Press Review — AI, Agents and Cloud Infrastructure — June 24, 2026
This June 24 edition spans AI developer tooling (VS Code offline mode, open source under agentic pressure) to new cloud infrastructure layers (Lambda microVMs, OpenAI and Broadcom’s Jalapeño chip, analytics-converged Postgres for agents), alongside enterprise AI governance gaps and the widening environmental footprint of data centers.
Using Visual Studio Code’s ‘air-gapped’ AI model mode
VS Code 1.122 ships a BYOK feature letting developers plug in locally hosted models (Ollama, LM Studio) without signing into GitHub Copilot. This opens fully offline or air-gapped development workflows, a recurring need for teams under strict data-confidentiality requirements. Key limitation: local models don’t cover inline autocomplete or next-edit suggestions, which remain tied to Copilot. For developers with adequate hardware (8 GB VRAM or more, suited to models like Gemma4 or Qwen3.5 9B), it’s a credible path to keeping code data off external servers.
Open source grapples with agentic coding
Agentic coding is hitting the open source community from two angles: a flood of low-quality AI-generated pull requests overwhelming maintainers, and unresolved legal uncertainty around copyright for AI-produced code. David Heinemeier Hansson has pushed back against blanket bans, calling them unjustified discrimination, while some projects have gone the other way and banned AI submissions entirely to protect code quality. The legal angle is significant: copyright law requires human authorship, leaving AI-generated contributions in a persistent grey zone. How the open source ecosystem settles these questions will shape contribution norms for years.
UN Chief calls for more transparency on AI’s environmental impact
UN Secretary-General António Guterres launched an AI environmental transparency initiative at London Climate Action Week, urging the tech industry to shift toward renewable energy. A UN study published in June found that data center electricity consumption now exceeds that of 90% of countries worldwide, with fossil fuels still dominant in the mix. The tension is stark: Lex Coors, speaking for European data center operators, argued that maintaining competitiveness against the US and China requires leaning further into fossil fuels. The gap between climate commitments and infrastructure reality in the AI race is widening.
Article in French.
Windows 11: Optional update brings past restore feature and delayed updates
Microsoft shipped optional update KB5095093 for Windows 11 on June 23, adding two noteworthy features. The first is an extended system restore capability: the OS snapshots its full state every 24 hours (every 4 hours on Enterprise) and allows rollback up to 72 hours back, a meaningful improvement over the previous event-triggered approach. The second is indefinite update deferral, letting users postpone system updates on their own schedule. Both address longstanding requests from system administrators, with restore granularity varying depending on the Windows license tier in use.
Article in French.
Lambda boosted by MicroVMs closes in on AgentCore
AWS has introduced microVMs for Lambda, built on Firecracker and SnapStart, pushing execution limits from 15 minutes to up to 8 hours without event-driven handlers. They ship with shell access, Dockerfile-based filesystem setup, and five configurations scaling up to 8 GB RAM and 4 vCPUs (expandable fourfold). The target use case is multi-tenant environments (AI sandboxes, third-party code execution) where isolation per user is non-negotiable. The positioning alongside AgentCore signals that AWS is converging its serverless stack toward agentic AI workloads, with microVMs providing the compute isolation layer agents need.
Article in French.
OpenAI and Broadcom unveil LLM-optimized inference chip
OpenAI and Broadcom unveiled Jalapeño, a custom inference chip purpose-built for large language model workloads. The stated goals: better performance, improved energy efficiency, and scalable inference capacity. By developing its own silicon, OpenAI joins the club of Apple, Google (TPU), and Amazon (Trainium/Inferentia) in reducing dependency on NVIDIA GPUs and gaining control over inference costs at scale. If Jalapeño delivers on its promise, it could shift dynamics in the AI hardware supply chain and give OpenAI a structural cost advantage in serving its models.
EDB converges analytics on Postgres to support AI agents
EnterpriseDB has extended its managed EDB Postgres AI service with converged analytics capabilities, using Apache Iceberg as a shared catalog layer to bridge engines like ClickHouse and Spark. The differentiation from Databricks LTAP is clear: EDB starts from the operational layer (Postgres) rather than the analytics side, letting enterprises keep data sovereignty without proprietary cloud platforms or duplicated data copies. For CIOs managing regulated data or hybrid deployments, precisely the environments where AI agents need real-time access to operational data, this approach reduces both risk and architectural complexity.
The missing layer in enterprise agentic AI
AI agent frameworks from OpenAI, Anthropic and others are effective at orchestrating complex, multi-step tasks, but they lack a governance layer capable of evaluating each action against business or regulatory policies before execution. Gartner predicts that more than 40% of agentic AI projects will be abandoned by the end of 2027 due to inadequate risk controls. In production environments subject to compliance constraints, this gap is concrete: an agent can execute an action coherent with its goal but non-compliant with the environment’s rules. The missing piece isn’t in the LLMs or orchestrators, it sits in the execution layer that translates policies into operational guardrails.
AWS Continuum offers devs help with securing code
AWS launched Continuum, an AI-driven service for code security that goes beyond vulnerability detection. It continuously analyzes first-party and third-party code, surfaces threats, suggests fixes integrable into existing development workflows, and in enforce mode, can apply remediations autonomously once it has learned the environment’s context. New capabilities include automatic threat modeling from source code and design documents in STRIDE format, alongside penetration testing and code scanning inherited from its predecessor, Security Agent. The goal, per AWS VP Chet Kapoor, is to shift security teams from alert triage to full remediation lifecycle management.
Why open infrastructure will define the AI era
According to the Jellyfish 2026 State of Engineering Management Report, Claude Code has become the most widely used AI coding tool, ahead of Gemini Code Assist and GitHub Copilot, with 91% of developers reporting productivity gains. But this adoption is creating structural lock-in: Gartner forecasts that AI coding token spend will surpass developer salaries by 2028. Peter Farkas of Percona argues that open infrastructure (open source, open models, open protocols) is the necessary counterweight. Vendor portability is both a negotiating lever and a long-term resilience guarantee as AI tooling becomes load-bearing in software development.
