AI Radar for Business — Friday, July 17, 2026

Yesterday's radar explained how AI is starting to be managed as just another budget line item; Friday is the day to see who gets each piece of the business. Anthropic and OpenAI are setting up their own companies to implement their AI in their clients' offices, Apple is proving that the model is an interchangeable part—its AI in China will run on Alibaba's Qwen—and a 27-billion-parameter model now fits on a phone under an Apache license. Three signals of the same shift: value is moving from the model to the implementation.

Anthropic and OpenAI are no longer just selling the model: they are selling implementation, and they call it "the trillion-dollar business"

TechCrunch reveals the scale of the bet: Ode with Anthropic, the implementation company that Anthropic launched in May with Blackstone, Hellman & Friedman, and Goldman Sachs, is already worth $1.5 billion, employs 100 engineers embedded in client offices, and has just acquired the startup Fractional AI; its CEO states that "it's easy to imagine it as a trillion-dollar company." OpenAI has its equivalent, The Deployment Company, and Microsoft set up its own this month with $2.5 billion. For your company: the labs themselves are confirming that the bottleneck isn't the models but making them work in real businesses—the same thesis as yesterday's network of certified partners. These companies will go after large accounts; for an SME, the takeaway is to demand from your integrator what Ode practices: a focus on the business process, not the model, and agnostic architecture—Ode is "Claude-first" but declares itself technologically agnostic, and you should demand the same. Source

Apple can only launch its AI in China with Alibaba's model: the model is interchangeable, regulation is not

China's cyberspace regulator has finally approved Apple Intelligence for launch in China: it will run on Qwen, Alibaba's model, with contributions from Baidu, after Apple also explored DeepSeek and ByteDance. Same product, same text and image generation features in iOS and macOS—but with a different brain underneath because local regulation does not allow theirs; there is no launch date yet. For your company: if the most valuable company in the world has designed its AI to be able to swap the underlying model without rebuilding the product, your system should be able to do the same—treat the model as a replaceable part from the initial design, which is the only real defense against changes in price, provider, or regulation; and if you operate in multiple markets, assume that your AI stack may vary by jurisdiction. Source

A 27-billion-parameter model already fits on a phone, for free and with an Apache license

PrismML has published Bonsai 27B, 1-bit compressed and ternary versions of Qwen3.6-27B: the 3.9 GB variant runs on an iPhone 17 Pro and the 5.9 GB version runs on any modern laptop, retaining 90-95% of the performance of the original model, including reasoning and tool calls. The weights are on Hugging Face and GitHub under the Apache 2.0 license—free commercial use. For your company: "serious" local AI is no longer theoretical: to classify documents, extract data from invoices, or answer queries with information that cannot leave the company (GDPR, professional secrecy, client data), you can now pilot a capable model on an ordinary laptop, with zero cost per token and without depending on anyone's cloud—and if yesterday we talked about rationing token spending, this is the extreme version of that control. Source

What to watch for tomorrow?

Two fronts open today: the World AI Conference in Shanghai, with over 300 product launches and an opening speech by Xi Jinping, and the date that leaks suggest for Gemini 3.5 Pro—Google still hasn't confirmed anything. And the digital Omnibus of the AI Act remains pending in the EU's Official Journal; once published, you'll read about it here with its definitive timeline.