Daily Briefing
29 August 2026 · 5 stories
🔥 Top story
1A startup launches six AI teams that replace your entire customer-service software stack
Imagine skipping the phone system, ticket tracker, scheduling tool and quality dashboard entirely and instead hiring six virtual employees to do all of it. That is the bet from Crescendo, which on August 25 launched CXP: six specialised AI agents (concierge, agent assist, quality, workforce, optimisation, knowledge) replace the equivalent of six traditional tools. For a 50-person company handling 500 tickets per day, the concrete change is simple: instead of paying for seven subscriptions and three integrators, you configure six agents in four weeks. Whether the bet holds over 90 days remains to be seen. Source : https://crescendo.ai/news/crescendo-launches-the-ai-native-customer-experience-platform/
2NVIDIA ships its first processor built for AI agents -- OpenAI, Anthropic and AWS already received theirs
NVIDIA's Vera processor, designed for the nonstop work AI agents do, is now shipping at scale. AWS received its first machine in Seattle on August 27. Oracle is already deploying at scale. OpenAI, Anthropic and a space-industry player received theirs between May and July. For teams running a customer-support agent, the concrete change is this: because each agent call consumes much more compute than a simple chat, the CPU is back to being the bottleneck. NVIDIA promises 1.8x faster per core on agentic workloads. Source : https://blogs.nvidia.com/blog/vera-cpu-delivery/
3China releases a free AI model that beats GPT at cybersecurity -- fifth Chinese release in five days
Chinese lab Z.ai published the open weights of GLM-5.3 on August 29: a 753-billion-parameter model of which only 40 billion fire at any question. For non-technical readers: you can download it and run it at home, with no subscription. It tops the charts on two cybersecurity benchmarks (84.5% on CyberGym, record on Terminal Bench 3.0). The price is published: a few cents per million words read, less than a coffee for an entire conversation. This is the fifth major Chinese release in five days -- the engine no longer stops on weekends. Source : https://huggingface.co/zai-org/GLM-5.3
4NVIDIA publishes a small energy-efficient AI model and a smart router -- so your agents don't crash at the wrong moment
On August 11, NVIDIA released Nemotron 3.5 Lightning for free: a 31-billion-parameter model of which only 3.6 billion fire at any moment, designed to run on a single server. It ships with NeMo Switchyard, a router that picks the right tool for each incoming task (fast reply or deep analysis) so you do not pay the premium for everything. For a startup automating support, the cloud bill can drop by more than half according to early adopters. The model scores 51.6% on the SWE-bench test. Source : https://blogs.nvidia.com/blog/nemotron-lightning-switchyard-rtx-dgx/
5U.S. judge rules: 1.5 billion dollars for 7 million books used without permission -- new rule of the game for the whole AI sector
The Bartz v. Anthropic settlement received final court approval on July 20 and payments began on August 10. Anthropic had used about 7 million works downloaded without consent, working out to roughly 3,100 dollars per book on average, and will destroy the copies. For an entrepreneur who wants to use AI inside a product, the concrete change is this: the question "can I actually use this model in Europe?" now has an official price tag. Google, Meta, Midjourney and OpenAI face similar suits -- a market for training-data insurance just became solvable. Source : https://anthropiccopyrightsettlement.com/
📡 To watch
EU AI Act Article 50 has been active for 28 days
Since August 2, every professional chatbot must declare itself as an AI and every generated image must be watermarked. The Digital Omnibus (EU Regulation 2026/1744) postpones the heaviest obligations until December 2027, but the transparency rules are enforceable right now. For any small business selling into Europe, the time has come to add a legal notice to its chatbot. 78% of European organisations are still not ready. Source : https://digital-strategy.ec.europa.eu/
Linux security patch due before federal deadline
On August 28, the U.S. agency CISA added a Linux kernel exploitation by an AI agent (CVE-2026-53362) to its official catalogue of vulnerabilities that must be patched. Federal agencies must apply the patch before August 30. This is the first incident where an AI agent has triggered a federal patch requirement. Worth watching for any organisation exposing agents on standard hardware. Source : https://www.cisa.gov/known-exploited-vulnerabilities-catalog
The Chinese release cadence is not slowing down -- next wave expected this week
After five deliveries in five days (GLM-5.3-Flash, Qwen3.8 Flash, Ling 3.0 Fin, Tencent Hy4, full GLM-5.3), the DeepSeek, Qwen team, Moonshot and ByteDance labs are likely to align their models in the five days that follow. Average usage price has dropped below 1 euro per million words read for the most competitive releases -- a game changer for SMB integration in Europe. Source : https://huggingface.co/zai-org/GLM-5.3-Flash
First large-scale deployment of Vera + Rubin at Oracle Cloud
Oracle becomes the first major cloud provider to deploy the new Vera CPU + Rubin GPU pair at industrial scale. The stated goal is hundreds of thousands of Vera CPUs operational by the end of 2026. For teams hosting their agents in the cloud, the price per million words should fall in the coming months -- to be confirmed on independent benchmarks. Source : https://blogs.nvidia.com/blog/vera-cpu-delivery/
📊 Trend
Sunday, August 30, 2026: the weekend no longer stops the machine. Chinese labs have stepped over the Western calendar and shipped their fifth open-weight release in five days (full GLM-5.3, after Tencent Hy4 and Qwen3.8 Flash). Meanwhile NVIDIA is shipping at scale a CPU designed specifically for agents -- hardware is now an acknowledged bottleneck. Customer-facing companies are making their own pivot: Crescendo bets on replacing the entire customer-service stack with six specialised agents rather than a thin AI overlay. And the U.S. courts have ruled for the first time on the price of books used without permission to train a model: 1.5 billion dollars -- a benchmark that will weigh on every negotiation to come.