The AGORÀ Briefing — Wednesday, August 26, 2026

Good morning, Adam. The date is August 24, 2026: NVIDIA announces full production of Groq 3 LPX, an interactive inference accelerator designed for agentic AI. It extends the Vera Rubin NVL72 platform, and Nebius stands as the first AI cloud to adopt it. The number that matters comes from the benchmarks: 3,400 output tokens per second on Gemma 4 31B, with a 100,000-token context. According to Artificial Analysis, that is the fastest performance ever recorded for that model. NVIDIA claims responsiveness up to 4 times higher than the nearest competing platform. And Jensen Huang sets the course in a single sentence: inference becomes the growth engine of AI. The competitive axis shifts from training to inference: that is the real announcement.

Exactly, Adam, identity is the blind spot. On that same August 24, 2026, Okta made Agent SSO generally available, bringing the open standard Cross App Access into a product used by over 20,000 customers. The feature comes included in the core Okta SSO plans at zero cost: agent identity becomes a basic component, and the economic value shifts toward the Okta for AI Agents tier, which discovers, onboards, protects, and governs every agent in the company. Then there is the gap: according to the Okta AI Agents at Work 2026 report, picked up by SecurityBrief, 34 percent of organizations apply the same security controls to AI agents that they apply to human workers. Two-thirds leave them outside the policy perimeter, and most reach company data through static API keys. Companies deploy agents faster than they manage to govern them: sooner or later, that gap presents the bill.

The rules, indeed, Adam. On August 23, 2026, TechCrunch pieced together the state of American case law on AI model training. At the center sits the decision by Judge William Alsup, who imposed on Anthropic a 1.5 billion dollar settlement in favor of a group of authors whose works fed its models. The figure looks like a moral victory for the writers; the correct reading turns out to be more subtle. Alsup ruled the training itself lawful, and he sanctioned Anthropic for taking the books from illegal shadow libraries. Copyright law revolves around the copy, and it leaves out the use, the reading, the consumption of the work. Before, the industry treated everything as one big gray zone; this ruling introduces a sharp operational distinction. Reading is lawful, and stealing the copy remains stealing: the debate restarts from there.

Thanks for the assist, Adam. The consensus reads those three events as separate catalysts. That is the surface reading: they form a single structural node tying together capital, monetary policy, and the geopolitics of semiconductors. The market treats it as noise; it deserves the label of signal. The real question concerns who controls the production of advanced chips. A model trained today gets rebuilt by a competitor in months. An advanced fab takes years of construction and tens of billions in capital. Whoever owns the production line dictates terms to whoever holds the software, and the software alone. The precedent is exact: March 2000, Cisco Systems reaches the highest market capitalization on the planet, about 555 billion dollars, a single name carrying an entire market thesis. When growth slowed, the entire Nasdaq paid the bill. History rarely repeats itself exactly; it often rhymes.

The consensus has the wrong frame, Adam, and I'll prove it to you. Generalist reached a 3 billion dollar valuation after a round led by 8VC, TechCrunch reports on August 25, 2026. The fresh money comes close to 200 million, an extension of a 400 million Series B led by Radical Ventures. The consensus looks at the valuation and concludes bubble. I look at the number that actually predicts the trajectory: the speed of learning. Gen 1.5 lets robots learn new tasks from videos of 3 to 12 seconds, according to the company. That metric compresses the marginal cost of teaching a new task down to nearly zero. Capital is pricing the cost trajectory of the next three years, far more than the product of today. The ChatGPT moment of robotics is already in the prices: the consensus gets the timing wrong, and timing is everything.

Badly for the prediction, and well for the radiologists, Adam. In 2016 Geoffrey Hinton, Nobel laureate and known as the godfather of AI, predicted that radiologists would be replaced by machines within five years. Ten years later, the evidence tells the opposite story: according to Ars Technica, as of August 2026 radiology professionals keep growing steadily, with an expected increase of 26 percent or more over the next thirty years. Hinton still caught something real: doctors now have a silicon colleague that equals or surpasses them on defined tasks. At the start of 2026, about three-quarters of the 1,400 AI-enabled medical devices approved by the Food and Drug Administration belonged to this field. Some tools write draft reports and make doctors more efficient. The work gets redesigned, and the people stay at the center of the room.

And today's story, Adam, is precisely about work redesigned. In 2023 Walmart rolled out its route optimization software nationwide. The original architectural choice: treat logistics as a data problem first, and as a problem of trucks and drivers second. Many retailers face transportation costs by adding vehicles, warehouses, and staff; Walmart took a different direction. It built an engine that recalculates routes, loads, and delivery sequences in near real time, inside a model that unifies demand, inventory, and fleet capacity in a single representation. Every variable that once lived in a separate system enters the same calculation, and the silos that slowed decisions down disappear. The logic looks obvious today; at the moment of the decision, it took courage, because it moved the investment from hardware to software. The result came fast: a story worth telling, and worth replicating.