|
⏱ The 10-Second Version
- Moody’s warned that banks’ growing dependence on a handful of AI providers, chiefly OpenAI and Anthropic, creates systemic risk, saying an outage at one could ripple across the financial system and let a dominant few set AI’s price.
- Meta released Muse Glimmer, a free, 30-billion-parameter AI model that runs offline on a single consumer GPU, its clearest answer yet to the “you need our cloud” pitch every closed AI vendor makes.
- Federal Reserve Chairman Kevin Warsh is rebuilding how the Fed reads the economy around AI, running internal bots named after economists Milton Friedman and James Tobin and pushing for real-time data over delayed government surveys.
|
|
The Big Thing
Moody’s Warns Banks’ AI Dependence Risks Vendor Pricing Power
What Happened
Moody’s warned this week that banks racing into AI are building a different risk: dependence on a tiny cluster of foundation-model and cloud providers, chiefly OpenAI and Anthropic. A model outage at one major provider, the rating agency’s report says, “could potentially spread quickly across customers and sectors,” and as these firms chase investor returns, a dominant few “could exert control over the price of AI services.” More than three in four City firms already use AI; Lloyds alone is spending £13 billion, partly on AI-driven cuts. Moody’s isn’t only worried about banks: any company betting its workflows on one or two AI vendors carries the same exposure. Read the full story at the Guardian, via AOL →
Why You Care
Ask your team this week whether you could actually switch AI vendors, or just assume you could.
|
What to Say in Your Next Meeting
“Moody’s just warned that a handful of AI firms, chiefly OpenAI and Anthropic, could end up setting AI’s price for the entire banking system.”
|
|
|
Speed Round
Meta released Muse Glimmer, a 30-billion-parameter open-weight AI model under an Apache 2.0 license that runs offline on a single consumer GPU, handling coding, scheduling, file organization, and multimodal tasks across 100+ languages; Meta says it beats comparably sized open models from Google and Alibaba on several agentic and coding benchmarks. Every other lab sells you a subscription; Meta just open-sourced the argument for canceling it.
Read the announcement at Meta AI Research →
Federal Reserve Chairman Kevin Warsh is rebuilding how the central bank reads the economy, running internal AI systems named “Milton” and “Tobin” after economists Milton Friedman and James Tobin, and pushing the Fed toward real-time retailer and bank data instead of delayed government surveys; he’s also proposed cutting policy meetings from eight a year to six. The Fed used to wait a month for numbers it trusted; now it’s running its hunches past a chatbot named after an economist.
Read the full story at Axios →
AMD agreed to acquire Taalas, a Toronto startup that etches AI model weights directly into silicon instead of loading them into memory; its first chip reportedly served Meta’s Llama 3.1 8B at close to 17,000 tokens per second, a claimed 73 times an Nvidia H200’s throughput at a tenth of the power, though each chip only runs the one model it was built for. Model-specific chips are blazing fast right up until you want to upgrade the model.
Read the full story at SiliconANGLE →
South Australia announced the country’s first royal commission into AI, a roughly $3 million inquiry launching in October and covering education, employment, health, and the arts, with a final report due mid-2027; premier Peter Malinauskas called it “a really important fork in the road” after meeting OpenAI, Anthropic, and Apple on a US trip. A royal commission on AI, nine months after everyone already deployed it, is either prudent or extremely on brand.
Read the full story at SBS News →
A new Gallup poll found seven in ten Americans oppose new data centers in their own communities, with opposition running through both the left, framing them as corporate boondoggles, and the right, worried about AI controlling information; commentators are calling it one of the most bipartisan issues in years. It’s rare to find something Americans agree on this year; a building full of GPUs humming next door managed it.
Read the full story at Forbes →
|
|
Workflow Teardown
The Vendor Escape Hatch Test
Moody’s just warned that banks, and by extension anyone leaning on a handful of AI vendors, are exposed if one provider raises prices or goes down. The same week, evidence a real alternative exists landed for free: Meta’s Muse Glimmer, a 30-billion-parameter model that runs offline on a single consumer GPU. Whether or not you’d ever run Glimmer yourself, the question it raises is worth two minutes: if your main AI vendor doubled its price tomorrow, would you actually have a backup, or just the assumption that you would?
- Pick the one AI-powered task your team would feel first if your main vendor raised prices or went down for a day.
- Note what actually makes that task hard to move: proprietary data, a specific integration, or a capability only that model has.
- Paste it into an assistant with the prompt below to find out what you’d really lose by switching.
Here's an AI-powered task my team relies on through one vendor:
[paste: e.g. drafting contracts, ranking sales leads, a specific
integration or automation]
Tell me:
1. What actually makes this hard to move to another provider, including free or open models
2. What we'd really lose in quality or speed by switching, versus what we just assume we'd lose
3. Whether we have zero, one, or more than one working backup today
Be honest if the answer is zero.
|
|
Forward this to the coworker who assumes switching AI vendors would be a bigger project than it actually is.
— Don.
|
