Currently0322026·06·234 MINread first
○Journal
Notes from inside the work.
Working notes from inside the studio — what we learn the week we learn it, written for operators who'd rather read it than be sold it.
○ All entries31 prior
- 0312026·06·234 MINThe frontier gets fencedThe US government ordered Anthropic to pull its two strongest models offline for all users — not because the models broke, but because the government can't figure out how to fence off the people it wants to exclude.On June 9, 2026, Anthropic launched Claude Fable 5 and Claude Mythos 5. Fable 5 scored 95 percent on SWE-bench Verified. Mythos 5 scored 95.5. By any available measure, these were the most capable coding models…
- 0302026·06·144 MINThe phone chooses the modelApple's WWDC announcement that iOS 27 will let users replace Siri's AI brain with Claude, Gemini, or ChatGPT is a platform story, not a features story.On the afternoon of June 9, inside Apple Park's Steve Jobs Theater, Craig Federighi gave the WWDC 2026 keynote and spent a surprisingly small fraction of it talking about hardware. The announcement that will matter most…
- 0292026·06·144 MINThe dev tool that brought its own keysOpenCode reached 160,000 GitHub stars and the #1 spot in developer tool rankings by being entirely model-agnostic — and its rise marks the moment the AI coding tool market split into two distinct categories.For the past two years, the dominant question in AI coding tools was which company would win. Cursor, built on frontier models with a polished interface, attracted developers who wanted speed. GitHub Copilot, backed by…
- 0282026·06·063 MINThe code runs fine until it doesn'tA cluster of security research from spring 2026 finds that AI-assisted coding has quietly accumulated a vulnerability backlog. The teams that adopted fastest may have the most to audit.The productivity story is real. Developers using AI coding tools spend less time on mechanical tasks, ship features faster, and — in McKinsey's February 2026 survey of 4,500 developers across 150 enterprises — reduce…
- 0272026·06·063 MINOpen weight is not open sourceMiniMax released M3 this week — frontier coding, one-million-token context, native multimodality. The weights are coming. The training code is not.On June 1, MiniMax published a model called M3 and described it as the first open-weight model to combine frontier-level coding, a one-million-token context window, and native multimodal capability in a single package.…
- 0262026·06·044 MINNear a trillionAnthropic filed a confidential S-1 on June 1. The company is worth nearly $1 trillion. The more interesting question is what that number asks of everyone who builds on it.The S-1 that Anthropic filed with the SEC on June 1, 2026, is a document nobody outside the company and the SEC has read yet. That is by design — "confidential" means the review happens before the prospectus goes…
- 0252026·06·043 MINMicrosoft builds its ownAt Build 2026, Microsoft released seven in-house AI models, including its first reasoning model. It is a hedge against the dependency that has shaped the company's AI strategy for five years.The relationship between Microsoft and OpenAI has never been what it appeared to be in the press releases. Microsoft invested — ultimately approaching $14 billion — and OpenAI built. In exchange, Microsoft got API…
- 0242026·05·304 MINWhen the exclusivity endsOn April 27, Microsoft and OpenAI unwound the terms that had bundled frontier AI access with a single cloud provider for five years. The cloud decision just got harder to dodge — and more honest.The announcement on April 27, 2026, from Microsoft and OpenAI didn't look like much from the outside. A joint blog post. A few revised terms. No new product released. If you were watching the ticker, nothing moved the…
- 0232026·05·303 MINThe audit firm is now the AI firmThree of the four biggest professional services firms now deploy Claude to their entire workforce. The story isn't the headcount — it's what the product becomes.When KPMG announced on May 19 that it was deploying Claude to all 276,000 of its employees across 138 countries, the number landed as it was designed to land: very large, unambiguously significant, useful for a press…
- 0222026·05·214 MINThe web becomes a toolGoogle announced WebMCP at I/O 2026 — a proposal to give AI agents structured access to any website. It is a bet on who controls the plumbing of the agentic web.There is a recurring gap in how AI agents interact with web services. An agent that needs to query a database, file a form, or complete a transaction has two options: it can use a well-designed API if one exists, or it…
- 0212026·05·214 MINThe steppe finds its capitalVenture funding in Central Asia reached $320 million in 2025. The money is small by global standards. The direction of travel is new.The Central Eurasia Venture Forum met in Tashkent in April 2026. Eight hundred people came: investors from London and Singapore, venture funds from the US and MENA, technology companies that had been watching the region…
- 0202026·05·204 MINThe management layer nobody builtIBM's AI divide data is right about the wrong thing: the gap between AI spending and AI value is not an operations problem, it's an accountability problem.IBM held its annual Think conference in Boston on May 5 and led with a statistic that should be uncomfortable for anyone who has approved an AI budget in the past two years: only 25% of AI initiatives deliver expected…
- 0192026·05·203 MINAgents get eyes in the browserFor the first time, AI coding agents can see what they build — not just write it.For most of the past year, coding agents have had a fundamental blind spot. They can read your files, edit your code, run your tests, and push to a branch. What they cannot do — or could not, until recently — is see…
- 0182026·05·194 MINThe pod and the paycheckMeta posted $56 billion in quarterly revenue on May 14 and started cutting eight thousand jobs on May 20. The math is not about the money.Meta posted $56 billion in quarterly revenue on May 14. The next week, it started cutting eight thousand people.
