Week 6 Intelligence Briefing: When Infrastructure Becomes the Product

by Pål Erik Waagbø | Week 6, 2026

THE 30-SECOND TAKE

This week, three things converged that haven’t overlapped before. Models are now building models (OpenAI’s Codex debugging its own training runs). Agents are managing agents (Anthropic’s Opus 4.6 with agent teams, McKinsey running 25,000 AI agents among 60,000 total “workforce”). Infrastructure became the actual product (Google committing £185 billion, Meta £600 billion through 2028, SpaceX-xAI merging at £1.25 trillion to build space-based data centres).

The measurement systems for this new layer are rudimentary at best. OpenAI’s selling £200,000 ad packages with basic click tracking. No one knows how to measure agent-assisted outcomes. That gap is your opportunity.

Why this matters for executives: The infrastructure race is already won by those with capital. Your competitive advantage isn’t compute anymore, it’s how you structure context so AI systems understand and transact through your organisation correctly.

Why this matters for digital practitioners: Traditional attribution is breaking. When agents discover, recommend, and transact, the entire vocabulary of “position 1” becomes meaningless. You need new frameworks, fast.

KEY DEVELOPMENTS

Models Building Models Is Now Production Reality

OpenAI released GPT-5.3-Codex this week. Here’s what matters: early versions of 5.3-Codex were used to find bugs in its own training runs, manage rollout, and analyse evaluation results. This isn’t aspirational anymore. It’s in production.

Anthropic’s Dario Amodei has said Claude helps design its successor. The threshold from theoretical to practical recursive improvement just got crossed.

The model achieved 64.7% on OSWorld (testing AI control of desktop computers), nearly double the prior version’s 38.2%. Hours later, Anthropic launched Claude Opus 4.6 with “agent teams” that let multiple agents split projects and work in parallel. This is the new baseline, not the future state.

Why this matters for executives: If your competitive advantage is “we’re good at configuring systems,” you’re about to get commoditised. Systems that configure themselves don’t need configuration expertise.

Why this matters for digital practitioners: Your expertise doesn’t disappear, but it shifts. Attribution modelling skills don’t map to old platforms. They map to building measurement frameworks for agent transactions.

The Infrastructure Race Locked In Hundreds of Billions

Google announced capital spending up to £185 billion in 2026. Meta committed £600 billion US infrastructure by 2028 through Meta Compute. Microsoft launched “Good Neighbor” data centre pledges responding to community backlash. SpaceX and xAI merged (£1.25 trillion valuation) with Elon Musk stating “in 36 months, the cheapest place to put AI will be space.”

This isn’t experimentation. This is infrastructure becoming the product itself. The companies winning aren’t the ones with better algorithms. They’re the ones who can deploy compute at scale.

Why this matters for executives: The AI race is fundamentally an infrastructure race now. If you’re betting on algorithm superiority, you’ve already lost. The question is whether your organisation exists in AI’s understanding of the world at all.

Why this matters for digital practitioners: Platform-dependent strategies (optimising for Google, Facebook, X) are actively being replaced. Publishers saw 33% Search traffic drops, 21% Discover drops in 2025. Expect 43% decline over three years. Your job isn’t platform optimisation anymore.

Agent Economics Arrived While You Weren’t Looking

McKinsey revealed 25,000 of their 60,000 “workforce” are now AI agents. OpenAI launched Frontier, a platform for enterprises to deploy and manage AI agents like employees (complete with onboarding, permissions, performance reviews). Universal Commerce Protocol (co-developed by Shopify and Google) provides standards for agent-to-merchant transactions.

The infrastructure for the agent economy is being built now. Not in the future. Now.

Why this matters for executives: When agents discover, recommend, and transact, who owns the measurement? Who gets credit? How do you attribute revenue when AI systems both find AND buy? These questions aren’t answered yet. First movers define the frameworks.

Why this matters for digital practitioners: Your CDP architecture experience maps directly. Context at scale is hard. You’ve spent years structuring personal context across systems. The substrate changed (AI instead of humans), but the expertise didn’t.

