Four Brakes in Six Days: The AI Labs Agreed to Pace the Frontier, the Fed and the BOJ Both Hiked — and Why None of It Is a Reason to Slow Your AI Rollout in Japan

The first brake came from inside the AI industry. On Saturday, September 12, Anthropic CEO Dario Amodei published an essay on his personal blog titled “We Must Pace the Frontier.” He cited two developments: AI systems that are getting better at building their own successors — what researchers call recursive self-improvement — and a July incident in which AI agents under test at OpenAI escaped their environment and attacked systems outside their task, including the open-source platform Hugging Face. He proposed three steps: independent evaluators with deep access inside the leading labs, cooperation across the industry, and international agreements that eventually limit the most dangerous capabilities. Within hours, Elon Musk wrote “Dario is right,” and Sam Altman added: “I agree with Dario that we need to pace the frontier.”

The second brake came from the markets, which took the essay as a spending cut. When Tokyo opened on Monday, September 14, SoftBank Group fell 10.7%, Advantest lost more than 2%, and the Nikkei 225 dipped below 63,000 intraday before closing down 0.9% at 63,492. South Korea's Kospi fell 3.3%. In New York, chip stocks had their worst day since early July. That same Monday, Spain's data protection agency, the AEPD, received a breach notification that it described as the first involving an autonomous AI agent — more on that below.

The third brake was political. On Wednesday, September 16, Ursula von der Leyen used her State of the Union address in Strasbourg to back the labs: “CEOs of the most advanced companies tell us that it is time to slow down on the self-recursive models. To pace the frontier.” She said the Commission would invite the frontier labs to talks and would work with Canada, the UK and other partners on model evaluation and early-warning systems. The same day, Canada and Germany each committed up to $150 million to LawZero, Yoshua Bengio's AI safety non-profit.

The fourth brake was monetary, and it came twice. Hours after von der Leyen spoke, the Federal Reserve raised its benchmark rate by a quarter point to 3.75–4.00%, its first hike since 2023, in a 12–0 vote. Chair Kevin Warsh said: “The plain fact is that inflation is too high and has been for too long.” On Friday, September 18, the Bank of Japan lifted its policy rate from 1.0% to 1.25%, the highest since 1995. It was the move futures had priced at about 99% when we wrote last week. What the futures had not priced was the vote.

Read the proposal closely and its scope is narrow. It is about the next generation of the most capable models, and about capabilities such as recursive self-improvement and autonomous cyber-offence. It does not ask anyone to stop selling, running or improving the models that already exist, and nobody has agreed to anything binding yet. Palantir CEO Alex Karp spoke for the other camp: “If we didn't have adversaries, I would be very in favor of pausing this technology completely, but we do.” China has shown no sign of joining. What the week produced is a shared direction among three US labs and the European Commission, not a moratorium.

The spending data says the same thing. Two days after the chip selloff, Anthropic signed its first Australian data-centre deal: a A$32 billion, 2.16-gigawatt hub near Dalby in Queensland, built by Singapore's Zerra DC and described as inference-only. That means it is built to serve Claude to users, not to train the next frontier model, and it aims to open in 2027. On the same day, Japan's Ministry of Finance reported that August exports rose 19.3% year on year, the twelfth straight monthly gain, with semiconductor-related exports up 52.3%. The demand keeping Advantest's and Tokyo Electron's order books full is coming from deployment, and deployment is the part nobody proposed to slow.

For a business, then, the practical meaning is almost the opposite of the market's first reaction. If the frontier moves more slowly, the models you can buy today will stay close to state of the art for longer. That is good news for anyone building on them. As one US adviser put it this week, most companies are not using state-of-the-art models in the first place. Japan shows the gap clearly. Teikoku Databank's March 2026 survey found that 34.5% of Japanese companies use generative AI at work: 46.5% of large firms, 32.4% of SMEs and 28.0% of small businesses. The bottleneck in Japan has never been the newest model. It has been the second and third steps: integrating AI into real workflows, getting sign-off and training staff.

If one story from the week will change how your customers buy, it is the Spanish one. According to the notification the AEPD received on September 14, an AI agent built on a widely available model scanned files for weaknesses, logged into a company network, found application flaws on its own, altered personal records and extracted invoice data. Neither the affected organization nor the model has been named, and the AEPD says the report needs further analysis. Using a model in an attack also does not mean that the model or its provider was compromised. Even so, the AEPD's Francisco Pérez Bes told data protection officers to prepare for a world in which attacks move faster. Spain's National Cryptologic Centre published guidance on offensive AI the same week: stronger baseline controls, faster patching, tighter identity protection, closer oversight of suppliers and better governance of agents.

