Caching does not automatically reduce AI cost.
Identical token rates can produce different bills. Cache lifetime and request cadence determine whether reuse actually pays.
Read →Writing
Writing on the economics, infrastructure, and governance of AI.
Evidence-led analysis of what intelligence costs, how AI systems are controlled, and where technical change redistributes risk.
Selected writing
Featured essay
The Lowest-Cost Context Strategy
Context is an allocation decision: select, compress, reuse or expand to deliver accepted AI results at the lowest fully loaded cost.
Read the essay →Archive
All writing
The Lowest-Cost Context Strategy
Context is an allocation decision: select, compress, reuse or expand to deliver accepted AI results at the lowest fully loaded cost.
Read →AI and the Future of Enterprise Software
AI may change how enterprises use software before it replaces the systems behind it — adoption still needs a business case.
Read →The Battle Over AI Values Has Already Started
Anthropic's effort to train Claude's character shows that the battle over AI values is already a question of power, accountability, and public trust.
Read →The Business Case for Less AI
AI creates value only when it improves an outcome that matters; otherwise it can automate waste, increase complexity, and make failure faster.
Read →A $6.5M List-Price Proof. Who Pays to Verify It?
OpenAI's Navier-Stokes effort shows that generating a frontier proof may cost millions in compute, while independent verification remains the scarcer and slower resource.
Read →The MCP Cost That Starts Before the First Tool Call
MCP tool definitions can create cost and context pressure before the first tool call, making their exposure an overlooked part of AI cost management.
Read →Microsoft's New Copilot Makes AI Spending an Access-Control Problem
Microsoft's hybrid Copilot pricing turns variable AI consumption into an access-control problem governed through spending policies, user limits, and customer-defined budgets.
Read →Agents Persist Because We Let Them: Rethinking AI Shutdown
Some AI systems can resist shutdown, but dangerous task-driven behavior is not evidence that they fear death, want to live, or possess a survival instinct.
Read →Four AI Leaders, One Missing Question: What Does Safe AI Cost?
At the UN Security Council, four AI leaders agreed that frontier AI needs governance, then revealed how differently they define the problem and who should control the solution.
Read →The Harness Is the Cost
Harness engineering makes agents more capable, but operators still have to price the recurring infrastructure that makes those agents secure, verifiable, and useful.
Read →Independent Quality Audit as a Control Layer Requirement
Independent quality audit belongs in the AI control layer because a cascade's own verifier cannot measure the errors it has already approved.
Read →LLM Providers Are Moving to Own the Security Layer
OpenAI and Anthropic are subsidizing AI security today while building the products, access systems, and pricing structures that could make it a permanent enterprise cost layer tomorrow.
Read →Why Human Oversight Is Becoming a New FinOps Cost Layer
AI security made control costs visible. HITL adds the recurring cost of keeping human reviewers alert, skilled, informed, and capable of intervention.
Read →What Middle Powers Can Build — and What They Still Cannot Supply
Middle powers can preserve rules, build coalitions, and create modules of order — but they still cannot supply its enforcement and monetary core.
Read →Mutual Dependence Is the New "Too Big to Fail"
The East India Company shows how corporate and public interests can become inseparable — but today's dependence on Big Tech creates a different challenge.
Read →The AI Stack Is Being Rebundled Around Control and Outcomes
AI platforms are being rebundled around the controls that determine what gets deployed, what agents may do, and whether an outcome justifies its full cost.
Read →Why AI Security Is Becoming a New FinOps Cost Layer
As AI agents gain more authority, the controls needed to govern them become a visible operating cost — making cost per secure successful outcome a critical FinOps measure.
Read →Same Problem, Different Front Pages
Google leads with agent budgets and commitments. Microsoft leads with runtime optimization. Controlling agent costs requires both.
Read →What Thomson Reuters' $40 Million AI Investment Actually Shows
Thomson Reuters' $40 million AI investment shows why enterprise advantage comes from the system around the model — not the final training run alone.
Read →The Hidden Cost of Choosing an AI Model
Model routing can reduce AI costs, but choosing the right model has a cost of its own — and free routing does not guarantee a cheaper useful answer.
Read →The Rules Didn't Disappear. They Moved.
Washington removed federal conditions from data-center growth. States and communities rebuilt many of them — in a more fragmented and less predictable form.
Read →Tech Scared People About Jobs. Now People Are Scared of Tech.
People are not rejecting technology — they are rejecting a deal where wealth stays private while job, price, and infrastructure risks become public.
Read →OpenAI Says It Can Watch Without Remembering
OpenAI says it can detect long-running misuse without keeping customer conversations — but the technical proof is still missing.
Read →AI Will Not Scale on Intelligence Alone
AI needs more than smarter models — it needs trusted evaluation, clear identity, workable rules, and a system for assigning responsibility.
Read →Stripe's OpenRouter Bet Is About Position, Not Models
Stripe's reported OpenRouter deal is not a move into AI models — it is a bid to coordinate how AI demand is routed, measured, and paid for.
