The top things worth knowing about in AI today.
Twenty of the world's most valuable chip companies shed US$1.3 trillion in market value between Friday's close and Wednesday, with Nvidia alone down $238 billion and SK Hynix falling almost 10% despite posting record quarterly profit. Analysts described it as a repricing of expectations rather than a fall in AI demand. Nothing has changed about what the models can do, but the cost of the compute behind them is now a live question for anyone budgeting an AI project.
Read more →Nvidia is in talks to guarantee roughly US$250 billion of lease and construction financing so OpenAI can occupy a 10-gigawatt campus SoftBank is building on a former uranium-enrichment site south of Columbus, according to the Wall Street Journal. A separate discussion covers up to $350 billion for the chips themselves, and the full site could exceed $500 billion. OpenAI is not yet profitable and cannot borrow at investment grade on its own, which is why its chip supplier is being asked to stand behind the debt.
Read more →The Beijing lab behind Kimi K3 closed a US$3.5 billion round at a $35 billion valuation, well above the $1-2 billion it first sought, with China's state AI investment fund among the lead backers. Annual recurring revenue reached $300 million in June, up from $100 million in March, and the company briefly paused new subscriptions after K3 shipped because it ran short of compute. Open-weight Chinese models are now a commercial force rather than a benchmark curiosity.
Read more →OpenAI has opened limited availability of Presence, sold as a project rather than a product: its engineers build one workflow agent at a time - billing disputes, insurance claims, IT requests - with the customer setting approval and escalation rules. Early users include BBVA for voice banking support and SoftBank for Japanese-language conversations, and OpenAI says its own support line now resolves 75% of inbound issues without a human. Access is limited by delivery capacity rather than demand, which says a lot about where the real constraint on enterprise AI sits.
Read more →Eligible Microsoft 365 commercial tenants were switched to GPT-5.6 across Word, Excel, PowerPoint and Copilot Chat from 24 July unless an admin opted out, and Claude Sonnet 5 is now available in Copilot Cowork and PowerPoint. Copilot Chat in Outlook also expanded from reading a single thread to reasoning across a whole mailbox and calendar, with no paid Copilot licence required. If you administer Microsoft 365, the model doing your users' work may have changed without them noticing.
Read more →Article 50 of the EU AI Act applies from 2 August: people must be told when they are talking to a chatbot, synthetic output must be machine-readable as AI-generated, and deepfakes must be labelled. Penalties reach 15 million euros or 3% of global turnover. Unlike the high-risk provisions, which slipped to December 2027, this deadline was not extended - so anything you ship into the EU with generative output needs its disclosures in place this week.
Read more →Cyera, valued at US$12 billion after a $600 million round, signed a letter of intent to acquire non-human identity firm Oasis Security for about $1 billion, mostly in cash. It is Cyera's third acquisition this year. The logic is simple: every agent you deploy is a new identity holding credentials and permissions, and somebody has to govern them.
Read more →OpenAI analysed more than 800,000 US work-related messages and found 43.5% of occupation-specific requests covered tasks normally belonging to a different role - marketers debugging websites, salespeople querying datasets, owners reviewing their own contracts. It calls the pattern task crossover, and it shows up most in customer experience, design, HR, legal and marketing. Role boundaries in most organisations are already looser than the job descriptions admit.
Read more →A short read every week — the few things worth your time, and nothing that isn’t.