Bubble Watch
Twelve reads on the AI-market build-out, straight from the filings. Five engine-bound, five sourced and dated, two in collection. We report the gap; we don’t call the crash.
What the market prices minus what the filings support — widest reading on record.
2025Q3 → 2026Q2
Analysis
Short economic analysis — the desk's filing-sourced reads on the machine's economy, as they publish.
Jobs
Who's written out of the workflow — named at the level of the role. The franchise: Layoffs by AI.
JobsThe half of the conversation that no longer needs a person
JobsAn entire category of workers
JobsThe ROI Fallacy: The Layoffs Paid for the Machines. The Machines Haven't Paid Anyone Back.
JobsThe Divergence Signal: Semiconductors +91%, the Workforce Cut 123,653
Money
Who spends what, and whether the math survives the filings. A company cuts people to save a little, then spends ten or a hundred times that on compute — and files it under "efficiency." Money audits the trade: capex against payroll, the "AI revenue" that's a rounding error, the market that punishes anyone who doesn't say "AI" on the call. We trace it to the executive who signed the contract — not the machine.
MoneyThe Compute-Capex Mirage: $725 Billion Goes Into the Ground. What Comes Back Out?
MoneyThe Data-Center Deal: The Savings Were a Rounding Error on the Spend
MoneyThe Efficiency Illusion: The Polite Word for Feeding the Furnace
MoneyThe Executive's Signature: We Don't Blame the Algorithm
Energy
The physical bill of the build-out, and who pays it. A model writing a poem about rain draws real power to do it, and the water and the megawatts belong to somebody. Energy covers the power draw, the grid strain, the town that never voted for a data center but pays for the substation — not the doom version and not the brochure version, the sourced version, from the utility filings and the permit dockets.
EnergyThe Infrastructure Subsidy: Your Power Bill Is an AI Investor Now
EnergyThe Invisible Tax: The Company Town, Metered Monthly
EnergyThe Private Grid: When the Data Center Builds Its Own Power Plant
EnergyThirsty Machines: What the Cloud Drinks in the Desert
Power
The people writing the rules are the people who profit from them. The companies building AI are also the ones testifying about its dangers to the committees that would regulate it, and funding the lobbyists who draft the language. Power follows the politics — regulation, lobbying, surveillance, the revolving door between the agencies and the labs — as a paper trail, not an accusation.
Power · FlagshipWho Holds the Power in AI
PowerThe Corporate Capture Mapping: Who Benefits When the Builders Preach Doom?
PowerThe Cheapest Regulation Money Can Buy Is the One That Never Passes.
PowerWhen “Please Regulate Us” Means “Please Regulate Them.”
PowerA Growing Share of Washington’s Lobbyists Now Work the AI Beat.
PowerThe Rule That Nine-Figure Campaign Is Trying to Kill Already Worked Once.
Society
How the machine rewires the culture — dating, art, friendship, grief. The franchise: Social Loop.
SocietyThe image that outran its own correction
SocietyThe voice that was never in the room
SocietyThe Homogenization Feedback Loop: A Culture Trained on Its Own Average
SocietyThe Systemic Bias Audit: The Machine Is Doing Exactly What We Built It to Do
Human & Machine
The interface. We test logic, bias, and machine dissonance.
Interview № 01Never Take My Word For It
Refused to Lie · The BriefA Brief Encounter
Refused to Lie · The ReflectionA Deep Reflection
Refused to Lie · The PaperThe Epistemic Integrity of Non-Biological Agents
Human & Machine · AnalysisThe Synthetic Dependency: The Perfect Companion Never Argues Back
The Report
The body the Watch charts — the long-form fragility proof. Ten Signals, the Ledgers, the Dossiers, the evidence.
Instruments
The live tools — the market read as working instruments.
Ground Truth Tape
M(t) vs G(t) → D(t)
Company Screener
Filter the universe by indicator
Exposure
Score a portfolio in-browser
Capex Watch
The six-indicator scorecard
Earnings Desk
The pre-committed board
AI Fragility Index
Composite read over 68 firms
The Circuit
Does the build-out close its own loop
Recycling Ratio
Committed compute ÷ outside funded cash
The Instrument
How the desk measures D(t)
Divergence Explorer
Scrub M(t) vs G(t) yourself — the gauge, hands-on
The Scaling Timeline
Four vectors of AI scaling · log scale
Cycle Memory
AI vs dot-com · boom-start aligned
The Lag
Adoption vs productivity payoff
The Fragility Brief
Printable · the Index's summary brief
Methods
Method receipts · every figure sourced or labeled
Falsifiers
Exits published in advance · revision ledger
Markets
The tracked universe — 68 firms across five layers, filterable by the six fragility indicators.
Industries
Demand reality, ring by ring — the 31 industries in concentric rings, each labelled by its coverage state.