AEO Consensus Formed This Week

SparkToro research showed AI brand recommendations change almost every run. Rankings are meaningless. The new signal is visibility percentage across many runs, not position. Google Gemini averages 9 fan-out queries per prompt (creating roughly 900 queries per 100 prompts tracked). Sites must cover many related query variations.

Ahrefs data shows AI Overviews reduce clicks by 58%. Marketing Brew, publisher data, SparkToro all validate the same pattern. AEO is consensus, not speculation.

Why this matters for executives: Traditional SEO declining in parallel with platforms being replaced. You need infrastructure that ensures AI systems cite you correctly, not tactics to rank on dying platforms.

Why this matters for digital practitioners: The “Product Perception Loop” framework from TLDR Product nails it. “AI perception is now a product signal. PMs need to measure what AI believes about their product and deliberately improve it over time.” This isn’t content optimisation. This is measurement infrastructure.

THE BIG QUESTION

When AI models can improve themselves, agents can manage agents, and infrastructure becomes the actual product, what happens to the expert who built their value on knowing how to configure systems that no longer need configuration?

Here’s the uncomfortable answer. Traditional expertise is breaking in three specific ways this week revealed.

Attribution expertise? OpenAI’s ad platform measures clicks and impressions. That’s it. After years building sophisticated attribution models, the measurement systems for the new layer are rudimentary.

CDP architecture? When Gemini creates 900 queries per 100 prompts, traditional customer data platforms aren’t built for this interaction pattern. The journey isn’t linear touchpoints anymore. It’s population-level agent behaviour.

Platform optimisation? Publishers facing 33% Search traffic drops aren’t dealing with algorithm changes. They’re dealing with platforms being actively replaced.

But here’s where leverage sits. McKinsey has 25,000 agents working with zero attribution framework. OpenAI’s selling £200,000 ad packages with basic click tracking. Organisations have no framework for measuring AI-assisted outcomes.

Your attribution expertise doesn’t map to the old systems. It maps to building the measurement frameworks for agent transactions. When Universal Commerce Protocol enables agent-to-merchant transactions, who defines success? You could.

Context engineering is CDP architecture applied to AI. Google Personal Intelligence, the Apple-Gemini deal, SpaceX-xAI merger, all premised on the thesis that personal context is the competitive moat. The newsletters validate that context at scale is hard. Your CDP experience (structuring, connecting, activating personal context) maps directly.

AEO is confirmed, but the deeper opportunity is perception infrastructure. This isn’t content work. This is infrastructure that ensures AI systems understand, cite, and transact through accurate context. That’s architecture work.

Why this matters for executives: The frame shift required is this. Wrong frame: “I help companies optimise for AI systems.” Right frame: “I help companies exist in AI’s understanding of the world.” The first is tactical. The second is survival infrastructure.

Why this matters for digital practitioners: You know how to build attribution systems. You know how to architect CDP infrastructure. You understand cross-organisational data systems. The only thing stopping you from being the person who builds measurement and context infrastructure for the agent economy is calling it that instead of calling it “AEO consulting.”

YOUR MONDAY MOVE

Pick one thing your organisation does that AI should know about. Run the same query 20 times across ChatGPT, Claude, and Gemini. Track which answers appear and how often (visibility percentage, not ranking).

The exercise takes 45 minutes. You’ll discover two things. First, how wildly AI recommendations vary run-to-run (SparkToro was right). Second, where AI’s mental model of your organisation is wrong or incomplete.

That gap between what AI believes and what’s actually true is your perception infrastructure opportunity. Most organisations don’t even know the gap exists yet.

Why this matters for executives: When SpaceX-xAI is betting on space-based compute, Google and Meta spending hundreds of billions, McKinsey running 25,000 agents, these aren’t experiments. They’re the new operating environment. You need to know if AI understands your organisation correctly before agents start transacting at scale.

Why this matters for digital practitioners: This 45-minute exercise becomes your audit methodology. You’re not learning a new field. You’re applying visibility tracking to a new transaction layer. The frameworks you build now define the category.

QUOTABLE TAKE

“We’re seeing the hotels.com moment of the whole internet. Your website just dropped from being the paper to becoming a reference in the AI’s paper. You don’t just want to get referenced. You want to be the main paper the AI’s rewrite is built on.”