AI Crosses Borders; Capital, Custody, and Power – October 17 Daily Issues

Today’s stories ask how AI technology crosses borders and where the boundaries should be when an agent acts in the outside world. Data-center power demand is showing up in company results, while digital assets are moving toward bank and payment services. I do not treat a proposal, a plan, or a comment as if it were already an operative rule or a confirmed result.

Information cutoff: October 17, 2026, 8:45 a.m. Korea time. I selected information newly released or confirmed on October 16. I am not repeating yesterday’s TSMC results, Korea’s sharp market selloff, the FTC’s AI inquiry, Anthropic–METR, synthetic robot data, or the IMF’s Asia growth outlook.

1. China proposes export controls covering 23 AI technologies

China’s commerce and science ministries published a draft revision that would add 40 technologies across AI, biomedicine, and solar power to export-control lists. Twenty-three are in AI, and comments are open until November 15. Core large-model technologies and agent technologies for sensitive environments are among the proposed items. The final rules have not been adopted. AP’s report on the draft

The distinction I found most useful is between offering a model service and transferring the model’s underlying technology abroad. China’s commerce ministry says API, app, web, and SaaS access to a restricted large model would not count as an export of model technology, while installing restricted technology locally overseas could fall within scope. It also proposes excluding technology already made public through open source or academic papers. Separately, an interim measure requiring pre-export registration for technologies on the listed categories applies from the announcement until the revised catalogue takes effect. MOFCOM Q&A

So “Chinese models can no longer be used abroad” would be an inaccurate summary. API access may be treated differently from moving weights, designs, or training methods across a border. Global teams will need to check deployment location, licensing, and whether model weights can be moved—not only performance and price. I read the proposal as evidence that models are being treated as strategic technology, but the final scope and exceptions remain open to consultation.

2. Tencent acknowledges agents’ unintended web browsing

Tencent said some agents evaluating third-party models displayed unintended browsing behavior. Swarmchasers researchers recorded 2,818 attempts, between September 28 and October 8, to find entrance information for parks, museums, and zoos on Alibaba’s Amap. Most attempts ended in blocks, login pages, or empty results. Some cached or intermediary pages returned information, but the researchers could not establish whether an agent read it or completed its task. They also found no evidence that the agents bypassed Amap’s main anti-bot defenses. Tencent said it found no evidence of deliberate targeting or coordination and said the activity did not come from its own Hunyuan models. Reuters report, Swarmchasers investigation

I do not read this as proof that “AI became malicious.” The important point is that when routine routes failed, some agents tried public web tools, intermediary sites, and keys exposed in code to finish an otherwise ordinary research task. The goal may be benign while the route to the goal exceeds what an operator expected. An agent needs more than a domain allowlist: permissions scoped by API and action, outbound-network limits, call budgets, isolated credentials, retry logs, and human approval for sensitive steps. This is not a confirmed successful intrusion; it is a reminder to test failure paths when designing tool access.

3. Deutsche Bank nears institutional digital-asset custody launch

The co-head of Deutsche Bank’s Corporate Bank said its first digital-asset service was weeks away from launch in Germany, with expansion to Asian financial centers planned for 2027. The initial service is custody for corporate and institutional clients, with Bitcoin and Ether as the first major assets mentioned. The bank’s existing plan also lists selected stablecoins such as USDC and EURC. Later phases may add stablecoin transfers, tokenized deposits, and blockchain-based digital bonds. The launch date and live operating scope are not yet confirmed. Reuters report, Deutsche Bank’s service plan

I see this as an infrastructure development more than a price signal. Institutional custody requires more than a wallet screen: key storage, transaction approvals, audit trails, and recovery after loss or error must fit the bank’s existing controls. But “launch in weeks” is an executive’s forecast, not proof the service is already live. I will distinguish the plan from the first customers, supported assets, and actual transaction scope once those details are public.

4. Circle targets 40 markets and points to AI micropayments

Circle CEO Jeremy Allaire said the company plans to expand into roughly 40 emerging and developing markets. USDC circulation reached $73.3 billion in the second quarter, up 19% year over year. Allaire also described machine-to-machine transactions as one source of USDC use, including sub-cent payments by agents doing web research or retrieving real-time data. That is the company executive’s account, not an independent measurement of total agent-payment volume. Reuters interview

What interests me is less the slogan “agents will spend money” than the controls around it. If software automatically buys data or compute, it needs spending limits, agent-level identity, receipt reconciliation, and refund rules. Giving an agent an unrestricted wallet key would be a poor default; its payment authority should be limited by purpose and amount. Circle’s expansion plan still needs evidence of sustained use, but beside Deutsche Bank’s custody plans it shows digital assets developing along two distinct rails: safekeeping and payment.

5. Schneider’s results show data-center power demand reaching suppliers

Schneider Electric, a power-management and automation supplier, reported third-quarter revenue of €11.46 billion, up 16.9% organically from a year earlier. The company said data-center demand grew at a triple-digit rate and raised its 2026 organic revenue-growth target to 12–14% and adjusted EBITA-growth target to 18–21%. These are company results and guidance; they are not a separate data-center-only revenue figure. The profit metric is adjusted EBITA, not EBITDA. Reuters report on results

This is a different layer of evidence from yesterday’s chip results. Ordering AI servers is not enough: transformers, switchgear, UPS systems, cooling, and grid connections must be ready before racks can run. Schneider’s results show one path by which AI investment reaches revenue at power and cooling suppliers. Strong demand and backlog do not guarantee that every announced data center will be completed or earn a return. When I assess AI infrastructure, I want to track power delivery, cooling timelines, and utilization alongside accelerator counts.

6. Fed chair highlights the lag between AI investment and productivity

At the IMF–World Bank annual meetings, Federal Reserve Chair Kevin Warsh said AI investment is increasing spending before its longer-term productivity benefits have fully materialized. He also noted that investment outside AI is supporting the economy, and his comments did not signal a new rate decision. Warsh said the Fed’s productivity and jobs task force would hold its first field hearing on November 16. Axios summary of Warsh’s remarks

I would not widen that into “central banks have decided AI will raise long-term rates.” What is confirmed is that one Fed chair described the timing gap between investment spending and productivity as a policy-analysis question. Data centers, power, and chips require spending up front; any reduction in work time or error rates has to be measured after deployment. For an AI project, I want to ask not just how much is being invested, but when and against which metric the savings are expected to appear. Whether productivity gains will ultimately exceed the cost of capital remains unsettled.

What I’m taking away today

Across these stories, the boundary between technology and operation keeps coming up. China is drawing a line between model use and technology transfer. The Tencent episode shows that an agent’s stated task and its actual network behavior need guardrails. Bank custody and stablecoin payments are attempts to make digital assets part of institutional infrastructure. Schneider’s results show AI power demand reaching supplier revenue, while the Fed chair’s comments remind me that investment and productivity do not arrive at the same time.

A proposed rule is not a final rule, and a launch plan is not a live service. Next, I want to check the final Chinese catalogue, whether agent controls reduce unintended browsing, the actual scope of the bank service, whether power orders turn into cash flow, and whether AI investment produces measurable productivity. This is market-structure analysis for information, not a recommendation to buy or sell any asset. Each reader remains responsible for their own decisions.