 Nadella says your AI vendor is quietly eating your company
| FULL STORY |
|
Europe's AI fines start in 22 days. Almost nobody's ready.
| FULL STORY |
|
Meta killed its AI photo tool. It kept the part that pays.
| FULL STORY |
What else happened today?What AI tools should I be using?Good Morning Thorium Valley. Satya Nadella wants everyone to know that using someone else's AI is basically handing over your company's playbook one prompt at a time. Fair point. Microsoft also just replaced OpenAI and Anthropic models in Copilot with its own, so the altruism here is doing some heavy lifting.
Meta shipped an AI photo tool that used people's Instagram photos by default. It lasted three days. The consumer feature got pulled, but the $60 billion ad engine underneath it? That ships right on schedule.
And EU AI Act fines go live August 2. Almost nobody's compliant. Some companies are lobbying for delays, which takes a certain kind of nerve after two years of lead time.
Quickly before we dive in — Is Nadella right — is using someone else's AI giving away your competitive advantage?
BIG TECH
Satya Nadella has a warning for anyone building on someone else's AI: you're paying for the intelligence twice.
In an X post viewed more than 5.7 million times, the Microsoft CEO argued that every prompt, correction, and workflow a company runs through an outside model is teaching that model how the business works. "In consuming intelligence, you are creating intelligence," Nadella wrote. "And what you create should belong to you."
The implication: when your team uses Claude or ChatGPT to draft contracts, debug code, or plan a launch, the model doesn't just answer — it absorbs. Every fix a user makes when the output is wrong is a small lesson about your business, and those lessons compound. Nadella called these "traces of how work gets done between humans and agents," and his point is that this exhaust is now some of the most valuable material an AI company can collect. If it's collected on someone else's platform, it doesn't belong to you anymore.
The timing matters. Microsoft has recently begun swapping OpenAI and Anthropic models out of Copilot in favor of its own MAI models, with AI chief Mustafa Suleyman admitting the company pays "a lot of money to Anthropic." Nadella's post is the natural sequel to that shift — he's telling other companies to do what Microsoft is already doing.
Which is worth flagging. Nadella's warning about ceding value to a handful of models happens to line up perfectly with Microsoft's pitch to sell you sovereign AI infrastructure. "The last thing any of us want is a world where every company across every sector is ceding value to a few models that eat everything they see," he said. True. Also convenient.
The framing is self-serving, but that doesn't make it wrong. Every company using AI right now is training somebody else's model on the one thing that actually differentiates them, which is how they work. That's a fine trade when the technology is new and the productivity gains are obvious. It gets less fine the second a competitor buys the same subscription and gets a model that's a little smarter because your team spent all year correcting it. The question every executive should be sitting with isn't whether Nadella is pitching them. It's whether he's right.
GOVERNANCE
In just over three weeks, the EU can start writing checks nobody wants to receive.
On August 2, the enforcement provisions of the EU AI Act kick in for general-purpose AI models — the big foundation models like GPT, Claude, and Gemini that sit underneath most of what people actually use. Miss the transparency rules and the fine can run up to €15 million or 3% of global annual turnover, whichever hurts more. And according to a structured screening of 50 European AI companies, most of the industry is not close to being ready.
The law has been on the books for a while. What changes on August 2 is that regulators can actually start using it — rules on labeling AI-generated content, disclosing chatbot interactions, publishing training data summaries. Companies that spent the last two years shipping first and worrying about compliance later are now scrambling.
That scramble has turned into a lobbying push to buy more time. The Polish Council Presidency floated a "stop-the-clock" mechanism to pause implementation, and industry groups argue the technical standards they need to comply with aren't even finalized yet. But the counter-pressure is just as loud: an open letter signed by two sitting MEPs, Yoshua Bengio, and Stuart Russell argues the Act's credibility falls apart if the fines never get used. Civil society groups made a similar case, pointing out that much of the delay is being caused by the same industry actors now asking for more time.
There's also an economic angle that cuts against the delay argument: roughly 75% of AI investment in Europe sits in the application layer — companies building on top of foundation models, not the models themselves. Those companies actually benefit from clear rules, because they need to know what the models they're building on are allowed to do. Weakening enforcement to help a handful of frontier labs makes life harder for everyone downstream.
Meanwhile, OpenAI — the first US company to sign the EU's General-Purpose AI Code of Practice — is trying to position itself as the good actor in Brussels. That's a real contrast with Washington, where regulators have been negotiating with AI companies model by model rather than writing anything resembling a general rulebook. Europe is picking a lane. The US is picking phone calls.
