Sep 14, 2026
Anthropic has called for a global effort to pace AI’s progress — third-party evaluators with employee-level access, common safety standards among labs in democracies, and verification extended to authoritarian states building the same capabilities. It is a serious proposal, made by serious people. The reaction was immediate, and telling: as Bloomberg framed it, the heads of Anthropic and OpenAI are now in a clash with investors, with the White House, and with much of their own industry.
The proposal deserves support. But the conversation around it is missing two things, and they turn out to be halves of the same problem.
Writing on LinkedIn, tech advisor Rob Zelinka called the situation “a fundamental breakdown of governance.” Saanya Ojha of Bain Capital Ventures put the bind more precisely: “The people building the frontier increasingly see the risk; the race gives them very little room to act on it.”
It matters who said that second part. Ojha is an investor. The clearest statement of the incentive problem came from the side of the table that creates it.
That is the duality. The engineers closest to the capability are the ones most able to see what it can do, and the least able to slow it down — because the pace is not set in the lab. It is set by capital. Valuations are underwritten on capability curves. Rounds are priced on what the next model can do, not on whether the last one can be audited. A firm that unilaterally decelerates does not get graded on its restraint; it gets repriced. Asking engineers to solve that is asking the wrong people, on the wrong floor of the building.
Which is why a voluntary pause is fragile in a way its critics understate and its authors clearly understand. It asks a group of competitors to hold a line that the capital behind each of them is actively paying to cross.
The second omission is that this debate is happening entirely upstream of the harm.
A pause among the labs does nothing for the hospital, the insurer, or the investor already running a model that shipped eighteen months ago. That is where AI touches people today — in an eligibility determination, a care authorization, an investment decision. The binding constraint here has never been model capability — it is the institutional capacity to evaluate, govern, and safely operate what has already been deployed.
That capacity is not improved by a slower frontier. A hospital does not become better at auditing its vendor’s model because the next generation of that model arrives in twenty months instead of twelve. The deployed system is already operating on the front lines. The governance debt is already on the books.
It is tempting to treat these as separate arguments — one about frontier safety, one about enterprise adoption. They are not.
The same pressure that will not let a lab slow down will not let an enterprise stop shipping. A hospital system that pauses its AI program to build evaluation capability watches competitors announce deployments. A CIO who insists on auditability before rollout is, in that quarter, the person slowing things down. The mechanism is identical: speed is legible to the people holding the purse, and governance is not.
One force, two failure surfaces. At the frontier it produces unmanaged risk. Downstream it produces unusable systems. The second surface has already been measured. The first… Welcome to the party we should have started eighteen months ago.
95% of enterprise generative-AI pilots deliver no measurable return, against $30–40 billion invested. That is MIT NANDA’s figure, and it is worth being precise about what kind of statistic it is: not a safety statistic, a capital-efficiency statistic. Tens of billions of dollars were deployed and produced nothing.
Those pilots did not fail because the models were not clever enough. They failed on governance, integration, and security — the parts nobody demos. Which means the market has already run the experiment on whether you can skip that work and still get paid. The answer came back at 95%.
So the framing of governance as a brake on returns has it exactly backwards. In the only large-scale test anyone has run, the absence of governance is where the returns went. That figure is the receipt for the downstream failure and the argument to the people funding the upstream one, in the same number.
We reached these conclusions about eighteen months ago and rebuilt Pivotal around them.
The bet was not that models would stop improving. It was the opposite: that capability would keep outrunning any institution’s ability to deploy it responsibly, and that the scarce thing would therefore not be intelligence but provable intelligence. Systems whose decisions can be reconstructed, audited, explained, and attributed have always been the operational requirement. Not because a regulator demanded it, but because in high-stakes domains a decision nobody can explain is a decision nobody can defend — and increasingly, a decision nobody will fund.
That is an architectural position, not a philosophical one. It determines what you build, and in what order.
Healist.AI is what the position looks like when you actually build it.
It is a HIPAA-compliant health-intelligence platform, architected with governance, security, and AI safety as primitives from the first commit rather than as a hardening pass before launch. Its Cloud MCP service exposes 270+ governed, consent-gated endpoints — every one permissioned, logged, and revocable, so a third-party agent reasoning over someone’s health data does so on terms that person set. Guidance is computed against a complete, persistent health record rather than a point-in-time snapshot, because an answer derived from a fraction of a history is not a careful answer. And when the system is uncertain, Healist.AI is built to fail closed and route to a human clinician rather than fail open and sound confident.
None of that is a feature list. It is the cost of operating where being wrong is not recoverable — and it is precisely the work a slowdown at the frontier cannot do on anyone’s behalf.
The labs are asking for time because they can see a risk the race will not let them address. The organizations deploying their models are running the same problem one layer down: they can see the governance gap, and the pressure to ship will not let them close it. Both are being paced by the same force, and it is not a technical one.
A pause helps the first group. It does nothing for the second unless the interval is deliberately spent building the capacity to evaluate, govern, and operate what is already in production.
Buy the time. Then spend it downstream, where the consequences land. We would rather be part of the group that spent it well — and that is the whole reason we build the way we do.