AI Radar for Business — Sunday, September 13, 2026
· CompaniesAutomation
Meta reappoints managers after its experiment with agents, Anthropic accuses five Chinese labs of distilling Claude, and the Pentagon negotiates 5 billion for the computing chain.
Three things from the weekend. First: Meta, which had spent a year removing management layers because AI was going to do the coordination work, is now asking for volunteers to become bosses again. Second: Anthropic has named and provided figures for the laboratories that were "milking" its model to train their own, and some of those names are on your list of cheap providers. And third: the money is no longer coming solely from venture capital, but from the Pentagon, which is negotiating a $5 billion loan to manufacture data center parts. Additionally, below, the unified command that Salesforce wants to sell you to govern agents from three different providers and 146 free AI activities for SMEs in Spain. Yesterday's radar was about agents attacking servers; today is about the organization that must support them.
Meta reappoints managers: AI multiplied code but product barely moved
Meta is asking engineers in its Applied AI division—created this year with about 7,000 reassigned employees—if they want to voluntarily return to manager roles after more than a year of flattening the organization and cutting around 10% of the workforce in May (about 8,000 positions), with middle management being the most affected. The context lies in internal data from the restructuring plan involving agents that the company canceled hours before execution: code changes increased by 220% while features delivered to the user only grew by 36%, technical incidents rose by 40%, and time spent fixing them by 70%. For your company: if you measure AI's effect on activity—lines, closed tasks, touched tickets—you will see a beautiful number that means nothing. Measure what is delivered and what it costs to maintain: features in production, incidents per week, and rework hours, before and after. And keep that coordination layer: automating execution without anyone to decide what gets done multiplies work in progress, not results. Source.
Anthropic points to five Chinese labs for milking its model: where your cheap model comes from
In its September threat report, covering December 2025 to August 2026, Anthropic accuses Alibaba, Moonshot AI, DeepSeek, Xiaomi, and Zhipu of illicit distillation: using Claude's responses to train rival models. Accounts linked to Alibaba generated more than 151 million interactions between May and July, with peaks near three million daily across more than 3,500 fraudulent accounts, and the total campaigns approach 200 million exchanges; what they sought was exactly what is expensive to copy: agentic reasoning and software engineering. This is the flip side of price reductions: DeepSeek just priced its V4.1-Flash at $0.15 per million input tokens during off-peak hours. For your company: price per token is not the only purchasing criterion. Request in writing from each provider the origin of the training data and who assumes liability if a third party claims; and review your own API keys, as part of that traffic entered through resellers and intermediary accounts. If you use a flagged model, have an alternative ready before a contract or a court decides for you. Source.
The Pentagon becomes a cloud lender: $5 billion for Fluidstack
The US Department of Defense is negotiating a loan of approximately $5 billion to Fluidstack through its Office of Strategic Capital, not to build data centers, but to secure the manufacturing of components that data centers need. This would be by far the largest loan from that office, which until now financed rare earths and drone manufacturers. The deal is not closed, and neither Defense nor the company have commented. The figure comes a week after Fluidstack closed a private round of 1.5 billion, so the same provider is receiving private capital and public credit simultaneously. For your company: when a State enters to finance the computing supply chain, it is because capacity is being reserved years in advance, and that is felt downstream in timelines and rates. Ask your cloud or model provider for two specific things in the next renewal: until what date they have committed capacity for your workload and what price revision clause applies. And do not sign three-year volume commitments without an exit window. Source.
Your agents already come from three different providers and no one is keeping track
Salesforce has introduced its Trusted Enterprise AI Harness, a suite of six capabilities—context, agency, action, governance, security, and model routing—with a control plane that promises to discover and register agents, give them identity and policy, measure them, and control their cost, including third-party agents. The detail that explains the move is that companies already run an average of 3.1 orchestration platforms and 85% have two or more. The essentials are available now, but the unified experience doesn't arrive until the start of fiscal year 2028, in February, and still without pricing. For your company: there is no reason to wait until February 2027, because 80% of the work is not buying a product but doing the inventory. A sheet with four columns—which agent, which credential it enters with, what permissions it has, and how much it spent last month—and one rule: no new agent enters production without a named owner. When the control plane arrives, that sheet is what you will upload to it. Source.
146 free AI activities for SMEs: Acelera pyme network launches its first AI Week
Red.es is holding the first AI Week from September 21 to 24 across the Acelera pyme Office network: 54 offices distributed throughout the territory and 146 activities including seminars, practical workshops, webinars, and diagnostic sessions on productivity, process automation, marketing, data analysis, and safe use of AI. The network, under the Ministry for Digital Transformation, has 76 offices, received 29 million euros in the 2025 call with ERDF funds, and claims to have reached more than 200,000 SMEs since 2020. For your company: it's free and in-person, so the real cost is two hours of one person's time. Go with a specific process written in one line—the one that most often forces you to copy data from one place to another—and leave with a diagnosis and a contact name, not a brochure. If no one from your team can go, check the program for your province and at least note which solutions they are recommending to companies of your size. Source.
What to watch tomorrow?
Two things: whether any of the labs pointed out by Anthropic respond—so far, silence—and whether the Pentagon's loan to Fluidstack is confirmed or falls through, as it sets the precedent for the rest of the computing chain.