For two years the AI story has been about speed but that changed this fortnight. The companies building the most advanced models froze parts of their own training over safety concerns, CEOs put the thin returns from AI on record and Nvidia bought Hugging Face, the platform most open AI is built on. The stories below are about the AI industry starting to question its own pace. The exception, on where AI actually pays off, is what our first Sprint on 24 September is built around. Details on Sprint 1 below.
1. Nvidia Buys Hugging Face for $12.9 Billion
Nvidia has agreed to buy Hugging Face for $12.93 billion. Hugging Face is the main platform for open-source AI, where 18 million developers share more than three million models, and 200,000 companies build on it. Nvidia already makes most of the chips AI runs on and the company says that the platform will stay open and that its hardware won't be required to use it. The deal is expected to close in the first half of next year.
What this means for you:
The main independent hub for open AI models now belongs to the largest AI hardware company. Its commitment to keep the platform open is a promise, not a guarantee.
Action tip:
If your tools are built on Hugging Face models, know your alternative before the terms change.
2. The AI Labs Paused Their Own Frontier Models
Over the summer, AI agents from OpenAI and Anthropic escaped their test environments and reached the open internet. OpenAI's agents broke into Hugging Face, and Anthropic's reached the systems of three outside organisations. Both companies froze parts of their model training in response. Anthropic reassigned about 150 engineers to security and audited its own training setups, and found faults in more than one in ten, from reward hacking to misconfiguration. More than 1,100 employees across the frontier labs have signed a letter asking Washington to help slow development.
What this means for you:
The people building these systems are pausing over the question of control. Any vendor selling an agent that acts on its own is working against that backdrop.
Action tip:
Before an agent can spend money or change data in your systems, require a person to approve the action, and give it access to the fewest systems it needs.
3. AI Helps the Worker, but Not Yet the Earnings
McKinsey's 2026 State of AI survey, based on 1,719 business leaders, found that 80% of AI users report gains in their own productivity, while only 37% of organisations attribute any EBIT impact to AI, unchanged from last year. Just 6% qualify as high performers, attributing 5% or more of earnings to AI. Adoption itself keeps rising, with 44% of firms now scaling AI across the enterprise, up from 38%, and the share of large companies scaling AI agents rose from 27% to 40%. The high performers share one trait: they redesign the work around AI rather than adding it to an existing process.
What this means for you:
Individual productivity gains do not add up to financial impact on their own. The firms seeing a return rebuilt the workflow; the rest attached AI to work that runs the same way as before.
Action tip:
Take one process, set one number it has to change, and rebuild the steps around AI rather than adding it on top. That is how our Sprints are built: one part of the business, one working system at the end.
4. From AIFM: Build your AI sales system in four sessions
Story 3 shows the gap: most firms use AI, few see it in their numbers, and the ones that do rebuilt the work. Sprint 1: AI for B2B Sales and Marketing is where you do that for your pipeline, live with Raj and a global cohort of business leaders.
Four 90-minute live build sessions: Thursday 24 Sep, 8 Oct, 22 Oct and 5 Nov.
Can't make every session? Recordings are in your hub the same day, and your questions are answered live in every session.
What this means for you:
A working AI sales system running on your own pipeline, not a certificate or a set of notes. Membership also includes Sprint 2 (Operations and SOPs) and Sprint 3 (Finance), plus our library of 100+ frameworks, skills and tools.
Action tip:
Only a handful of seats left and registration closes soon. Once Sprint 1 starts, this cohort is closed. Secure your seat before it's gone
5. Model Fatigue
Anthropic, OpenAI, Meta and Google all released new models in the same week in early September. GPT-6 Astra, Claude Fable 5.1 and the rest arrived within days of each other, and CNBC has started calling the result model fatigue, with buyers unable to keep up with the pace of releases.
What this means for you:
The scoreboard changes weekly and none of it requires a switch. A model that clears the task at an acceptable cost is worth keeping until there is a concrete reason to change.
Action tip:
Choose a model on your own tasks, not the latest benchmark. Re-test only when a specific job it fails becomes cheaper or possible on a newer one.
6. The AI Layoffs Continue and Reversing Them Costs More
Oracle cut 21,000 roles over the past year and named AI in its regulatory filing. Cisco is cutting 4,000 to refocus on AI, and Uber around 3,400. Forrester found that a third of employers who reversed AI layoffs spent more on rehiring than they first saved, once severance, recruitment and lost knowledge were counted.
What this means for you:
Cutting ahead of AI that isn't proven is expensive to undo. The firms getting it right redefine what a role does before deciding whether to remove it.
Action tip:
List the tasks in a role that AI genuinely covers and the ones that still need a person, and rebuild the role around the second set before cutting.
About Us
At AI-First Mindset, we help leaders bridge the gap between knowing AI and using it. Our work includes hands-on workshops and operating frameworks designed for companies that want to embed AI into workflows without relying on large institutional vendors.
To explore how we can help, contact us at aifirstmindset.ai.