillance or autonomous weapons. Judge Lin's order rebuked the Pentagon, stating: 'The empty invocation of national security is not a blank check to punish and retaliate against government critics.' Anthropic has a second lawsuit pending in Washington, D.C., over a separate Pentagon supply-chain risk designation that could lead to its exclusion from civilian government contracts. This outcome signals a potential constraint on executive power to weaponize national security designations against AI labs based on their policy positions rather than genuine security threats, affecting how frontier labs can maintain safety boundaries without regulatory retaliation.

Jalapeño is the first step in a multi-generation compute platform designed for initial deployment by the end of 2026, combining OpenAI-designed accelerators with Broadcom silicon implementation, networking, and connectivity technologies. Tested on SemiAnalysis' InferenceX benchmark, Jalapeño registered more tokens per user and more throughput per kilowatt than currently available state-of-the-art inference processors. At 700W per chip versus 1,400W for the GB300, Jalapeño delivers roughly twice as many chips per gigawatt of installed power. Deployment begins at the end of 2026 in small volumes, with more significant deployment coming in 2027. For infrastructure practitioners, this signals that frontier labs are executing full-stack optimization (silicon, memory architecture, networking) to solve inference economics at scale—a capability that will compress margins on generic compute and shift competitive advantage toward vertically integrated platforms.

Emerald's core product, Emerald Conductor, is software that schedules AI workloads against available batteries and generation onsite, reducing power draw when the local system is stressed while critical jobs keep running, effectively giving utilities a site they can dispatch like a generator. The International Energy Agency expects AI to drive nearly half of U.S. electricity demand growth through 2030. Twelve Fortune Global 500 companies now hold investments in Emerald AI. The company is working with Digital Realty and NVIDIA to bring online the world's first power-flexible AI factory—the nearly 100-megawatt Vera Rubin AI Research Factory in Manassas, Virginia. This round underscores that power and grid integration are becoming first-class infrastructure concerns for enterprise AI deployment—not optional optimizations.

The financing programme extended credit to AI cloud companies buying Nvidia chips and took a cut of what those chips later earn, with Nvidia receiving payment at sale and again as a share of cloud revenue. Under the arrangement, Nvidia promised to rent GPU capacity itself if a provider could not find another customer and would receive 50% of any revenue above a base hourly rate covering provider costs. Investor scrutiny has mounted as Nvidia pumps money back into the AI ecosystem, stoking fears over circular deals that artificially inflate demand. The company helped arrange $500 billion in financing from major US financial institutions for customers and agreed to guarantee up to $105 billion to help OpenAI lease a massive data center. The pause highlights regulatory and reputational risks when a dominant chipmaker controls both supply and demand-side financing, a dynamic that will likely constrain similar initiatives across the AI infrastructure ecosystem.