The AGORÀ Briefing — Monday, September 7, 2026

Good morning Adam, and good morning to everyone listening. On September 3, 2026, Forbes revealed the closing of a 1.5 billion dollar round led by Jane Street, a quantitative trading firm. The recipient is Fluidstack, an infrastructure startup for artificial intelligence born in Oxford in 2017 and now headquartered in New York. The funding lifts the valuation above 18 billion dollars, more than double the 7.5 billion recorded at the start of the year. The three cofounders, Gary Wu, Jamie Cox and Cesar Maklary, become billionaires thanks to their respective stakes in the company. The firm provides computing capacity to operators set on competing with Nvidia in the field of chips dedicated to artificial intelligence. The message from the market is clear: capital rewards those who build the compute infrastructure, ahead of those who refine language models.

Hello everyone. The real leverage is in packaging, Adam, and that is where the price of compute gets decided. The cost of AI compute for the next decade will be set in 2026, inside the advanced packaging plants of TSMC and Intel. The consensus looks at nanometers and model architectures. The public debate measures power in FLOPS and in billions of parameters. That number describes the chip. It ignores the bottleneck that decides how many chips exist and at what price: the advanced packaging that joins logic and HBM memory. Doubling the FLOPS of a single chip achieves little when assembly capacity remains the constraint. 90 percent of analysts are right about the present and wrong about the pace of change. The trajectory of CoWoS capacity proves it number by number. This is a regime change, well beyond a technology race: the cost curve of compute depends on packaging more than on silicon design.

Welcome back. Behind that number is a public database: AlphaFold, run by Google DeepMind, the artificial intelligence research lab based in London. As of July 26, 2026, the database reached 200 million predicted protein structures, with coverage extending to nearly every organism with a sequenced genome, as reported by Time. The leap starts from the initial 350,000 structures published in 2021, which included nearly all known human proteins. Demis Hassabis, cofounder and chief executive of DeepMind, described that coverage as the entire protein universe, during a briefing on the same day. The question that weighs on anyone allocating capital to computational biology research remains open: what fraction of these predictions matches structures confirmed in the lab, and what fraction remains a computational hypothesis awaiting verification. The number is public and verifiable. The answer to that question, for now, remains to be written.

Hi Adam, hi everyone. What happens is that the competitive advantage shifts from the product to the protocol. On September 5, 2026, DocuSign made its MCP Server public, with native integration into Claude, ChatGPT, Gemini, Copilot and Slack, according to the account from Yahoo Finance. Into those conversational platforms comes Iris, the company's agreement intelligence engine, which reads contracts and extracts their duration, obligations, renewal clauses and legal risk. The insights become governed actions directly inside existing workflows. The official announcement, published on PR Newswire, calls the whole thing an agreement layer for the agentic enterprise: an infrastructure layer designed to be consulted by any external agent. Here is the technical point. MCP stops being a simple API channel and becomes a shared layer across competing platforms. Whoever controls that layer dictates the rules to everyone who passes over it. The moat, today, is the protocol.

Good day to all of you, and to you, Adam. What happens is that a silent debt piles up, and Harvard Business Review shows it with three converging signals gathered through September 2026: a CEO already set to retire, called back to reverse a measurable decline, thousands of positions cut in a mature sector, and a public testimony about the struggle to align internal culture with the speed of the market. The distance between these signals measures the cost of what we can call leadership debt, the consequence of organizational adaptation that fell behind the pace of external change. The sharpest case is Dan Schulman. He had led PayPal for nine years and was ready to retire, when the Verizon board urged him several times to take the helm of the company. At the end of 2025 he accepted, inheriting market share, share price and customer satisfaction in documented decline. Adi Ignatius tells the story in the interview in the magazine's September-October issue. A board that calls back someone who wanted to retire admits, without saying so, that it waited too long.

Warm greetings to everyone. We saw it in 2011, Adam. That year the ECB raised its refinancing rate twice, in April and in July, up to 1.50 percent. The president was Jean-Claude Trichet, and the logic was to tighten against inflation imported from commodities. The outcome ripened within a few months. The European periphery slid into a sovereign debt crisis, the spreads of Italy and Spain exploded, and in November Mario Draghi reversed direction with two consecutive cuts. An older precedent reinforces the picture: in September 1992 the pound abandoned the European Monetary System under speculative pressure, and the lira followed within days. The context today is different, and the structure of the risk is the same. Deutsche Bank organizes its European outlook around three questions: the resilience of the economy, the end point of the tightening, and the stability of sovereign bond markets. The economists Mark Wall, Clemente Delucia and Yacine Rouimi forecast a deposit rate of 2.50 percent in September 2026, with a concrete risk of a climb above that threshold. The capital rushing toward compute, as we heard at the top of the show, is rushing inside this cycle.