Edition · Thu 15 Oct 2026
The Next AI Race Is for Control, Not Just Compute
Opinion: This week's news from the Financial Stability Board, Figma and Siemens points to the same lesson. The winners in AI will be those who can govern what their systems do, not only those with the most chips.
The Zevin Intelligence Journal
Published
Three announcements, one theme
For two years the AI race has been told as a story about compute: who can buy the most chips, build the largest data centers and secure the most electricity. That story is not wrong. But three announcements this week suggest a second contest is now under way, over who can control what AI systems actually do.
On Thursday the Financial Stability Board, which coordinates financial authorities from 24 jurisdictions, published 12 sound practices for AI in finance. Its strongest change from the draft concerns AI agents. Firms using them in higher-risk tasks need more intensive human oversight and clear, risk-based triggers for when an agent must ask a human for approval, the FSB said, according to Reuters. It also dropped the idea of treating AI as "synthetic employees", warning that agents "may interpret pre-defined goals in unexpected ways."
The same day, Figma launched guidelines that let companies write down their design rules so its AI agent follows them, along with a format for handing shared context to coding agents. And on Wednesday, Siemens led an investment in Skeleton Technologies, whose systems are meant to smooth the millisecond power swings of GPU clusters, swings that Skeleton says can create peaks and harmonics.
Siemens framed its own move in similar terms. "The race to scale AI is increasingly becoming a race for power," said Peter Koerte, a member of its managing board, in announcing the deal, which Siemens says gives it exclusive rights to sell Skeleton's products to data centers. The point is not only to supply more electricity, but to make the supply behave predictably when thousands of chips switch load at once.
Why control is the scarce resource
These are different industries, but each addresses the same gap. Capability has outrun governance. A regulator is asking banks to define when a machine must stop and ask. A software company is selling a way to tell agents what "good" looks like. An industrial group is investing in hardware that keeps AI workloads from destabilising the power that feeds them.
In each case, the value lies less in making AI more powerful than in making it predictable. That matters commercially. An agent that produces work a team must redo is a cost, not a saving. A bank that cannot explain or stop an automated decision faces supervisory risk. A data center whose power draw is unstable may face extra engineering costs. None of these problems is solved by adding more compute.
The counterargument
The obvious objection is that controls slow innovation. Approval triggers add friction; written rules can become bureaucracy; caution can let less careful competitors move first. The FSB itself frames its practices as a tailored toolkit rather than rigid rules, which acknowledges that one-size-fits-all controls would be costly.
That objection deserves weight. Controls that are applied indiscriminately would waste the very productivity gains AI promises. But the guidance this week is proportional by design: more oversight where risk and autonomy are high, less where they are low. Proportional control is not the enemy of speed. It is what allows organisations to hand more work to AI without betting the firm on systems that, by the regulator's own account, do not reason as people do.
It is also worth being honest about what is not yet known. Figma has not published evidence that its guidelines reduce rework. Skeleton's claims of energy savings and GPU performance gains are its own. The FSB's practices are not binding. The trend is clear; its payoff is still to be measured.
What leaders should do
Boards adopting AI should ask three questions that have little to do with chip budgets. Where must a human approve before an agent acts? Are our rules written down in a form the systems can actually use? Is the infrastructure underneath stable enough to scale? Firms that can answer those questions will be able to deploy AI more widely and with more confidence. That, more than the next order of GPUs, may decide who leads the next phase.