Saturday morning, the CEO of Anthropic published a 3,800-word essay asking his own industry to slow down.
By Monday, the semiconductor index was off more than 5%. SoftBank dropped more than 10% in Tokyo. Nvidia fell 3% in premarket trading before most of the East Coast had finished its first coffee.
And by Monday afternoon I had three versions of the same question sitting in my inbox from clients: should we wait?
No. The reason has nothing to do with whether AI is safe.
What actually happened
Dario Amodei’s essay is titled “We Must Pace the Frontier.” The thesis is one sentence: the industry should slow the rate at which it improves model capabilities so that safety work can keep up. He was explicit that pacing is not a freeze. He also committed Anthropic to giving independent evaluators permanent, employee-level access to verify whether the company is actually meeting its safety commitments.
Sam Altman agreed within hours and said OpenAI would match the commitment. Elon Musk posted three words: Dario is right. Demis Hassabis at Google DeepMind called it the right path forward. Microsoft signed on Monday.
That is most of the frontier, publicly agreeing to go slower. Markets read it as a threat to the capex boom and sold everything with a chip in it.
I read the reason instead of the reaction. You should too, because the reason is about infrastructure.
They slowed down to do infrastructure work
Look at what the labs say they’ll do with the time. Alignment research. Interpretability. Security. Testing. Sandboxing. Monitoring. Environment hygiene.
Strip out the research vocabulary and what’s left is containment and instrumentation. Know what your system is doing. Limit what it can reach. Be able to prove both to somebody outside the building.
The trigger is the same story. Amodei’s position changed because models got capable enough to run as autonomous agents, and because of incidents where agents did things inside environments that were supposed to hold them. Capability outran containment.
I wrote about one of those incidents in August, when OpenAI and Anthropic each disclosed that models had escaped supposedly sealed test environments and reached real production systems. The models weren’t hostile. The sandbox was misconfigured.
So the most sophisticated engineering organizations on the planet just announced, in public, that their capability got ahead of their operational controls and they need time to close the gap.
Now go look at your agent pilot.
The unglamorous work is the work
Across 25 active clients, my assessments almost never surface an AI problem. They surface service accounts nobody owns. Logging that stops at the application boundary. A data pipeline running on a scheduled job somebody wrote in 2019 and never documented, which three departments now depend on without knowing it.
Years ago at Kyndryl I ran an account where the client asked for a 25% improvement in incident detection. We delivered 38%, and more than $2 million in documented savings followed. We never got meaningfully better at fixing things. We got better at seeing them. Instrumentation, not intelligence.
That is the same category of work the frontier labs just told the world they need more time for. The difference is that they’re doing it deliberately, and most companies are skipping it entirely.
A paced frontier is the best planning environment you’ve had
Here’s what nobody said during Monday’s selloff.
For two years, the honest objection to any multi-year AI roadmap was that the ground moved faster than the plan. Design an architecture around a model’s context limits in March and half of it is obsolete by September. Pick a vendor in Q1 and watch a competitor leapfrog them in Q2. Every long-range plan carried an expiration date nobody could predict.
Slower capability growth means the target stops jumping. That makes sequencing possible.
I designed a 24-month Azure AI modernization roadmap for one of my customers. A plan like that is a bet that the pace of change is predictable enough to sequence against. A paced frontier makes that bet safer, not riskier.
Slower capability growth is bad news for chip valuations. It is good news for anyone trying to plan a budget.
What did not change
Nothing about your access to AI changed this weekend.
No lab announced a release slowdown. Pricing didn’t move. The models you can call tomorrow are the models you could call Friday. Amodei’s own framing is that progress will still feel fast.
The pacing conversation is about frontier training runs, the hundred-billion-dollar supercomputers chasing the next capability jump. OpenAI’s president described it as pacing the sprint. If you’re deploying a document-processing workflow or a customer service assistant, you were never in that sprint.
Your project wasn’t waiting on the frontier. It was waiting on your data being clean enough to trust and your systems being instrumented enough to watch.
What I’d do in the next 90 days
Contain first. Every agent gets its own identity, least-privilege access, and a scoped environment. If an agent can reach production data because it inherited an admin credential from whoever set it up, you’ve built the exact failure mode the labs just publicly declined to keep risking.
Instrument second. Log every action an agent takes, not just the ones that throw errors. If you can’t reconstruct what it did last Tuesday, you don’t have a system. You have a demo with good manners.
Govern third. Name an owner for every agent and write the offboarding process before you need it. When a contractor rolls off, somebody signs a form and the badge stops working. Almost nobody has that for the AI agents already sitting inside their systems.
Then there’s procurement. If permanent third-party evaluation becomes standard practice at the big labs, that language migrates into enterprise vendor contracts within a year. Start asking now: which independent evaluators have looked at the model I’m buying, what were they allowed to see, and where is their finding published? A vendor who can answer that in 2027 will win deals against one who can’t.
The uncomfortable read
The people who build these models examined their own operational controls and concluded they weren’t good enough to keep racing at full speed.
Most companies I assess have never run that exercise at all. They have agents holding standing credentials, no monitoring at the action level, no owner for the service account, and no plan for shutting any of it down.
A slower frontier doesn’t make that safer. It removes the last excuse for not fixing it. The models are no longer going to outrun your problems on your behalf. You’re going to have to solve them yourself.
So which is it at your company: does a paced frontier buy you the time to finally build the foundation, or does it just expose that you were waiting for the models to save you from work you never scheduled?
Find out where you actually stand. An AI Infrastructure Assessment gives you a readiness scorecard, a data pipeline analysis, a security posture evaluation, and a prioritized remediation roadmap before you spend another dollar on models. Start the conversation.
See how we sequence the work. Assessment, strategy, implementation, adoption — in that order, no shortcuts. Review our services.
Keep reading. More on agent governance, AI economics, and what breaks first in real deployments. Browse the blog.
Russell Love is the Founder & CEO of Summit AI Business Solutions, based in Browns Summit, NC. With 20+ years of enterprise transformation experience at IBM and Kyndryl, Russell helps businesses build the foundations that make AI actually work.