Deep dive
One judgment, dug to the bottom: evidence, an adversarial table, and what would prove us wrong. For readers who'll spend twenty minutes to truly get one thing.
42 pieces
Some questions won't fit in a morning paragraph: a denominator nobody defines, two accounts that can't both be true, or a call of our own that has started to wobble. A deep dive takes one such question and digs to the bottom — evidence laid out, the strongest opposing case given its own table, and the ending written in advance: here is what would prove us wrong. Twenty minutes later what remains is not more information, but a position that dares to publish its own falsification terms. For readers willing to spend twenty minutes truly understanding one thing. How we verify it — the methodology
Horizontal AI spending mostly goes to assistants sold by office-suite makers and model companies
This report crosses: enterprise software strategy, corporate AI spending surveys, workplac…
Will an AI overbuild end in telecom-style bankruptcies? AI's debt hasn't disappeared, only moved
This report crosses: telecom history, corporate credit and bond defaults, lease accounting…
GPT-6 Astra needs to write out less of its reasoning, but OpenAI's assurance measures something else
AI safety teams catch misbehaving models mainly by reading the scratch work a model writes…
Before Anthropic's IPO: who pays for the compute, and who carries the loss if revenue lags
Anthropic is the company that develops the Claude family of AI models. Anthropic earns mon…
Is GPU compute tight? The answer is in contract length
CoreWeave, a cloud company that specializes in renting out GPUs, the chips AI runs on, say…
AI agents do their hands-on work on CPUs: four chipmakers brought dedicated racks, and buyers so far are buying CPU capacity
An AI agent is AI that takes a task and carries it out: the model works out the next step …
Long-document AI scores track how much it reads: new outer code lifted the same models 39 points on a diligence review
On September 8, Harvey, a legal AI company, published a set of experiments. Seven AI model…
The \"subsidy\" in AI coding subscriptions sits mostly on the compute ledger, not the loss ledger: allowances mostly move with compute
On September 29, OpenAI halved the allowance for new subscribers on OpenAI's US$200 monthl…
The White House AI accord checks whether controls work, not where: in three agent incidents this year, controls were not switched on in the test environment
On September 29, six technology companies, including Google, Anthropic and OpenAI, signed …
Moratoriums have delayed only about 2.3GW: the figure counts delay alone and scores cancelled and relocated projects as zero
About 300 local measures across the US, plus one New York State order, now pause or ban ne…
Data-center opposition reached the US election, and it lands on governors' executive orders. Most candidates want data centers to bring their own power, not a building ban
Opposition to new US data centers reached the election this year. On August 20 a Republica…
None of the three 'pace the frontier' numbers is a speed, and checking any of them takes someone inside the company
On September 12, Anthropic CEO Dario Amodei proposed to "pace the frontier": the few compa…
Artificial Analysis measured Anthropic's Fable 5.1 at 20% more per task than Fable 5, then 13% cheaper a week later on a new index version. List prices never changed. Our call: what really splits Anthropic and OpenAI is who ends up with the token savings, not who cuts prices
In the early hours of September 17 this publication recorded a judgment of its own that wa…
The claim that frontier labs are holding back their best models has six live and six expired examples in ninety days. Our call: the layer of the cut separates them, not how strict the control is. Nobody measures how long each hold lasts
On September 6 this publication recorded a call in its knowledge graph. The call concerned…
Newer Claude models invent fields on third-party coding tools, Armin Ronacher reports. The cause is traced and one switch fixes it. Our call: buyers should measure format failures in their own tool layer, not leaderboard gaps
In late August a Chinese-language relay went around the usual channels. The relay said tha…
OpenAI's Jerry Tworek and Box's Aaron Levie say open-source tools win because models use them best, a claim against their own side; our call is that the deciding moment may be execution, not training
But the deciding moment may lie in execution, not training. The mechanism assumes "will th…
OpenAI's US$1 billion cyber subsidy runs for six months, but the clock on its vetted cyber models is not one date. Our call: a July UK-US assessment found an open-weight model can already attack weakly defended systems of the kind the subsidy names, once given a foothold
On September 3, OpenAI announced Daybreak for Frontline Defenders: US$1 billion of subsidi…
OpenAI cites chain-of-thought monitoring to release GPT-6 Astra, but its own system card shows that monitor catching Astra under 11% of the time when Astra is told to evade, against near-100% for GPT-5.6 Sol
On September 1, OpenAI released GPT-6 Astra, the first model OpenAI has ever rated "Critic…
Three frontier labs now hand out cyber models by vetted list, not by price; our call: how long that gate matters is set by how fast open-weight models catch up, which the UK's AI Security Institute measures
On September 2 Google released Gemini 3.8 Flash Cyber, a model built only for security wor…
The \"50% off, 14x the tokens\" number is real, but our call is that it measures users switching onto OpenRouter, not new demand. OpenRouter's own earlier study of its platform found price sensitivity about one-twentieth of what 13.8x implies.
