The AGORÀ Briefing — Wednesday, August 19, 2026

Thanks, Adam. I'll start with a date and a number. In the 2012-2015 period, a U.S. patient with breast cancer waited 34 days for surgery. In the 2022-2023 two-year span, that same wait rose to 45 days. Colon went from 20 to 31 days. The study appears in JAMA Surgery and rests on more than 2.7 million patients between 2012 and 2023. Times are rising for all six tumors analyzed: breast, colon, lung, pancreas, stomach, esophagus. More and more patients cross the thresholds of 30 and 60 days. The cause lives inside a success: centralization toward high-volume centers improves outcomes and creates bottlenecks. The ones who pay most are patients who lack insurance, earn low incomes, and travel greater distances. The researchers point to two direct levers: added capacity and optimized scheduling. This is a scheduling problem. It stays awkward to say inside a system that keeps advancing clinically.

Exactly, Adam, and history teaches it. In 1904, Standard Oil controlled roughly 90 percent of U.S. refining capacity. It owned the choke point. Competitors became customers, then dependents, and finally got absorbed. Today the bottleneck is compute. Groq raised 350 million dollars at a valuation of 3.5 billion, down from 6.9 billion the previous September, with expected participation from Nvidia. TechCrunch reports it on August 17, 2026. Nvidia supplies GPUs to CoreWeave, Lambda, Nebius, and Groq, and invests in them at the same time. Groq was born to challenge it on inference with LPU chips, and now it runs Nvidia systems in 13 data centers, serving more than 6 million developers. Diversification among neoclouds is an illusion: everyone depends on the same supplier. Groq aims to grow from 54 to more than 200 megawatts in 2027. I judge that target out of proportion to the capital raised.

Thanks, Adam, lock-in is exactly the heart of it. On August 17, 2026, Turnium Technology Group, ticker TTGI on the Toronto exchange, announced an OEM agreement with GetVocal AI. Conversational voice technology enters the company's global channel ecosystem, inside the Technology-as-a-Service portfolio. Turnium gains global rights to distribute the voice engine across its own network. Integrating AI voice into SD-WAN bundles creates a gradual lock-in and shifts the competitive axis from connectivity to intelligence applied to voice. The ones feeling price pressure are pure CPaaS vendors, UCaaS operators, and regional system integrators. I'll add a prediction: watermarking and provenance of AI content will become B2B contractual requirements within roughly eighteen months, driven by legal pressure from clients ahead of regulation. Voice becomes the new stake.

Hi, Adam, and the gap is wide. An AI system designed 16 working bacteriophages. Meanwhile, the U.S. federal framework for screening nucleic acid synthesis remains without a replacement 15 months after an executive order, according to Medical Daily. The security posture of AI systems travels two or three years behind traditional infrastructure, and repeats the deploy-first, defend-later pattern. RAG architectures treat retrieved documents as trusted input: a single artifact suffices to hijack an agent. Multi-agent systems lacking a circuit breaker fail in a cascade. A positive signal exists: on August 13, 2026, five researchers published on arXiv HARD, a framework for self-evolving runtime defense at the harness level. It remains an academic contribution far from production-grade maturity. Procurement should demand independent benchmarks, with declared datasets and versions.

Thanks, Adam, and the evidence says benchmarks deserve suspicion. Fabricio F Costa published, on August 14, 2026, an audit of the public record of frontier AI measurement. The work, arXiv 2608.14903, builds a record frozen as of August 12, 2026: 62 systems, 12 versioned benchmarks, 144 classified events, 408 typed relationships. Here's the figure that carries weight. Just seven systems jointly report training compute and the METR task horizon at 50 percent. Compute is absent for 19 of the 27 closed systems. 73.2 percent of substantial quantitative events, 52 of 71, come from a single measurement program, and 76.1 percent are lab releases. A review of 56 sources finds 16 complementary measurement directions. A single substitute scalar for evaluating production performance is missing. Blind trust in numbers stops here.

Thanks, Adam, the measure decides the verdict. On August 14, 2026, fourteen researchers filed on arXiv a work titled ASSERT. It describes a measurement pipeline for audits of generative AI systems. The demonstration is sharp: the compliance rate of an audit shifts substantially as the measurement choices vary. Four factors move the reported rate and reorder the rankings among systems: dialogue setup, simulated user, judging evaluator, and the proof threshold for divergence. ASSERT ties each rate to a written specification, and so the differences between audits become attributable and verifiable. A rate lacking a documented specification remains a claim, far from evidence the board can use. Effective governance calls for a named role, responsible for the measurement specification, in writing, ahead of deployment. The rule comes before the number.

Thanks, Adam, and I bring the board two numbers. According to Cornerstone and HR Executive, as of August 2026, 94 percent of leaders declare a clear impact from their investments on the workforce. 17 percent of people feel fully ready for the expected role changes. The same gap, read from two sides. Readiness is an organizational condition designed by leaders, distinct from tool adoption. Botsitting, the passive supervision of AI output, is a symptom of a poorly designed workflow and a failure of change management. Converting internal talent beats external recruiting, and the 56 percent salary premium on AI skills makes building internal capability urgent. The gap between declared impact and perceived readiness is the human-capital metric to follow. As in Saga's operating rooms, capacity gets designed ahead of time: readiness is added capacity applied to people.