AI Business Brief

The AI industry has a Turing point problem. Not the Turing test — the chasm. The moment when the technology stops being something that early adopters evangelize and starts being something that the early majority quietly adopts, because it solves a real problem at a price they can accept and through an interface — as […]

Part one defined the constraint: AI is an intelligence factory governed by an exponential cost curve. Part two defines the response. OpenAI and Anthropic are not iterating within the same strategy; they are choosing different survival architectures. One maximizes scale to capture monopoly rents. The other maximizes defensibility to survive the cost curve. The divergence

Every major technology platform transition in history has followed the same structural pattern. A new capability emerges. Companies rush to build at the most visible layer — the user-facing application — because that is where attention flows and where early revenue is easiest to see. But over time, value doesn’t stay at the top. It

Three years ago, I introduced the idea of “AI Convergence”: the point at which AI becomes the missing layer that finally enables long-promised technologies to work. Many of those technologies stalled because they never crossed the chasm from early adopters to real-world utility. Augmented Reality — as explored in the interface layer wars reshaping consumer

On April 7, 2026, Anthropic published a 240-page system card for a model it isn’t releasing to the public. That single fact — a frontier lab documenting a capability leap in exhaustive detail while voluntarily withholding the product — is the signal worth paying attention to. The Mythos Preview system card is simultaneously a technical

If the previous piece showed where the power in AI is accumulating, this one shows what that power actually does when it is fully exercised. Google’s TPU flywheel made one thing clear: the frontier is no longer constrained by access to compute — as explored in the economics of AI compute infrastructure — . Closed-loop

Most people using AI are getting faster at being wrong. They’re producing more — more text, more slides, more analysis — at higher velocity, with higher confidence, toward conclusions that were never examined in the first place. The model is fluent. The thinking underneath it isn’t. And fluency, it turns out, is an excellent disguise

Here is the most important structural fact in AI infrastructure — as explored in the economics of AI compute infrastructure — right now: Google’s TPU fleet grew 11.5x in seven quarters, it now draws more power than Microsoft’s entire AI compute stack, and the quarterly rate of additions is itself accelerating. By Q4 2025, Google

Most practitioners interact with AI as if they are querying a search engine or instructing an employee. Both models are wrong in the same direction: they treat the AI as a passive executor of explicit commands. The AI Orchestrator Playbook establishes the correct model. A large language model is a conditional probability distribution — P(output

On April 7, 2026, Anthropic published a 240-page system card for a model it isn’t releasing to the public. That single fact — a frontier lab documenting a capability leap in exhaustive detail while voluntarily withholding the product — is the signal worth paying attention to. The Mythos Preview system card is simultaneously a technical

On April 2, 2026, Anthropic published an interpretability paper titled Emotion Concepts and Their Function in a Large Language Model — as explored in the intelligence factory race between AI labs — . The paper studied Claude Sonnet 4.5. Its findings are being discussed mostly in terms of “AI has emotions,” which is the least

The question driving most AI market commentary in 2026 is still the wrong one. Benchmark comparisons, model leaderboards, parameter counts — these measure capability. The competitive terrain that actually determines enterprise value capture is control infrastructure — as explored in the economics of AI compute infrastructure — : who owns the layers that make capability

Most people use AI the way they use a search engine. They ask a question. They get an answer. They move on. What they rarely do is build a system, a set of thinking tools so deeply embedded into how they work with AI that the output quality compounds over time rather than resetting with

A large language model — as explored in the intelligence factory race between AI labs — is not a search engine. It is not an employee you instruct. It is a conditional probability distribution — P(output | context) — and when you write a prompt, you are not sending a command. You are conditioning that

There is a number that keeps doubling, and nobody outside a handful of data centers fully appreciates what it means. In Q1 2024, the total tracked AI compute capacity across the world’s major technology players stood at roughly 2.5 million H100-equivalent units. By Q4 2025 — eight quarters later — it had reached 21.3 million.

The previous piece established the core shift: system prompting is not instruction-writing but systems intervention. A prompt does not compel a model to comply. It conditions a probabilistic system with its own attractors, feedback loops, and resistance dynamics. That reframing explains why surface-level prompt refinement so often fails. But it also surfaces a deeper bottleneck.

AI capability is not a single curve; it is four overlapping paradigms, each with distinct inputs, mechanisms, walls, and economic structures. Understanding which paradigm you are in determines whether you invest in data, alignment, inference compute — as explored in the economics of AI compute infrastructure — , or orchestration. The transitions between them are

Every analyst covering AR glasses is asking the wrong question. They are asking whether consumers will buy the product, whether the battery lasts, and whether the frames look good enough to wear to dinner. These are reasonable questions about a consumer device. The consumer device is the wrong unit of analysis. The right question is

ASML’s Q1 2026 landed with a guidance raise to €36–40 billion, memory customers sold out for 2026, logic ramping to 2 nm for AI products, and the unit plan moving from 60 Low NA EUV scanners this year to 80 next year. Read structurally, the quarter tells a single story: AI’s bottleneck has a postal code