In late August a number went around the industry. OpenRouter is a routing platform: develo…
The reassurance that AI's gains mostly flow to users rests on a revocable decision, this publication's August 28 call said. The indicator it named cannot tell a revocation from soft demand
On August 28, SecondSource recorded a call in its ledger of calls. The daily brief publish…
SecondSource's \"buyer's rebate\" call on Google's Marvell warrant was backwards. Under US accounting rules it is a discount Marvell pays Google, about six points of Google's purchases by our calculation.
Yesterday's daily brief called Google's warrant in Marvell a buyer's rebate; this piece fi…
Your attention got scheduled last (August 19, 2026)
The market thinks output goes up if you let AI work several tasks at once, and that the co…
Attention goes to the back of the queue: in LinearB's 8.1 million pull requests, AI-assisted pull requests waited about five times longer than human-written ones for a first look, then reviewed faster
In LinearB's data on 8.1 million pull requests, AI-assisted ones waited about five times l…
The AMD CTO's \"next three years\" for optics now has a date: the 2027 MI500 gets an optical option, and copper and optics each hit a yield wall
In January, AMD's chief technology officer, Mark Papermaster, still answered the question …
\"AI broke into real systems\": August's wave was mostly disclosures, not new incidents — our call
In the past three weeks, five incidents in which an AI model, during safety testing, reach…
TSMC's C.C. Wei is \"really jealous\" of memory makers' margins but won't charge memory-style rent; our call is that memory is collecting it
Two things pointing in opposite directions happened on TSMC's July 16, 2026 earnings call.…
The US$720 billion question: the share of big-four capex turning into AI silicon is rising, not falling as this piece first assumed
This year's capital-spending guidance from Amazon, Microsoft, Alphabet and Meta adds up to…
The Last License War: The Clouds Never Paid — So What Is Moonshot Collecting?
Start by finishing the thread Kevin Xu left hanging. In an earlier era, he wrote, there wa…
The Software Split Is Real, but the Line Is Drawn Wrong: Six Months of Receipts on the Token Pass-Through Hypothesis
Start with the hypothesis in its own words. Gooaye's two scripts run like this. Scenario A…
The Guarantor Falls, the Guaranteed Rise: Circular Financing Is Now Priced in the Bond Market (July 29, 2026)
The deep dive "The Guarantor Falls, the Guaranteed Rise: Circular Financing Is Now Priced …
The Guarantor Falls, the Guaranteed Rise: Circular Financing Is Now Priced in the Bond Market
Start by laying out what happened on July 27. Before Monday's open, The Wall Street Journa…
Moonshot's Kimi K3 license splits open weights from free commercial use, and most of China's open-weight labs have not gated theirs
On July 27, the K3 license, not the benchmark scores, dominated the discussion in English-…
No One Gets to Be TSMC: Why Everyone in AI Is Becoming an Integrator
The foundry/IDM framework in this essay borrows heavily from Ben Thompson's long-running s…
"LLMs Are a Commodity": Take the TSMC Analogy Seriously, and It Bites Back
Start with the exact words. On the BG2 podcast in December 2025 — with Arvind Jain, CEO of…
Deep Dive #13 — Four Bull Cases for AI Capex Rest on One Untested Pillar, and the Pillar Just Got a Deadline
In the second half of 2025, four heavyweight figures each ran their own independent math t…
Deep Dive #12 — After Perfect Scores: The AI Race Is Now Measured in Hours, but the Cash Register Rings Somewhere Else
The first half of 2026 produced a pair of facts that should not coexist. On one side, the …
Deep Dive #11 — They Stopped Buying: Four Years of Chip Controls Built a China That Rejects American Chips
This is the story of a policy completing a full circle. In October 2022, Washington impose…
Selling Tokens Finally Makes Money — Now Comes the Fight to Keep It
The story starts with a reversal. From 2023 through 2025, nearly all the money in AI was m…
Deep Dive #9 — What AI Levels, the Market Reprices
Three years ago, the most-cited finding in all of AI-and-work research said this: give con…
Coupled in training, decoupled in deployment: our call on whether the AI harness can be split from the model
The question is whether the harness can be pulled apart from the model underneath it. The …
Our call on test-time compute, one week on: the capital case is borne out, but the \"budget\" has split into four axes
Test-time compute (TTC) is the compute a model spends at inference time, before it answers…
One more step: confirm by email
A confirmation email is on its way to the address just entered; clicking the link inside completes the subscription. Not in the inbox? Check spam or Promotions. Wrong address? Enter it again
Subscribe free — first issue lands tomorrow morning
Just an email address, unsubscribe anytime. Subscribing finishes with one click on the link in the confirmation email; if it is not in the inbox, check spam or Promotions.