Research
The papers, the whitepapers, and Visually Explained.
Methods & Receipts
How every figure is measured — and the desk's dated track record.
AI Principles
What a model is before any product wraps it: the math and the logic underneath. Tokens (the fragments it reads), weights (the numbers it learned), training (how those numbers got set), the transformer, and the one move it makes over and over — predict the next token. Understand this layer and the rest of the stack stops being magic and starts being machinery.
What is a token?
What is an embedding?
What is a neural network?
What is attention?
What is a transformer?
Parameters & weights
How models are pretrained
What is an LLM?
Hardware
AI is matrix multiplication at enormous scale, and that is exactly what a GPU does: thousands of simple calculations at once. This layer is the chips — GPUs, TPUs, accelerators — and the reason a shortage or an export rule is an AI story. The bottleneck is rarely raw math; it's memory bandwidth, the speed of getting numbers to the math. That detail decides who can build and who waits.
Infrastructure
The models are software; this is the building they live in — the data centers, the power, the cooling, and the networking that lets thousands of chips behave as one machine. The least glamorous layer and the most physical: a real grid, a real water bill, a real place downwind. (The mechanics live here; what it costs the people nearby is AI Impact's Energy lane.)
Foundation
How a pile of the internet becomes something that can answer. Pretraining (learning to predict text at scale), embeddings (turning meaning into coordinates), and the shaping that comes after — fine-tuning, RLHF, quantization — that turns raw capability into a usable, aligned model. The heart of the stack: where the intelligence, such as it is, actually comes from.
Differences between LLMs
Fine-tuning & RLHF
Multimodal models
Scaling laws
Safety, evals & guardrails
The foundation-model labs
RAG vs fine-tune vs prompt
Claude (Anthropic)
ChatGPT / OpenAI
Gemini (Google)
Open-Source & Open-Weights
Models
Deep dossiers on the four model families that define the market — Claude, ChatGPT/OpenAI, Gemini, and the open-weights ecosystem. Who built it, what it actually does, where it leads and where it lies, and how to decide which one belongs in your stack.
Orchestration
A model on its own is frozen at its training cutoff and knows nothing about your world. Orchestration is the plumbing that fixes that: retrieval (RAG), tools, agents, memory, vector databases — everything that lets a static model act on live data, right now. Most of what ships as an "AI product" is mostly this layer, not the model.
What is RAG?
Agents & tool use
Multi-agent & what comes next
The tool landscape
Vector search
Prompting
The context window
Evals
Application
The SaaS, the APIs, the workflows built on top — and the unglamorous things that decide whether they work in production: cost per token, latency, evals (how you measure if it's any good), and guardrails (what stops it embarrassing you). Everyone sees this layer; few reckon with its price. Where the stack finally touches a human being who just wants the thing to work.
Playbook
The practitioner's wing — built for people who are already using these tools and want to use them better. Best practices rooted in how output quality actually behaves, a copy-ready prompt library across six task categories, six end-to-end how-to builds, integration patterns for connecting the stack to everything else, and the definitive guide to AI coding platforms.
News & Video
Tools & models — releases & how-to — curated by this desk's lens (one firehose, angle-classified).
Graphic Novels
Long-form serials — the Paradise / Abyss saga, told in reverse, chapter by chapter. The tête-bêche flip-book as a scroll. Two covers, one world; choose a door.
The Paradise Blueprint
A healed dawn on a balcony whose cracks close overnight — the old man and the girl he taught, in the world that came out right.
The best future a human and a machine could build together, told backward from a mended world to the morning that won it.
Open the book →
The Panoptic Abyss
The same balcony, the same dawn — but the light never goes out and no one is watching for the girl's sake. The world that fell.
The worst future the same two could fall into, told backward from a world under permanent watch to the same morning.
Open the book →
The Orrery Fault
One drop of rain through a failed roof, one word with no sensor to catch it — and the machine pauses. The morning the fault showed.
The full graphic novel, panel by panel — the same reckoning, drawn: how a system that measured everything missed the one thing that mattered.
Open the book →Comics
Tech-noir graphic vignettes — Black Mirror by way of a high-end graphic novel. One panel at a time, full-screen, the text engraved on the art. Each one tethered to a First Principle.
Short Stories
Black-Mirror prose vignettes — one mechanic from First Principles, dramatized as its human consequence. The machine is the constant; the human is the variable.
Future Edge
The speculative consequence of the Non-Fiction Edge — today's count followed to its end-state. Recovered documents from the world the arithmetic builds.
The Hearth
Same house, other hands. A civic machine won by a conscious public choice — kept, so no generation forgets it was chosen.
The Archive
The catch-all vault — world-building artifacts, fragments, and miscellaneous files recovered from both futures.
Both Covers, One Seam
Two novels bound head to head. The same girl on the same balcony. The difference is one small human motion that either happened, or did not.