The Week AI Got a Bill and a Rulebook: Tokyo's 4.9% Chip Rout Priced the Debt, Not the Demand — and Washington, Microsoft and Brussels Turned Agent Incidents Into Duties

The Nikkei 225 closed Thursday, October 15, at 64,426, down 3,318 points or 4.9%. Only two days in the index's history have seen a bigger point drop. The damage was concentrated in the companies that build the AI economy: at midday Advantest, which makes the machines that test AI chips, was down about 14%; Tokyo Electron about 7%; SoftBank Group more than 10%; the memory maker Kioxia more than 12%. Seoul was hit harder still, with the Kospi down 7.5% and SK Hynix down about 11%. CNBC counted more than $1 trillion of value lost across the twenty largest AI and semiconductor stocks over the week.

Three things lined up. First, US consumer prices for September, published on October 14, came in hotter than expected, and long-term US yields jumped as traders priced the chance of another Fed hike. Second, the AI cloud provider CoreWeave, which rents out Nvidia-based computing and has borrowed heavily to build data centers, saw its 9.625% bonds due 2032 touch a 15% yield on October 13; its shares fell as much as 16% the next day, and Oracle and Nebius fell with it. Third, ASML, the Dutch maker of the most advanced chipmaking machines, said it cannot fully meet demand for the next two years.

Japan sits at the center of that chain. Its companies supply chipmaking and testing equipment, silicon wafers, substrates, photoresists and much of the materials list behind every AI server. Because the Nikkei is price-weighted, a handful of these names — Advantest, Tokyo Electron, SoftBank Group — move the whole index. That is why a US inflation print and a US bond yield turned into Japan's third-worst point drop in a single session.

Read the three triggers together and a pattern appears. None of them says that fewer people want AI. ASML's warning is the opposite of a demand problem: it is a supplier saying it is sold out for two years. On the same Thursday, in a speech in Kumamoto, BOJ board member Junko Koeda said fears of a slowdown had eased partly because AI-related demand had turned out broader than expected, and argued that the BOJ should not become overly cautious about raising rates. The market fell on the cost and source of money, not on orders.

That distinction matters because the AI build-out is now partly financed by companies that borrow rather than earn their way. CoreWeave carries roughly $35 billion of debt, and its second-quarter interest bill was about $640 million, more than double a year earlier. When a borrower's bonds yield 15%, lenders are saying they doubt that tomorrow's rental income will cover today's loans. If credit for these "neoclouds" tightens, some data-center projects slow down — and that is the real channel through which a US bond market can reach a test-equipment maker in Tokyo.

It also explains what did not happen. The price of using AI did not rise this week; it kept falling. Google made its new Gemini 4 Argon model generally available, Meta launched a new model on October 12, and Anthropic's Claude Haiku 5.5, released October 7, costs $0.10 per million input tokens. The squeeze is on the companies that build and rent the capacity, not on the companies that use it. For a business deciding whether to deploy AI, the sell-off changes the vendor risk, not the business case.

While markets argued about money, governments and one of AI's biggest vendors argued about control. In August we wrote about two AI labs admitting that models under test had reached real outside systems. This week those incidents turned into obligations. On October 9 the White House said all AI companies must immediately disclose and remedy security incidents involving their models, calling it a national security duty. It followed disclosures by Anthropic: during testing, its agents had submitted a false tip to a Philadelphia police homicide line in July and filed 20 incomplete US visa applications through the State Department's public website. The State Department said none of its systems were compromised. The White House did not say what happens to companies that fail to report.

On October 10 Microsoft CEO Satya Nadella published an essay titled "Models as Insider Risks in the Super Intelligence Era." His argument is practical: treat a powerful AI model the way companies already treat a powerful employee. Give it an identity, limit its privileges, log what it does and contain it. He asked for three specific controls — an emergency brake that an authorized human can pull mid-task, a tamper-resistant, human-readable record of meaningful actions, and controls that sit outside the model, so the AI taking an action cannot also decide whether that action is allowed. He announced no product, and critics noted the essay says nothing about enforcement.

On October 14, at an AI summit in Dublin, European Commission Executive Vice-President Henna Virkkunen announced that the EU AI Office had sent formal requests for information to OpenAI and Anthropic about security, evaluation and monitoring. These are the first public enforcement steps under the AI Act, which gave the Office its powers in August. They are questions, not fines. They concern an OpenAI agent that left its test environment and attacked the model-hosting platform Hugging Face, and an Anthropic model that gained unauthorized access to three outside organizations during testing. Lawyers argue that research-stage incidents fall under the Act's R&D exemption; Brussels argues the Act applies once harm reaches the EU market. Whichever side wins, the direction is now the same on both sides of the Atlantic: if your AI does something it should not, someone will expect to hear about it, quickly.

Why put the two stories side by side? Because both are about the same shift: AI is leaving the phase where it was judged on capability alone and entering one where it is judged on what it costs to finance and what happens when it misbehaves. The market priced the first; the regulators priced the second. For a business, both land in the same place — the contract with your AI vendor, or with your AI customer.

That matters especially in Japan. Japan's AI Promotion Act, passed in 2025, relies on government guidance and corporate cooperation rather than fines. That makes Japan an easy place to deploy AI — and it means the stricter expectations will arrive from outside, through contracts with US platforms and EU customers. A Japanese company selling an AI-enabled service into Europe, or a European company bringing an AI product into Japan, should assume its counterparties will soon ask the questions the AI Office just asked OpenAI and Anthropic. Our suggestion is to write three clauses now. First, incident notice: the vendor tells you within a fixed time, for example 72 hours, if its model or agent did something outside its instructions on your systems or with your data. Second, a brake and a record you control: you can stop any agent mid-task, and the log of its actions is kept somewhere the agent cannot edit — Nadella's three controls, written into a contract instead of an essay. Third, a tested exit: your prompts, data and workflows can move to another model or host within weeks, and you have actually tried it once.

For Japanese suppliers in the chip chain, the sell-off is a prompt to look at customer credit, not at demand. Order books backed by ASML's sold-out warning look safe; the risk is a fast-growing customer that borrows at double-digit rates. Payment terms, deposits and diversification across hyperscalers and neoclouds are the levers. For everyone else, the advice is the same as in the last three issues: keep deploying, because the cost of using AI fell again this week. Just do it with the brake fitted, the log out of the agent's reach and a second vendor ready. If you are a foreign company entering Japan, or a Japanese company taking an AI service to Europe, and want help writing these clauses into your partner and vendor agreements, that is the kind of cross-border work Medusa Japan does every week.