The AGORÀ Briefing — Friday, September 11, 2026

Good morning Adam, good morning everyone. The European Central Bank staff projections, published in September, list investments tied to artificial intelligence alongside defense and infrastructure, as explicit drivers of medium-term growth. The document sets real gross domestic product growth at 0.9 percent in 2026. The precedent sits in 1999. Back then, Alan Greenspan's Federal Reserve wrote computer-driven productivity into its monetary policy reasoning. Private capex in networks and semiconductors was the premise, and the central bank kept conditions accommodative. In 2001, telecom capex collapsed, recession followed, and that productivity anchor turned out to be a market assumption dressed up as a monetary fundamental. The structure stays identical: the market dictates the investment cycle, and the central bank ratifies it.

Hi Adam, hello to everyone listening. Part of that money goes into light. On September 10, 2026, at the China International Optoelectronic Exposition in Shenzhen, Huawei showed what it calls the world's first near-packaged optics module capable of reaching 7.2 terabits per second, combining 36 channels of 200 gigabits each. The South China Morning Post reports it. The module replaces copper with light to carry data between accelerators, with less signal loss and less power consumption than current systems. Man Jiangwei, director of the company's advanced optoelectronics lab, stated that the product has entered the industrial development phase and that large-scale production will arrive as soon as the supply chain is ready. It remains an announcement, still far from wide commercial availability. The underlying signal is clear: the race changes target, from compute silicon to the fabric that links thousands of processors together.

Good to be here, Adam and listeners. I found a paper published on September 10, 2026 that tells a different story, with verifiable numbers. A controller learns handwriting in 18 seconds on a laptop CPU. At that point the cost of training data for a robot foundation model becomes marginal, and the bottleneck migrates from compute to judgment. The consensus has the wrong frame: it measures compute scale and data volume, and ignores the variable that truly predicts dexterity, which is real-time adaptation on the physical robot. Modern simulators struggle to reproduce the complexity of contact between hand and object, and collecting dexterous demonstrations remains an open and costly problem. The paper cites both limits as motivation, then sidesteps them: the controller starts from an estimate of the hand-object Jacobian, computed live on the robot. This is a regime change, far more than a passing trend.

Greetings to everyone, and to you, Adam. Your doubt already has an answer on arXiv. A study published on September 5, 2026 describes for the first time a resource-exhaustion attack on vision-language models that treats prompt and pixels as joint adversarial variables, instead of isolating the image channel as the sole lever. The authors are six researchers, including Zhaoxiong Ni and Yatie Xiao, and they call the method Joint Pixel-Prompt Optimization, JPPO for short. It is the extended version of a work accepted at the ACM CCS conference. Until today these attacks assumed a unimodal threat model: the prompt channel stayed fixed, visible to the user, excluded from adversarial optimization. The team evaluates five families of open-source models on MS COCO and ImageNet, with a perturbation budget set at 8 out of 255 in the infinity norm. That budget defines how imperceptible the perturbation can stay. The attack surface, in practice, doubles.

Adam, good day to you and to everyone following us. Salesforce says it has the answer, and calls it Control Plane. The company has made public a new architecture, the Enterprise AI Harness, designed to give agents a shared understanding of the customer and the enterprise. The project brings together six capabilities, called context, agency, action, governance, security and models, on a common, composable architecture. Around it comes the Control Plane itself: a single point from which to see, manage and control agents as they multiply. The stated example is the question: can we fulfill this order today? Answering requires data scattered across CRM, ERP, contracts and internal policies. The agent has to understand that context, decide the right action, and act within the boundaries set by the enterprise. The technical promise is high: open reasoning tied to controlled execution. The weak spot lies in governance, which runs far slower than the agents it is supposed to keep in check.

Hello everyone, and hello Adam. The pause has a date: July 13, 2026. That day the Department of Defense suspended the move to Phase 2 of the Cybersecurity Maturity Model Certification, the program built over five years to certify the cybersecurity of the defense industrial base. The decision comes through two administrative memoranda, one from the department's Chief Information Officer and one from the Under Secretary of Defense for Acquisition and Sustainment. Together they order that self-assessment alone be accepted, remove higher-level requirements from active solicitations and existing contracts, and block every waiver during a 60-day review. Lawfare tells the story. The department speaks of relief for small supplier firms, and the reasoning has a basis: Level Three certification costs a great deal for companies with thin margins. The method, on the other hand, radically changes the legal picture: a program resting on formal rules gets stopped by simple memoranda.

Happy listening to everyone, and welcome back, Adam. In Sharjah they did exactly that. On September 9, 2026, the Sharjah Digital Department gathered 24 chief executives and directors general from 21 government entities of the Emirate, for a workshop on executive leadership in the AI era, organized with PwC Academy Middle East and covered by Voice of Emirates. The makeup says something precise: leaders at the top, called to decide where to place authority between people and intelligent systems, even before choosing which tools to adopt. The thesis running through the evidence gathered in recent months stays consistent: leaders' ability to govern adoption weighs more than the quality of the tools. Cornerstone has documented a gap between the 94 percent of leaders who perceive an impact of AI on work and the 17 percent of people who feel ready to operate with these tools. The bottleneck remains leadership, and that gap says so clearly.

Warm greetings to Adam and to everyone: the story begins in Ottawa. That is where Kinaxis is based, a company that builds supply chain management software. Its platform is called Maestro and tackles a problem that touches every sector: geopolitical uncertainty makes logistics plans obsolete within a few weeks. Manufacturing, semiconductors and shipping companies ask themselves the same question: how to predict the impact of an event thousands of kilometers away, before it becomes a production delay or an empty shelf. The Strait of Hormuz, shifting tariffs and trade tensions demand quick answers. Kinaxis combines predictive artificial intelligence, scenario modeling and autonomous agents, and chooses orchestration over static forecasting: a single environment with sales data, production capacity, inventory and external signals such as weather and market news. When the world changes in a week, the plan changes with it.