Power, Principles, and Plumbing: The Vatican Takes on Big Tech While AI Infrastructure Gets Real

Today brought an unexpected voice to AI governance discussions as the Vatican issued its first papal encyclical on artificial intelligence, while the technical community wrestled with the unglamorous but critical infrastructure challenges that determine whether AI systems actually work in practice.

The Vatican's AI Power Challenge

Pope Leo XIV made headlines with his first encyclical "Magnifica Humanitas," but the 200-page document's real focus isn't artificial intelligence—it's about power concentration and inequality that AI amplifies. The Pope argues that technological power concentrated among elites creates dependencies, manipulates information, and undermines democracy, calling for an end to the "AI arms race" and community-based oversight.

The timing is significant, coinciding with President Trump's delayed AI executive order reportedly influenced by tech investor David Sacks. The encyclical emphasises that AI must serve humanity, not "the powerful few," with the Pope advocating for AI to be "disarmed" from military and economic interests while subjecting tech companies to stricter regulations. Notably, he presented the document alongside Anthropic co-founder Chris Olah, emphasising that moral principles must be embedded in AI development from the start.

For organisations adopting AI, this religious intervention highlights growing global concern about power concentration in AI development. The Pope's call for community oversight suggests that responsible AI deployment may increasingly require demonstrating how systems serve broader human flourishing, not just efficiency or profit maximisation.

Security Gaps Between Promise and Practice

Google Cloud's COO Francis de Souza advocates for proactive AI security built into platforms from the start, but Google itself struggles with AI security implementation. Recent reports show developers hit with surprise five-figure bills from unauthorised Gemini API calls, while compromised API keys remained active for up to 23 minutes after deletion. This exposes the gap between what cloud providers prescribe for AI security and their own practices.

The security challenge extends to everyday AI tools. A reviewer testing Amazon's Bee AI wearable found significant privacy concerns due to extensive data collection and cloud storage requirements, despite promising professional use cases like meeting transcription. The device requires broad mobile permissions including location, contacts, and calendar access, highlighting how AI convenience often comes with expansive data exposure.

These examples underscore a critical reality: even tech giants are navigating AI security in real time. For organisations deploying AI, this suggests the need for robust security auditing, careful vendor evaluation, and assumption that best practices are still evolving rapidly across the industry.

AI Infrastructure Gets Serious

As AI systems become more autonomous, infrastructure challenges are moving from theoretical to urgent. The Model Context Protocol (MCP) has rapidly evolved from Anthropic's internal experiment to an industry standard with 97 million monthly SDK downloads, creating new authentication challenges as AI agents perform tasks like updating CRMs and calling APIs.

WorkOS addressed this with auth.md, an open protocol enabling AI agents to autonomously register with services without human intervention. Unlike traditional API keys, it provides scoped, auditable credentials that can be selectively revoked—critical as agents perform sensitive tasks like code commits and system queries.

Meanwhile, observability is becoming essential for production AI systems. A comprehensive Langfuse tutorial demonstrates building complete monitoring pipelines for LLM applications, covering tracing, prompt management, and evaluation scoring. The Hugging Face team clarified confusing AI agent terminology, defining key distinctions like "scaffolding" versus "harness" to help practitioners communicate more effectively.

These infrastructure developments signal AI moving from prototype to production. Organisations need to invest in proper authentication, monitoring, and governance systems now, not after deployment problems emerge.

The True Cost of AI Hardware

New research from Epoch reveals that memory (HBM) costs have dramatically increased to nearly two-thirds of AI chip component expenses, rising from 52% to 63% between Q1 2024 and Q4 2025. Total AI chip component spending doubled from $22 billion to $52 billion during this period, with HBM memory alone accounting for roughly $20 billion of that increase.

This shift indicates that memory bandwidth has become the primary bottleneck and cost driver in AI chip manufacturing, potentially impacting the economics of AI model training and deployment. For organisations planning AI infrastructure investments, this suggests that memory costs—not just compute—will be the dominant expense factor, requiring careful consideration of model architectures that efficiently use available memory bandwidth.

Quick Hits