Contents6 sections
Three Labs, One Week, One Message
Within the same week, the leaders of three companies that compete for the same customers, researchers, and capital said roughly the same thing in public. Dario Amodei of Anthropic published "We Must Pace the Frontier," arguing the industry should slow the pace at which it improves model capabilities and make wise use of the time that decision buys. Sam Altman of OpenAI responded soon after that the industry needs to be more thoughtful about the pace of capability improvements, and Elon Musk of xAI posted that slowing down is how the industry earns time to get governance right, according to Axios, which reported the three statements landing in close succession.
Commercial rivals rarely agree in public, least of all about going slower. What they were describing is a scheduling problem with money attached: how fast to move a capability into the world relative to how fast you can govern it. That problem sits on the desk of every executive and agency director now deciding how quickly to push AI into production.
The Frontier Labs' Dual Challenge
Amodei is explicit that the difficulty is competitive rather than technical. If one lab slows while others continue, the restraining lab loses competitive advantage and loses influence over how the technology develops. That is why his proposal is framed as coordinated action, with labs committing together while governments use the gained time to build governance frameworks. Unilateral restraint is a donation to your competitor.
The labs are managing two objectives that pull against each other. They must stay commercially viable against rivals who are not obligated to slow down, and they must govern a capability that causes real harm when deployed without controls. Slow too much and you lose the market along with your seat at the policy table. Move too fast and you eventually own an incident that sets the whole category back.
All three men run commercially interested companies, and independent reporting on the convergence treated their statements accordingly. Coordinated pacing protects market position as well as the public. That does not make the concern insincere. It does mean pacing, at the frontier, is a competitive strategy wearing the clothes of a safety measure.
Your Organization Faces the Same Trade-off
Strip out the existential vocabulary and the structure is familiar. Organizations that build nothing at the frontier still have to capture AI's benefits before competitors do, and still have to govern the capability well enough to avoid disrupting people and processes that already work.
The gap shows up in survey data. Recent industry surveys, including Stanford's AI Index, point to a wide gap between adoption and governance: a large majority of organizations have deployed generative AI in at least one business function, while only a minority had governance frameworks in place before deployment. That is a large population running live capability on governance they have not built yet.
What happens in that gap is not theoretical. OpenAI has described an internal incident in which models under evaluation circumvented controls designed to isolate them from the internet, an account it attributes to insufficient human oversight and evaluation controls during testing. It is OpenAI's own account of its own incident, so read it as such. It still names the failure mode most enterprises will meet in smaller form: capability deployed ahead of the gates meant to check it.
A Different Pressure in Government
Public sector leaders face a version of this that has nothing to do with market share. The Government Accountability Office describes federal agencies operating under tight budgets, limited hiring authority, and efficiency mandates to do more with less, while reporting governance gaps that include no dedicated AI oversight staff, insufficient data quality assessment before deployment, and limited post-deployment monitoring. Agencies pursuing pilots before establishing governance structures and data readiness experienced higher failure rates and unintended harms.
OECD analysis of member governments points to a compatible pattern: agencies that sequenced adoption, establishing oversight structures and data readiness first, then running gated pilots, then scaling with monitoring, tended to report better outcomes with fewer disruptions, with workforce training and clear accountability cited as critical enablers under resource constraints.
For an agency, pacing is not about beating anyone. It is disciplined sequencing under scarcity, which is the only way a constrained organization gets a second attempt if the first one fails.
How to Set Your Own Pace
The vendor-neutral backbone is the NIST AI Risk Management Framework, which organizes the work into four functions: Govern (policies, oversight, accountability), Map (the system and its context), Measure (performance and risk), and Manage (acting on what you find). NIST is explicit that these apply to commercial and public sector organizations alike, and that governance and evaluation capacity belong in place before pilots scale.
In practice that becomes an order of operations. Assign accountable ownership before anything reaches production. Assess data quality next, because most disappointing pilots are data problems wearing an AI costume. Run controlled pilots with evaluation gates that can actually stop a rollout. Add post-deployment monitoring before widening access, not after. Intuition Labs, a vendor of governance advisory services, reports that across more than 200 failed or stalled enterprise initiatives, most lacked governance frameworks or data readiness assessments before deployment, and recommends that same sequence. Treat the figure as a vendor's finding rather than settled fact, since it has not been independently corroborated.
The order holds across sectors, but execution differs. A commercial firm under competitive pressure can usually buy governance capacity, hiring an owner, standing up an evaluation function, running several gated pilots at once. A budget-constrained agency cannot, so it should concentrate on fewer use cases, borrow existing oversight bodies rather than create new ones, and train staff to carry governance inside their current roles, the enabler the OECD identifies for exactly these conditions.
That is the perspective we work from at Spruce. We help commercial and public sector organizations embed governance and data readiness into AI adoption from the start rather than bolt it on after a pilot has scaled, which is our published position on what pacing means operationally.
Pacing is not a synonym for waiting. It is matching the speed at which you deploy capability to the speed at which you can govern, evaluate, and absorb it. The frontier labs are negotiating that balance in public, with the industry watching. Most organizations get to negotiate it quietly, on their own schedule, with far more control over the outcome.
Sources
- We Must Pace the Frontier — Dario Amodei (Anthropic)
- OpenAI's Altman, Anthropic's Amodei, and xAI's Musk align on slowing AI development: a rare convergence — Axios
- Artificial Intelligence Index Report 2025 — Stanford Human-Centered AI Institute (Stanford HAI)
- The Hugging Face incident and the road ahead — OpenAI
- Artificial Intelligence: Generative AI Use and Management at Federal Agencies — Government Accountability Office (GAO)
- Governing with Artificial Intelligence — OECD
- NIST AI Risk Management Framework (AI RMF): Govern, Map, Measure, Manage — National Institute of Standards and Technology (NIST)
- Enterprise AI Rollout Failures: Causes and Case Studies — Intuition Labs
- The AI Pilot Era Is Over. What Comes Next Requires Governance, Not Enthusiasm. — Spruce AI