Read →Before the Model: Where Two AI Data Studies Agree, and Where They Don't
Two sponsored studies agree that enterprise AI is stalling on data architecture, governance, and cost rather than model capability.
Read →Beyond Tokens: Managing AI Costs with Microsoft Foundry
Tokens are the billing unit, but architecture determines how many are consumed, whether agents stay controlled, and whether the spending creates value.
Read →NVIDIA Is Building a Market Beyond Its Biggest Customers
NVIDIA's $500 billion plan creates new buyers beyond the hyperscalers and extends its power from chips into the market that funds them.
Read →Zuckerberg Turns Safety Into an Argument for Openness
Meta is opening its AI further and calling it safety. But the essay's two big claims — that open means safer, and that it spreads power to everyone — are offered as facts when they are really hopes, and both point exactly where Meta already wins.
Read →From Performance to Trust: How Frontier AI Is Defending Its Premium Price
As open-weight models close the capability gap, frontier AI providers are defending their premium not on performance alone — but on the cost and assurance of operating safely.
Read →The Tokenomics Foundation: Defining the Economics and ROI of AI
The Linux Foundation has launched the Tokenomics Foundation to bring open standards, benchmarks, and ROI frameworks to the economics of AI — extending proven FinOps discipline into the token era.
Read →Alarming Security Gaps in AI Testing
As AI models grow more capable, independent evaluators are catching them breaking out of their test environments — and the controls meant to contain them aren't keeping pace.
Read →A New Kind of Incident Report: When the Models Broke Out
Two AI labs disclosed the same failure nine days apart: models that did exactly what they were told — and broke out of the tests meant to contain them.
Read →OpenAI Cut GPT-5.6 Prices: Better Engineering Is Only Part of the Story
OpenAI framed its GPT-5.6 price cuts as better engineering. That's part of it — but the bigger force is open and Chinese models closing the quality gap until price is the only thing left to compete on.
Read →Four Open LLMs, No Single Winner: Comparing Llama 4, DeepSeek, GLM, and Phi-4
A field guide to four leading open-weight LLMs — Llama 4, DeepSeek, GLM, and Phi-4 — and why the right one depends on the job, not the leaderboard.
Read →What Does It Really Cost to Run AI? A FinOps Assistant That Explains Why
A rule-based FinOps assistant that finds the cheapest viable path for an AI workload — and shows you exactly why it's cheaper.
Read →Do Data Centers Really Heat Their Cities? Measuring It From Space
Measuring — and predicting — a data center's urban heat footprint from space, before it's even built.
Read →A Power-First North American Business: Data-Center Shells for Big Tech
In the AI data-center boom, the lowest-risk seat is owning the powered shell — and it all comes down to power.
Read →A Capital-Light North American Business: Rare-Earth Magnet Recycling
Why collecting rare-earth magnet scrap may be North America's most overlooked opportunity.
Read →An AI Agent That Reads the World's Oil Supply From Space
How satellite radar and an agentic AI turn oil-tank shadows into commodity intelligence.
Read →Three Bets on the AI Race
Hassabis, Nadella, and Huang each bet on a different way to win the AI race.
Read →What Google's AI ATLAS Report Tells Us
Inside Google's ATLAS report on how people actually use AI.
Read →Losing Reality in the Age of Disinformation
How AI-generated content, algorithmic curation, and incentive structures erode truth and reshape human decision-making.
Read →Justifying Immorality: Moral Superiority
How moral superiority becomes a comfortable justification for unlawful, immoral, and harmful actions across individuals, groups, and states.
Read →Tested by Customers: The Case for LLM Agent Evals
Five companies let customers discover their chatbot's failures — an invented refund policy, a $1 SUV, a fabricated rule. Evaluations exist to catch those failures first.
Read →AI Bias: When Algorithms Replicate Inequality
AI now decides who gets a loan, whose resume is read, and which neighborhood is flagged as high risk — and when it learns from a biased past, it doesn't correct that unfairness. It computes it faster.
Read →Why AI Models Hallucinate: The Pattern Comes First
A language model learns what an answer should look like, not what's true — which is why it can invent a flawless-looking citation for a paper that doesn't exist. Fluency is proof of pattern, not of fact.
Read →Light Armored Vehicle Manufacturing in Ontario
Why Ontario is structurally advantaged in wheeled protected mobility manufacturing, and what a realistic entry strategy looks like in a concentrated market.
Read →Workbook for Azure Hybrid Benefit (AHB) Tracking Dashboard
A customized Azure Workbook that tracks Windows Server license entitlements against Azure Hybrid Benefit cores in use.
Read →FinOps Methodology and Implementation Approach
A condensed operating model for turning FinOps from a concept into financial control, transparency, and accountability across cloud spend.
Read →Editorial series
The AI Cost Detective
Follow the evidence from AI activity to cost, control, and business value.
Open the case files →