Suprvisr TLDR - July 20 2026

Every business runs on follow-ups: the recap after the meeting, the nudge to the client who went silent, the status report nobody enjoys writing. They eat hours every single week, and they're exactly the work AI is already great at. Cumulus does the whole loop: it drafts every follow-up and report in your voice, updates the systems that need updating, and chases the people who owe you answers. Every step logged, so you can always see what it did. It's the first job our customers hand it, because the hours come back immediately.

No credit card required.

For years the AI story was a talent contest: whose model is smartest, whose demo is slickest. This week it turned into something older and blunter, a fight over who makes the rules.

Start in Shanghai. On July 16, 29 countries signed up to a new China-led body, the World Artificial Intelligence Cooperation Organization. Think of it as a trade club for AI. The membership skews toward the developing world (Russia, Pakistan, Indonesia, Kazakhstan, Laos), and the pitch is simple: build your own national AI, and if you can't, China will hand you a capable one for free. To prove the point, a Chinese lab called Moonshot released exactly that the same day: Kimi K3, the largest AI model anyone has ever given away. "Given away" is literal. Anyone can download it, run it on their own machines, and build a business on it without paying the American labs a cent. The UN Secretary-General attended the signing. The United States did not.

Days earlier, the other side moved. The heads of Google DeepMind, OpenAI, and Anthropic, three men who agree on roughly nothing, each published nearly identical plans for how AI should be regulated, essentially volunteering for supervision before Washington invents something they'll like less. All three want independent testing and a central referee. They differ only on what kind of referee: one modelled on the FDA (approve models before they launch), one on FINRA (the industry polices itself), or one on the IAEA (an international body with inspectors).

For now, most of this is flag-planting. But the rules being written this year will shape what companies can and can't do with AI for the next decade, and right now the two drafts don't match.

That's how many internal dials Kimi K3 turns to work out an answer, which makes it the biggest AI anyone has ever released for free. China spent the money to build a top-tier model, then handed out the finished product for a sliver of what the American names charge. For the first time, the free option genuinely competes with the paid ones. Hold that thought, it comes back at the bottom.

Here's an awkward dinner party. Since 2024, ChatGPT has been built into the iPhone under a high-profile Apple and OpenAI partnership. On July 10, Apple sued that partner in federal court, alleging OpenAI ran a coordinated campaign to pull Apple's hardware secrets out of current and former employees, directed, Apple claims, from the very top.

The filing reads like a corporate thriller written by people with terrible operational security. Apple alleges one former engineer kept his company laptop after leaving for OpenAI, discovered he could still reach Apple's internal file storage, and downloaded confidential documents about unreleased products. One message quoted in the complaint: "LOL, I found out I can access the [network storage], so funny." Roughly 400 former Apple employees now work at OpenAI, which is racing to ship its first consumer device this year, shaped in large part by Apple's former design chief Jony Ive. Elon Musk and Sam Altman then spent the weekend arguing about all of it in public, because of course they did.

The partnership that put ChatGPT on the iPhone is, remarkably, still alive, which should make the next Apple event delightfully tense. And 400 employees crossing the street is more pipeline than leak. Every company losing people to a rival is running a smaller version of this exact experiment.

🖥️ The most important AI company most people can't name. TSMC, the Taiwanese firm that manufactures the chips nearly every AI runs on, just booked a record quarter near $39.6 billion. The shortage is real enough that Google recently turned Meta away, no chips to spare. When even the famous names can't get enough, the whole race slows to the speed of one factory in Taiwan, which is why Amazon, Microsoft, and friends are now designing their own chips, so they never have to hear no again.

🐢 Google's flagship model is still missing. Gemini 3.5 Pro was promised for June, then July, and as of this weekend it still hasn't shipped. Reports say Google rebuilt large parts of it after early versions struggled with long coding jobs and multi-step accuracy. So in the same week China gave away its best model, Google couldn't get its own out the door. A useful reminder for your next vendor roadmap meeting: "coming next month" is a mood, not a date.

💼 The real money moved downstream. Anthropic, Blackstone, and Goldman Sachs just put $1.5 billion into a new company, Ode, betting that the next fortune in AI belongs to whoever makes the models work inside real businesses. Ode sends elite engineers directly into client companies to do exactly that, competing with OpenAI's own deployment arm and consultants like Deloitte. When the smart money stops paying for the model and starts paying for the results, you can see where the value went.

🧳 A startup hit $1 billion without cutting staff. Fora, a travel company, just crossed a $1 billion valuation on a $60 million raise, and how it got there is the interesting part. Instead of using AI to replace its human travel advisors, it used AI to let each advisor handle far more clients. After two years of investors rewarding tools that trim headcount, they just funded the opposite bet: keep the people, equip them better, and let the math do the rest.

Two things from this week are worth acting on. First, the Apple lawsuit is a blunt reminder that when people leave, knowledge leaves with them, and most companies only learn what walked out the door when it turns up somewhere else. Worth an hour with your team this month: what do departing employees still have access to, and for how long? Apple, with all its resources, found out via a still-working laptop login.

Second, Kimi K3 proved the free models are now genuinely good. That changes what's possible: you can run a serious AI on your own hardware, in your own building, where your data and your know-how never leave. A capable GPU server now costs less than one mid-level hire, and unlike a hire, it can't be poached. Suprvisr AI plans and implements open-source AI deployments end to end, and Cumulus runs happily on-prem against your own model endpoint, which we can set up too. If "our AI, our building, our rules" sounds appealing, start with a call.

Same time next Monday. We'll do the reading, you bring the coffee.