- 0172026·05·194 MINAfter mobile moneyAfrica's first fintech wave was payments. The second wave is about what happens when transaction data becomes collateral and credit reaches people who have never had a credit score.The first wave of African fintech had a legible story: mobile money reached people who had never had a bank account. M-Pesa gave Kenyans a way to send money from a feature phone. MTN MoMo spread across West Africa.…
- 0162026·05·163 MINTrained without NVIDIAOn February 11, Zhipu AI released a 744-billion-parameter frontier model built entirely on Huawei chips. The export controls were supposed to prevent this.The argument for semiconductor export controls rests on a compound assumption: that the best chips are American-designed and Taiwan-manufactured; that restricting their export imposes a sustained compute gap on Chinese…
- 0152026·05·164 MINThe stack nobody drewCursor, Claude Code, and Codex didn't merge. They stratified — and the shape of that arrangement is starting to matter.An engineer at a mid-size fintech company described her workflow in a thread earlier this spring: she opens Claude Code in her terminal, gives it the ticket and the relevant codebase context, watches it plan across…
- 0142026·05·153 MINThe workforce receiptMeta is cutting 8,000 people starting May 20. The same week, the company guides to $115 billion in AI infrastructure spending for 2026. The budget went somewhere.The announcement came in late April and landed with the precision of a fiscal quarter. Meta would cut approximately 8,000 employees — roughly 10 percent of its 78,000-person workforce — beginning May 20, with more…
- 0132026·05·153 MINThe correspondent bypassFasset raised $51 million to build stablecoin banking across 125 countries. For the businesses it serves, the alternative was a SWIFT wire, a three-day wait, and a fee that ate the margin.There is a type of business transaction that is unremarkable by outcome but exhausting by process. A supplier in Istanbul invoices a buyer in Almaty. The buyer's bank sends a SWIFT message to a correspondent bank, which…
- 0122026·05·144 MINWhat inference costs nowThe AI bill stopped being about access to models in early 2026. It became about what it costs to run them.There is a conversation happening inside engineering organizations right now that was not happening eighteen months ago. It is about the AI line on the cloud bill.
- 0112026·05·143 MINThe capex illusionMicrosoft announced $190 billion in AI infrastructure spending for 2026. Twenty-five billion of that is paying more for the same things.On April 29, Microsoft reported its third-quarter fiscal 2026 results and announced it would spend $190 billion on infrastructure this calendar year — up sixty-one percent from 2025, the largest capital expenditure…
- 0102026·05·124 MINThe cooling crisis has a name nowA withdrawn water filing in Utah and an unmetered tap in Georgia describe a constraint that has stopped being theoretical.The story of the next five years of AI infrastructure has stopped being told in GPU shipments. It is being told in water rights filings and substation queues.
- 0092026·05·124 MINRare earths from where we sitThe country has roughly two and a half percent of global reserves, no operating mines, and a Rare Metals Cooperation Center being built a short drive from our office.There is a particular type of meeting that has started happening in Ulaanbaatar this year. American officials, Korean executives, and Mongolian ministers sit around a long table and discuss minerals that none of them…
- 0082026·05·102 MINTwo stories from the same weekA breach you can name and a model you cannot use, taken together, describe the next phase of the security question.In one week, two stories arrived that, taken together, describe the security question every operator now has to answer.
- 0072026·05·092 MINThe grid has the last wordEvery AI capability story ends in a substation. The constraint binding the next decade of compute is not chips — it is power.Two years ago, the bottleneck on AI capability was getting hold of GPUs. A year ago, it was finding hyperscaler capacity to host them. This year, every story about AI infrastructure ends in the same place: a substation,…
- 0062026·05·092 MINOn running your own weightsA year ago, the question was whether open models could do the job. The question now is which one you keep on your own machines.A year ago, when an operator asked whether they could run their AI work on weights they owned, the honest answer was usually no. The open models lagged the frontier on the tasks that mattered, and the gap was wide…
- 0052026·05·082 MINTwo announcements, one questionAI infrastructure is splitting into a barbell. The question for an operator is which workloads belong at which end.In one week, the industry produced two announcements that, read together, describe the full shape of where AI infrastructure is going.
- 0042026·05·071 MINFrontier deals are not procurement signalsTwo outliers solving each other's unusual problem is not a standard for the rest of us to align with.This week, Anthropic announced a deal for the entire compute capacity of a SpaceX data center in Memphis — three hundred megawatts and over two hundred thousand GPUs, with a stated interest in extending to…
- 0032026·05·062 MINOn the data center in the broom closetWhen compute that used to live in a hyperscaler's hall starts fitting under a desk, the placement decision returns.For most of the last decade, the answer to "where does our compute live" has been a procurement decision. You picked a hyperscaler. You signed a contract. You got an account manager. The thing your data sat on was…
- 0022026·03·141 MINOn slownessWhy studios that care about the work produce less of it.There is a common wisdom, repeated mostly by people who have never shipped software inside an operation, that digital transformation projects fail because they are too slow. They say — you have to move fast. You have to…
- 0012026·02·021 MINWhat systems mean to usA system is what happens when people, forms, rules, and tools run together long enough that the output becomes predictable. Our job is to make that system legible.The word "systems" is in our name. This is not incidental. It is the argument of the studio.
○ Subscribe
The studio's working notes — sent when there's something worth reading.
If you came here from a search and you have an operation you can't read — start a project conversation with MonArch →