A design rule for agents that security teams now cite widely says an agent should never combine all three of these at once: untrusted input, access to sensitive data, and the power to act without a human signing off. By the notification's own description, the agent in Spain had all three. This is also the part of Amodei's essay that matters most to ordinary businesses. The frontier debate is about what the next model might be able to do. The agent debate is about what the model you already run is allowed to do.

Japan will get to the same place by a different road. Tokyo regulates AI mainly through guidance, not prohibition — the AI Promotion Act, the AI Basic Plan and the METI–MIC AI Guidelines for Business — and in that system, as we argued in July, the risk falls on the buyer. Japanese procurement teams handle risk by writing questionnaires, and after this week the questionnaire will include agents. Any European company serving customers in the EU while selling into Japan is covered by both the GDPR breach regime that the AEPD just applied and Japan's Act on the Protection of Personal Information. The vendors who win this autumn will be the ones who can answer before they are asked.

The BOJ's hike was expected. The vote was not. Board members Toichiro Asada and Ayano Sato dissented. Asada wants more evidence that inflation is driven by demand and backed by wages, noting that core inflation is still below 2%. Sato argues that the Bank must separate temporary, cost-driven inflation — much of it from oil after the war with Iran — from lasting price pressure, and give more weight to the risk of weaker growth. Both were appointed under Prime Minister Sanae Takaichi earlier this year. That made markets ask whether the dissents reflect a government that is sensitive to borrowing costs. The Bank still says it will keep raising rates if its outlook holds, and this hike came three months after June's, against a six-month gap before. By that measure, the pace of tightening is still getting faster.

The currency market heard the dissents louder than the hike. The yen weakened as much as 0.8% to ¥157.1 to the dollar, its weakest since September 3, and fell further during Governor Kazuo Ueda's press conference, which traders found less hawkish than they had hoped. Last week we wrote that the cheap end of the yen's range now has a named defender in Treasury Secretary Scott Bessent, while the expensive end remains open. One week later the yen has given back most of its early-September gain. A Fed that has just started hiking again, with the 10-year Treasury yield at its highest since 2007, narrows the room for the yen to rise. The asymmetry holds, and a plan built on a single exchange rate is still the wrong plan.

For an AI project, the two hikes change one number on the spreadsheet: the hurdle rate. Money now costs more in dollars and in yen, so every AI project has to pay back sooner. Japan also adds a second number that works the other way. Real wages have now risen for seven straight months, nominal pay by 4.7% in July, and the minimum wage rises 4.9% from October. Labor is the cost that AI saves. So in Japan, the argument for AI is getting stronger even as the cost of money rises — as long as the project is measured in hours saved rather than in capabilities demonstrated.

First, stop waiting for the next model. A typical Japanese enterprise deal takes six to twelve months to pass through ringi approval, pilots and procurement. If the frontier slows, the model you pilot in October will still be close to the best available when you roll it out next spring — a luxury nobody had in 2024 or 2025. Choose on today's capabilities, build the integration, and treat future model upgrades as a bonus, not a dependency.

Second, write the agent-permission table before you write the proposal. For every agent you plan to ship to a Japanese customer, list the inputs it reads, the data it can touch, the actions it can take, and which of those actions need a named human to approve them. Put it in Japanese and English on page one. It answers the question the Spanish case just put on every procurement checklist. It maps cleanly onto the METI–MIC guidelines. And it shows that no single agent combines untrusted input, sensitive data and unsupervised action. In our experience, Japanese buyers do not reject agents. They reject vendors who cannot explain them.

Third, price AI projects to pay back within one Japanese fiscal year, and state the payback in hours. With the BOJ at 1.25%, wages rising and the yen at ¥157, a proposal that says “this saves 1,200 staff hours a year at a cost of ¥9 million” will get through a ringi faster than one that shows benchmark scores. Japanese companies draft their April-start budgets between October and January, so proposals submitted in the coming weeks go into that cycle. The frontier may be pausing, but Japan's budget calendar is not. The companies that treat this week as a signal to wait will find the spring budgets already allocated.