🎉 Morning. The Financial Stability Board just sent the G20 a very expensive-sounding reminder: AI risk is not only a model-lab problem anymore.

It is a bank problem, a vendor problem, a cyber problem, and, worst of all, a backup-plan problem.

Quick hits before we get into it: the FSB says frontier AI could change the speed, scale, and economics of cyber risk. Big Tech's AI profit story has a lot of investment gains baked in. The UK wants to buy from homegrown AI startups before every public-service workflow becomes somebody else's platform. Meta's smart glasses are running into the oldest hardware problem in the world: other people in the room. And NASA just launched a $4.3B telescope with a hand-me-down spy mirror, which is objectively great procurement energy.

Let's ride. 🤠

The Financial Stability Board published a letter to G20 finance ministers and central bank governors ahead of this week's meeting in Asheville. The headline: frontier AI models are becoming autonomous enough to matter for financial stability, especially through cyber risk.

The cleaner version came in the FSB's own press note: Andrew Bailey warned that frontier AI could affect cyber risk and called for safer model release, better deployment controls, stronger response and recovery plans, and resilience among critical third-party technology providers.

That last phrase is the one to underline. The scary part is not one bank using one weird tool. The scary part is many firms leaning on the same models, the same cloud providers, the same data vendors, the same identity systems, and the same automated support workflows. A failure does not need to be cinematic to be systemic. It only needs to be shared.

The Financial Times framed the concern around cross-firm cyber disruption and backup infrastructure. The Guardian's read landed in the same place: the AI optimism trade and the AI risk file are now sitting in the same G20 folder.

The operator read is boring, which makes it useful. If your company is wiring AI into customer support, fraud checks, treasury, pricing, compliance, or incident response, the new question is not just "what can the model do?" It is "what happens when the model, vendor, workflow, or identity layer goes sideways at the same time as everyone else's?"

That means the AI budget has a shadow budget: logs, kill switches, vendor maps, recovery drills, human override paths, data-retention rules, third-party concentration reviews, and the kind of disaster-recovery plan nobody likes until the day it saves the quarter.

The market has spent two years asking who gets the upside of AI. The G20 version asks who has to eat the downtime.

The Financial Times reports that the largest U.S. tech companies have received a roughly $160B pre-tax profit boost from gains on stakes in AI companies.

That does not mean the AI cycle is fake. It means the scoreboard is getting harder to read. Operating profit, investment marks, vendor spend, cloud commitments, and strategic stakes are starting to blur into one giant ecosystem ledger.

Why it matters: when AI investments make AI earnings look stronger, the market can mistake circular momentum for clean demand. Great businesses can still have messy receipts.

The FT says the UK is opening a £100M procurement push for homegrown AI startups, with public-service contracts ranging from £250,000 to £10M.

The stated target is healthcare, cybersecurity, defense data, public compute, and AI risk. The strategic target is obvious: do not let every important government workflow become a renewal negotiation with a foreign platform.

Why it matters: procurement is industrial policy with invoices. The startups that win boring government contracts can become tomorrow's default infrastructure.

Axios reports that major AI players are leaning into healthcare and disease work as the image-repair story for AI's next phase.

That makes sense. "AI helps cure disease" is a much better public narrative than "AI eats entry-level work" or "AI makes phishing cheaper." It is also not empty. Bio, diagnosis, drug discovery, and hospital workflow are real use cases.

Why it matters: watch the gap between demo value and deployment value. Healthcare does not reward vibes. It rewards evidence, liability control, integration, and enough trust that people actually use the tool.

Business Insider has a smart piece on Meta's camera-enabled glasses and the privacy problem that follows them into gyms, bars, events, and workplaces.

The product question is not only whether the hardware works. It is whether everyone around the wearer believes the recording light, understands the signal, and accepts the social contract. That is a much harder launch requirement than battery life.

Why it matters: wearables fail in public before they fail in spreadsheets. If the room says no, the roadmap has a problem.

AP says NASA launched the $4.3B Nancy Grace Roman Space Telescope on August 30. It is headed toward an observation point about 1M miles away and has a field of view more than 100 times wider than Hubble's.

AP also reports Nissan and Honda are collaborating on next-generation car software and electronic control units for vehicles expected around fiscal 2029. The automaker fight keeps moving from horsepower to operating systems.

MarketWatch argues AI has not eaten software after all, with the sector's rally potentially having room to run. The useful question is not whether SaaS dies. It is which SaaS tools become the AI control surface.

AP's box-office note says Coyote vs. Acme opened with about $15.5M after being shelved as a tax write-off and later picked up by Ketchup Entertainment for roughly $50M. Hollywood accounting remains undefeated as content strategy theater.

The FT reports LIV Golf is preparing for a possible Chapter 11 filing while talks continue around fresh backing. Sports money is still money, even when everyone calls it vision.

Ars Technica found the most 2026 court filing story imaginable: a man suspected the court might use AI, so he inserted prompts into his filings that were apparently meant to influence any chatbot reviewing the case.

That is ridiculous. It is also useful. Prompt injection moved from security blogs into regular institutional paperwork because people now assume software is reading everything. Sometimes they are right. Sometimes they are just making the judge mad in a new font.

The funny part is the strategy. The serious part is the default assumption underneath it: every process is becoming machine-readable before every process has machine-readable rules.

That's the briefing. Now go build something.

- Michael

P.S. If you are adding AI to a workflow this week, write down the part that breaks when the vendor is down. That is usually where the real architecture starts.

🎧 Podcast note: latest Transistor API verification was blocked today, so the podcast stays footer-only until the authenticated readback works.