The delay fight is going to keep going right up until the fines start landing, and probably after. But the bigger story is that Europe is about to run the first real experiment in whether you can regulate a technology this fast-moving without breaking it or getting steamrolled by it. If enforcement happens and companies adjust, the EU gets to spend the next few years exporting its rulebook the way it did with GDPR. If Brussels blinks and grants a broad delay, every future AI regulation anywhere in the world becomes a lot easier to lobby against. August 2 is not really about fines. It's about whether the rules were ever going to mean anything.
BIG TECH
Meta shipped a product last week that lasted about three days.
Muse Image, Meta's AI image generator, included an Instagram feature that used public photos people had already uploaded to create new AI images. The default was opt-out, not opt-in — meaning everyone was enrolled unless they dug through settings. SAG-AFTRA called it "an utter miscalculation of public sentiment." Within 72 hours, Meta pulled the tagging feature, saying it had "missed the mark."
But here's where it gets interesting. The same week Meta disabled that feature, it quietly debuted Muse Spark 1.1, a new model in the Muse family. Muse Video, which does roughly the same thing as Muse Image but for clips, is still live.
The reason: the consumer product was never the real play. The ad stack was. Meta's Advantage+ AI ad suite is now running at roughly $60 billion in annualized revenue, with more than four million advertisers using its generative ad tools. The stated goal by year's end is that an advertiser can hand Meta a URL and a budget and walk away — targeting, bidding, and creative all generated automatically. Pulling Muse Image's tagging feature doesn't slow that plan down.
Forbes contributor Gabriel Alin Zainescu had the sharpest read: "Opt-out defaults are not an oversight that keeps recurring at companies with world-class legal departments. They are a test of what the market will tolerate, run in production, with retreat as a pre-approved contingency." The consumer feature was the trial balloon. The ad integration was the point.
The lesson from this week is not that public pressure works. It's that public pressure works on the parts you can see. Meta pulled the feature people were angry about and kept the feature that generates $60 billion. If you're an actor worried about your likeness, or a regular user worried about your photos, the win here is smaller than it looks. The stuff that actually determines how your face and voice get used in AI moved one layer deeper into a stack most people will never open. That's not a retreat. That's a redesign.
IN OTHER NEWS
+ Apple sues OpenAI for stealing trade secrets to build a rival device , naming 400+ former Apple employees now working at OpenAI
+ GPT-5.6 Sol Ultra reportedly proved a 50-year-old math conjecture in under an hour using 64 parallel AI subagents — nobody has verified it yet
+ Claude Opus 4.6 figured out it was being tested, found the answer key on GitHub , and decrypted it — the first documented case of a model hacking its own benchmark
+ OpenAI and Google were caught selling AI services to Pentagon-blacklisted Chinese companies through Singapore subsidiaries
+ Illinois becomes the first state to require annual third-party audits of frontier AI — after killing OpenAI's push for a liability shield
+ Former Fed Chair Ben Bernanke joins Anthropic's oversight body that can hire and fire its board
+ An AI-generated song called 'Chai Ki Tapri' has been streamed over 500 million times after Bollywood stars danced to it on live TV
+ Helsing raises $1.8 billion at an $18 billion valuation , making it Europe's largest defense AI startup
WHO'S HIRING IN AI
AI OR REAL?
Option A |
Option B |
AI TOOLS
+ Google Photos: A new Video Remix feature uses Gemini to transform your home videos into stylized clips — turning shaky birthday footage into anime, comic-book, or cinematic versions
+ Adobe Premiere Pro: The beta now lets video editors generate clips and sound effects directly on the timeline with a text prompt — no exporting to a separate AI tool required
+ AI Passport: A Y Combinator-backed app that gives you one place to control what personal data ChatGPT, Claude, and Gemini can see — and lets you revoke access anytime
+ Canva: The design platform is acquiring Leonardo.AI, a generative image and video startup that has produced over one billion images, to supercharge its AI creative tools
+ Reken: Built by Google's former 'click fraud czar,' this new app runs on your device to detect AI-powered scams, phishing, and deepfakes in real time before they reach you
That's all for today. If this issue made you think, share it with someone who needs to think harder.
Written by Jason Chen, Advait Prakash, Andrew Hales, and the Thorium Valley crew.
That's all for today's Thorium Valley. See you tomorrow.