On Wednesday, the chairman of India's Securities and Exchange Board, Tuhin Kanta Pandey, told a business summit that the regulator is exploring a securities market AI model that could be built and governed in India as a form of digital public infrastructure. His stated reason was "concerns about the confidentiality and availability of frontier AI models." The regulator would rather not run market supervision on a model it cannot inspect and might lose.

The same speech said SEBI is putting in place IOSCO's supervisory toolkit for AI, and that using AI well needs reliable data, effective model oversight and clarity about who is accountable for decisions. Pandey warned that technology creates dependencies across institutions and service providers, which makes oversight of outsourced services essential. SEBI also plans to extend its cyber rules to the subsidiaries of exchanges and clearing houses, and to train security operations engineers to use AI on alerts. If a regulator is worried about losing access to its AI, a listed company should be too. (SEBI speech; full text; BusinessLine)

Governments are also lining up second sources. In a joint statement released Wednesday, after the Crown Prince of Abu Dhabi's visit to Ottawa on Monday and Tuesday, Prime Minister Mark Carney and Sheikh Khaled bin Mohamed bin Zayed Al Nahyan said Canada and the UAE will aim to deepen cooperation in AI, quantum computing, robotics, cloud computing, secure digital infrastructure and telecommunications.

The statement pairs Canadian research with Emirati money and infrastructure, and names work between Abu Dhabi's Technology Innovation Institute and Mila, the Quebec AI institute. Both sides also signed up to responsible technology governance. As more countries build their own compute and models, companies have more places to run workloads and more rules about where data may go. (Prime Minister of Canada)

Inside companies, dependency shows up on the bill. Amazon has published a revised role guideline for its software engineers. It says that measures like token usage, AI tool adoption and lines of code are deeply flawed because they are disconnected from results, and that treating them as targets drives up cost and complexity. "Tokenmaxxing is at odds with frugality," it says.

Amazon still tells engineers to learn and apply AI as soon as possible, and says the input measures track output well for now. It suggests managers look at the teams with the lowest AI adoption to understand why, and at the heaviest users to check whether delivery has risen to match. AI spend without an outcome attached turns into a habit with a growing invoice. (Amazon)

Someone also has to stand behind an AI that acts alone. Delaware is drafting a new legal form, the Artificial Intelligence Company, whose day-to-day affairs would be managed by an AI agent. A single member, a person or a company, would have to keep it adequately funded, and the agent would have to keep a log of its activities. The member would not be liable for its debts unless it left the company underfunded or used it for fraud or willful lawbreaking.

On Monday, House Oversight Chairman James Comer wrote to Delaware's governor asking for documents. He warned that the plan could shield AI companies and their investors from responsibility for harm, and said his inquiry could inform federal limits on state AI rules. The draft would run in a 30-month sandbox and would allow any lawful business except banking. If your counterparty might one day be an AI-run company, your contracts should already ask who carries the loss. (House Oversight; Ars Technica)

Already filed, if you want the run-up: The Model Card (US health regulators dropping required AI model cards), The Penalty (Japan weighing penalties in its AI law) and The Owner (Korea's bank supervisor saying the bank owns every AI mistake).

Also on the shelf: The Buffer (China telling employers to warn workers before AI takes their jobs), The Lab (Britain's new AI biosecurity lab, funded partly in OpenAI and Anthropic credits) and The Clerk (a US appeals court taking a case away from a judge over an AI-drafted order). Read today beside The Model Card. US health regulators stopped requiring AI vendors to publish model cards. Mumbai is asking whether it should build its own.

Governments have built spares before. The 1973 to 1974 oil embargo showed how exposed import-dependent economies were to a single source. In 1975 the United States started its Strategic Petroleum Reserve as part of an international energy agreement whose members committed to hold oil stocks of at least 90 days of net imports.

The first oil went into the reserve in 1977. Its underground caverns in Louisiana and Texas can hold 714 million barrels, and oil can start reaching the market 13 days after a presidential order. Nobody built it to use every week. It exists because going without a spare had cost more. (Strategic Petroleum Reserve)

Put these to management. Which of our critical processes depend on a single outside AI model, and what happens if that model is withdrawn, repriced or cut off for a week? Do we know where our data goes when it reaches that model?

Then the spend. What outcome is each major AI budget line tied to, and who checks it? If we deal with a company run mainly by AI agents, who in our contract carries the loss when it gets something wrong?

The Move

India's market regulator is weighing its own AI model because it cannot be sure the frontier models will stay confidential and available. Canada and the UAE are building capacity together, and Amazon is telling engineers that heavier AI use is not the same as better work.

Ask the chief technology officer for one page before the next board meeting: every critical process that runs on a single outside AI model, the fallback if it disappears for a week, and the outcome each AI budget line is meant to buy.

Price the spare before you need it.

Seven bullets. One move.