Writing

The practical AI FinOps metrics that connect token, GPU, retrieval and agent cost to accountable operational and business decisions.

A practical control map for deploying Model Context Protocol safely: identity, permissions, supply chain, approval gates, logging and operational ownership.

Strengths frameworks sell the upside and bury the cost. Every strength is a default, and defaults misfire under load — here's how to read your results for the failure modes, not the flattery.

AI governance fails when it is a policy without decision rights. A practical operating model for lifecycle gates, accountable roles, enforcement and time-bound exceptions.

A local model region answers one question about location. Sovereign AI also requires control of prompts, embeddings, model provenance, operations, jurisdiction and a tested exit path when conditions change.

The Digitalisation World interview covered the AI Grid concept in eight minutes. Here's what that format couldn't fit: the three questions I get asked most often afterwards, and why the answers matter more than the headline.

Traditional cloud bills scale with traffic. AI bills scale with autonomy — and an agent stuck in a loop spends like an intern with a corporate credit card. The circuit breakers, caching, and chargeback discipline that keep agentic spend answerable to somebody.

Training built the centralised AI factory. Inference is quietly un-building it — because serving predictions to users is a latency, geography, and cost problem the industry already solved once, for content, twenty-five years ago.

Most people use "quantisation" and "compression" interchangeably. They aren't the same thing — and knowing the difference is what separates a deliberate self-hosting strategy from cargo-culting whatever ran on someone's laptop last week.

Streaming solved access and then broke it again — eight services, $1,100 a year, and a discovery experience worse than cable. The platforms that survive won't be the ones that spent the most on content. They'll be the ones that removed the friction.

The repatriation conversation has become tribal — cloud believers vs data-centre revivalists. Both camps are wrong. The right question is workload-by-workload: where do the economics, performance, and compliance of this specific system land?

MCP standardises how LLM apps call tools, moving integration from M×N towards M+N. It does not make a connection safe — identity, scopes and review stay yours.

Authentication was the part the industry solved. Agents break everything downstream of it — what the caller intends, how far its credential reaches, and how much it can destroy before anyone notices.

A principle that cannot point to the moment it says no is a preference. What separates governance from a PDF: controls in the delivery path, a data layer you can actually see into, and gates priced by reversibility.

Why the build-vs-buy decision is never binary, how to think about total cost of ownership, and the hybrid approach that actually works in practice.

API inference wins until about 25,000 queries a day — where $12,000 a month of self-hosted capacity meets $0.016 a query. Most workloads end up hybrid.

A decision guide to agentic architecture patterns — single vs multi-agent, sequential, routing, parallelisation, orchestrator-workers, evaluator-optimiser, and reflection — with trade-offs, failure modes, and the observability you can't ship without.