Open vs closed frontier models and the regulatory asymmetry your AI roadmap cannot defer

The headline is an asymmetry, not a debate

Ethan Mollick posted a single sentence that captures the structural problem underneath the noise about frontier AI policy: the US is building an approval regime for closed frontier models and not for open models, and that asymmetry will not stay frozen for long. The interesting question for enterprise AI leaders is not which side of the open-vs-closed debate they prefer. The interesting question is what the policy outcome looks like under each of three plausible resolutions, and what their AI program should be doing under each. Most enterprises I speak with are treating this as a 2027-or-later problem. The actual timeline is shorter than that, and the cost of being wrong is operational rather than political

Why an approval regime, and why asymmetric

Closed frontier models are a tractable target for an approval-style regime because they have a small number of accountable corporate actors, identifiable release events, and a clear line of responsibility back to a lab. That makes them look like pharmaceuticals or aviation to a regulator: a class of actors whose behavior can be enumerated, audited, and licensed. Open models are different. They have no accountable release event, no licensed distributor, and a release surface that forks, quantizes, fine-tunes, and re-shells within weeks. A regime designed for one class will, by construction, create a vacuum around the other

The asymmetry is not a regulatory oversight. It is a structural feature of what each class of model is. Closed labs are auditable; the open model ecosystem is, by design, not auditable in the same way. Any policy that closes the gap will have to either make open models resemble closed ones, which inverts their value proposition, or accept a two-tier regime in which open models carry less legal certainty than closed ones. Neither outcome is stable, and the friction between the two is the source of Mollick's "large consequences."

Three plausible resolutions

Scenario A: closed frontier models receive an FDA-style pre-market approval regime, and open models remain largely unregulated. The likely business consequence is that closed models become the default for regulated workloads (financial services, healthcare, federal contracting, anything touching consumer credit or employment) while open models remain the default for everything else. The boundary becomes a procurement question: for which workloads do we need a licensed model, and which ones can we run ourselves. This is the cleanest outcome for incumbents and the most punishing for open-source-first strategies

Scenario B: the approval regime extends to open models via distribution-side enforcement, where fine-tuning, hosting, and downstream commercial use become the regulatory hooks. This is harder to write and harder to enforce, but it is the path Congress is most likely to attempt because it preserves the appearance of parity. The likely business consequence is that open-model deployments become subject to a documentation and provenance burden roughly equivalent to closed models, and the cost advantage of "open" shifts from licensing to deployment flexibility

Scenario C: the status quo holds for longer than expected. The federal rulemaking process is slow, the open ecosystem is politically popular, and the political appetite for a fight with the open-source community is low. The likely business consequence is that enterprises operate under an extended period of regulatory ambiguity, in which both closed and open model deployments proceed but neither carries full legal clarity. This is the worst scenario for capital planning, because every major commitment made under it has a non-trivial probability of needing to be unwound

What the asymmetry costs you right now

The asymmetry shows up first in three places that rarely make it into the AI strategy slide. First, vendor selection: the closed-model vendors you sign multi-year contracts with today are buying you a regulatory posture you may not have wanted to buy. If the regime closes around closed models, your compliance bill goes up and your switching cost goes up alongside it. If the regime opens up, the licensing premium you paid looks wasted. Second, build-vs-buy: open models let you keep optionality on hosting, fine-tuning, and deployment topology, but they also let you defer the compliance work that would otherwise be your vendor's problem. The deferral has a cost. Third, data and inference infrastructure: if you build a serious inference footprint on open weights, you have made a multi-year commitment to a deployment shape that may not match whichever regime lands

Most enterprise AI programs are not making these tradeoffs explicitly. They are happening by default, inside vendor pitches and procurement documents that treat model selection as a feature comparison rather than a regulatory posture decision. The asymmetry is forcing them to be policy decisions whether anyone planned for that or not

A defensible posture under all three scenarios

There is a posture that ages well across all three scenarios, and it has three components. First, separate the AI stack into a thin vendor-coupled layer (the model and the inference API) and a thick vendor-agnostic layer (retrieval, orchestration, evaluation, observability, data lineage). The thick layer is the part you control and the part that survives a regulatory shift. The thin layer is the part you swap. Second, treat every model commitment as a one-to-two-year commitment with a planned exit, not a multi-year commitment with a renewal. The vendors that look cheapest on a five-year total cost are the ones whose pricing and licensing you should be most skeptical of under regulatory uncertainty. Third, invest in evaluation and red-teaming that does not depend on a specific vendor's tooling. Internal evaluation harnesses are the single most important procurement decision an enterprise AI program will make in this regime, and they are usually under-funded

None of this requires picking a side on open vs closed. It requires treating the question as a portfolio allocation under uncertainty rather than as a values statement

What to do in the next two quarters

Two moves are worth the effort before the end of the year. The first is an explicit inventory of which workloads in your AI portfolio would be subject to a closed-model approval regime, which would be subject to a distribution-side open-model regime, and which are not in scope at all. Most programs do not have this inventory, and it can usually be assembled in a few weeks by walking through the production model deployments with the legal team. The second is a written, dated thesis on your default model family for the next 24 months, with the criteria that would change it. The thesis does not have to be right. It has to exist, be reviewed, and be revisable. Programs without an explicit thesis are the ones most likely to be caught flat by whichever scenario lands

Both of these moves are cheap. Both of them are usually deferred because the regulatory question feels like someone else's problem until it is not. The asymmetry is real, the timeline is shorter than it looks, and the cost of preparation is much lower than the cost of unwinding a multi-year commitment under a regime you did not plan for

A closing note

If your AI roadmap assumes that the open-vs-closed frontier model question will stay where it is, your roadmap is fragile in a way that is not yet priced in. The asymmetry is the kind of structural feature that resolves one way or another in policy cycles, and the resolutions are not equally costly for every enterprise posture. Programs that have already separated the vendor-coupled layer from the vendor-agnostic layer, written a 24-month thesis, and built internal evaluation infrastructure will absorb the change. Programs that have not will spend the second half of the cycle doing it under time pressure

Book a strategy consultation if you want help mapping this asymmetry against your existing AI portfolio before the next rulemaking cycle lands