This was a week of movement at the top. OpenAI shipped a new frontier model and then held part of it back over safety concerns, Apple changed its chief executive for the first time in 15 years, OpenAI's advertising business crossed a billion-dollar run rate, and Nvidia paid to keep even its competitors building on its platform. The common thread was leverage: who holds it, and how they keep it. Here is what mattered.
OpenAI released GPT-6 Astra, and held part of it back
On September 3, OpenAI released GPT-6 Astra to a limited set of trusted partners, opening it to paying users the following day. The company calls it its most capable and most aligned model, and says it came from its largest training run yet, the first pretrained on more than 100,000 GPUs at its Stargate site in Texas. The notable part is the caution: OpenAI says Astra crossed the "Critical" threshold for cybersecurity under its own Preparedness Framework, so it is rolling the model out carefully, with a version that refuses certain security-related prompts and full access limited to vetted users at first. [CNBC, Al Jazeera]
Our take: the most telling detail is not the benchmark scores. What matters more is that OpenAI deliberately held part of the release back. When the company that builds the model decides the newest version is too capable to hand out freely, that is a signal worth catching: the most powerful option is not automatically the right one to wire into your product. Capability now comes with sharper edges. The sensible move is to match the model to the job, running most work on a smaller, cheaper, well-understood model and reserving the frontier for the tasks that genuinely need it. We choose models that way on purpose, so you pay for power only where it earns its place and you are not carrying risks the task never called for.
Tim Cook stepped down as Apple's CEO
On September 1, Tim Cook stepped down as Apple's chief executive after 15 years, handing the role to John Ternus, the senior vice president who ran hardware engineering. Cook stays on as executive chairman of the board. Under him, Apple grew from roughly a $350 billion company into the second business in history to touch a $5 trillion market value, and the handoff comes while Apple is working to catch up in AI. [MacRumors, 9to5Mac]
Our take: the quiet story is who Apple chose. Ternus came up through hardware engineering, not software or services, which hints at where Apple believes its next decade is won: in chips and devices, the physical layer that AI now runs on. For everyone else, a leadership change at a key vendor is a good moment to check your own dependencies. When a platform you rely on shifts direction, your plans should be able to bend with it. The habit worth keeping is to avoid staking a core process on one vendor's roadmap staying exactly as it is today. We design systems that can adapt when the ground moves upstream, so a change of leadership somewhere else does not force a rebuild on your side.
OpenAI's ChatGPT ads business hit a $1 billion run rate
On August 31, OpenAI said its ChatGPT advertising business had reached a $1 billion annualized run rate in under 200 days, and it opened self-serve ad buying to more markets across Europe, India, and the Middle East, with ads now running in more than 40 countries. Even so, that pace is well short of the roughly $2.5 billion in ad revenue the company had reportedly told investors to expect for 2026. [CNBC, Digiday]
Our take: ChatGPT is becoming an advertising channel, and that shifts the calculation for any business that cares how customers find it. When people ask an assistant instead of running a search, the assistant's answers, and soon its ads, become the new shelf space. It is worth watching closely, because the rules for showing up inside an AI answer are still being written and are cheaper to learn now than later. It is also a reminder that a free assistant makes its money from attention, so anything sensitive you type into a consumer AI tool deserves a second thought. When we build with these models, we keep client and customer data on the paid, non-advertising side of the line, where it is not feeding someone else's ad engine.
Nvidia put $3.5 billion into MediaTek
On September 1, Nvidia said it will invest $3.5 billion in MediaTek, its first investment in a Taiwan-listed company, buying convertible bonds and deepening a partnership in which MediaTek will adopt NVLink Fusion, Nvidia's technology for letting other companies' custom AI chips connect into Nvidia's rack-scale systems. MediaTek's shares rose about 10 percent on the news. [TechCrunch, Nvidia]
Our take: the move to watch is NVLink Fusion, the part that lets rival chips plug into Nvidia's systems. Even as competitors design their own AI chips to escape Nvidia, Nvidia is turning itself into the thing they all connect to. That is a masterclass in staying essential, and the lesson runs well beyond chipmaking: the durable position is to own the layer others build on. For most businesses the practical version is smaller but just as real. Hold on to the part of your operation that is genuinely yours, your data, your customer relationships, your workflow, and treat the rest as interchangeable parts you can swap. We build around that principle, keeping what makes you distinctive under your control and everything else replaceable.
The one thing to remember
Every big story this week was about leverage at the top of the industry. OpenAI shipped its most capable model and then held part of it back, Apple changed its leader for the first time in 15 years, OpenAI's ads business crossed a billion-dollar run rate, and Nvidia paid to keep even its rivals plugged into its platform. The same question sits underneath all of it: who holds the power, and how do they keep it. For a business using AI, that question applies in miniature. Know which parts of your operation are truly yours and which are rented, keep the rented parts swappable, and spend your energy on the advantage no vendor can hand you or take away.