This weekend Dario Amodei published “We Must Pace the Frontier.” Within hours, Sam Altman agreed. Musk backed it. Hassabis called the direction correct. Whatever gets formalized this week, the signal is unambiguous: the people building the most capable AI systems in the world just told the market that capability has outrun verification.
I run infrastructure for regulated finance, and I’ve spent the past 18 months rewiring Blockdaemon itself to be AI-first. So I read the essay less as an AI story and more as a finance story. Here’s why.
The labs didn’t discover a new problem. They discovered the problem regulated finance has been organized around for a century: past a certain level of stakes, trust cannot be asserted. It has to be proven, continuously, by someone other than the party making the claim. That’s why clearinghouses exist. That’s why DTCC exists. The frontier labs just reinvented the audit function.
Look at the first concrete commitment. Not a compute cap. Embedded third-party evaluators with employee-level access, verifying that labs actually do what they say. Humans, checking. That’s the right move for model development. It does not scale to what’s coming next: agents acting at machine speed against real systems, and increasingly against real money. Treasury, payments, settlement, custody. No bank will let an autonomous agent touch value transfer because an evaluator reviewed the lab’s safety culture last quarter. They will require proof per action.
That layer is what we build. Every agent action running through Blockdaemon’s control plane gets evaluated against institution-defined policy, authorized or escalated, enforced through MPC threshold signing, and recorded as a signed, exportable attestation: which agent acted, under which policy, at what time. Not a log entry a vendor promises is accurate. A cryptographic artifact the institution itself can hand to a regulator. It sits on the same infrastructure that supports 400+ institutional customers and over $110B in digital assets, with zero key compromise events in six years. DTCC, NYSE, JPM, GS, Citi, Fidelity, RBC, SocGen and LSEG don’t work with us because we move fast. They work with us because everything we do is verifiable.
So my position on the slowdown: it isn’t a retreat. It’s the market repricing trust. For a decade the AI race rewarded velocity. From here, the binding constraint on adoption in regulated finance is governance, not capability. The models are already good enough for most institutional workflows. What’s missing is the execution layer that lets a risk committee, an auditor, and eventually an examiner say yes.
I’ve argued before that if AI hasn’t changed the workflow, it has changed nothing. The corollary now: if an agent’s action can’t be proven, it won’t be permitted. The labs are pacing the frontier so safety can catch up. Our job is to make sure that when institutions are ready to deploy, the trust layer is already there. It is.
