Where AI Oversight Compliance Cost Falls by 2036 Depends on Which of Three Regimes Forms
The question is not whether frontier AI will be overseen but which regime sets the cost of release and who can fund it, and each of three scenarios answers differently.
Executive Summary
Whatever form oversight takes by 2036, the regime will decide where the cost of releasing a frontier model falls and who can fund it. Three structures fit the current record: a federal board that sets one national gate, a state patchwork of non-identical statutes, and an international standard that makes one assessment portable across borders. Each is a labelled hypothetical built on stated assumptions, with no probability attached because no named source supplies one.
In an illustrative cost model, the patchwork becomes the most expensive structure once a developer must satisfy four or more separate regimes, while the federal board and the international standard converge near the same total through different mechanics. The signposts that show which regime is forming are observable now.
Introduction: The Regulation Question Is a Funding Question
The conventional framing asks whether the United States will pass an AI law. That is directionally correct and analytically thin, because a law is only the visible part of a regime. The part that moves capital is the cost of getting a model out of the door: how many gates, how long each holds, and how much of the bill lands on developers who have not yet earned revenue.
This article treats oversight the way an operator or an allocator would: as a variable in the cost of release. It does not repeat what the proposed federal bill requires, which the pillar analysis of Warner's AI Safety Board sets out. It asks the next question: if each of three possible regimes were in force in 2036, what would a release cost, who would pay, and what would a reader see first if that regime were forming?
Every scenario below is a hypothetical and is labelled as one. A scenario here is a coherent chain of stated assumptions, not a forecast. An assumption is a condition that must hold for the scenario to occur, and a signpost is an observable event that shows the condition holding or failing. The verdict on the whole exercise is simple: oversight design is a cost-of-capital decision made in public, and the evidence for it is already on the record.
What Does the Current Record Support?
Four sets of facts constrain any scenario, and each is reported here with its source. The descriptions of the federal bill and the voluntary accord rest on published summaries and press reporting rather than a review of the enrolled bill text or the accord document itself, and they are worded accordingly.
The federal bill is introduced and contested
The Artificial Intelligence Risk Management and Security Act of 2026, S. 5576, was introduced on 29 September 2026 by Sen. Mark Warner, with Sens. Brian Schatz and Andy Kim as cosponsors, and was referred to the Senate Committee on Commerce, Science, and Transportation, according to Quiver Quantitative's bill tracker. Its design, set out in the pillar analysis, is a Commerce-housed board.
The same day, Sen. Ted Cruz, R-Texas, objected to a request to pass it by unanimous consent, saying it gave the board and the Commerce Department too much authority and that he preferred a role for independent judges, as reported by Implicator.
Two points follow. The bill has not passed and sits in committee. And the objection concerned the allocation of authority, not the principle of oversight, which matters for how Scenario A is built.
A voluntary accord runs in parallel
Implicator reports that on the same day, 29 September 2026, President Trump and executives from Anthropic, Google, Meta, Nvidia, OpenAI and SpaceXAI signed a voluntary accord. As reported, it commits signatories to internal controls that monitor model capabilities and alignment, an internal team to ensure those controls work, an independent external auditor or evaluator, and an independent board committee to oversee reports and remediation. The President called it "morally binding." The accord's text was not reviewed for this piece, so its terms are stated as reported, and no enforcement mechanism is described in the reporting.
A separate precedent sits earlier in the year. Tom's Hardware reported on 3 June 2026 that a signed executive order creates an opt-in framework under which participating developers give the federal government early access to covered frontier models for up to 30 days, and that the order creates no licensing or pre-clearance requirement. Voluntary federal access, voluntary private assurance and a pending mandatory bill therefore coexist in the same year.
At least one state has already legislated
California's SB 53 is a frontier-model statute. According to the Future of Privacy Forum's explainer, Governor Newsom signed it on 29 September and it is now law; it defines frontier models by training compute above 10^26 operations, applies a framework-publication duty to large frontier developers with annual gross revenue above US$500 million, and provides for civil penalties of up to US$1 million per violation, enforced by the Attorney General. Only this statute's text and summary were reviewed for this article; no claim is made about other states' laws.
The European Union has a statute and a deferred timetable
The EU's Regulation (EU) 2024/1689, the Artificial Intelligence Act, is dated 13 June 2024. A Gibson Dunn note of 27 May 2026 reports a provisional agreement of 6 May 2026 that would move the Annex III high-risk date to 2 December 2027 and the Annex I date to 2 August 2028, with the 2 August 2025 date for general-purpose model obligations under Articles 51 to 56 unchanged and new investigation and fine powers for the AI Office. That note records formal adoption as still pending at its date of 27 May 2026; the omnibus was then published in the Official Journal on 24 July 2026 as Regulation (EU) 2026/1744 and entered into force on 27 July 2026, moving Annex III obligations to 2 December 2027 and Annex I obligations to 2 August 2028.
The verdict on the record: a federal mandate is proposed but stalled, a state mandate exists, a private accord exists, and a foreign statute exists. None of the three scenarios is excluded and none is established.
How the Scenarios Are Built
Each scenario below carries four elements: its assumptions, what the current record says about them, the cost-of-capital consequence, and the signposts. To keep the cost comparison honest, all three use the same illustrative developer, called Developer H. Developer H is a hypothetical, not a real company and not a forecast about any named firm.
The inputs are invented for illustration and are not estimates of real compliance costs. Developer H burns US$1.0 million per month with no revenue before a model's release, and a month is treated as 30 days, so daily burn is US$33,333. It makes one frontier release a year, so annual burn is US$12 million.
The cost of a regime to Developer H has two parts: an assurance cost per regime per release, and the cash burned during any delay the regime adds beyond the existing release plan. The mechanism by which delay becomes lost runway is worked through in the analysis of the pre-release window as a cost of capital; this piece reuses that logic and varies the regime, not the mechanism.
The verdict on method: the numbers below are arithmetic on stated inputs. What they show is the shape of each regime's cost, not its size.
Scenario A: A Federal Board Sets One National Gate
Hypothetical. In this scenario, Congress enacts a statute that creates a permanent federal body with authority to set binding technical standards for frontier developers, and that statute displaces most state frontier-model rules.
What assumptions must hold?
- Assumption A1: Congress passes a frontier-model statute that creates a standard-setting body with enforceable penalties.
- Assumption A2: The statute preempts or substantially narrows state frontier-model rules, so one national standard applies.
- Assumption A3: The authority question that prompted the Senate objection is resolved, for example by adding judicial review or shifting enforcement powers.
- Assumption A4: The federal gate adds a defined, bounded delay to the release calendar rather than an open-ended one.
What the record says: A1 is unmet today, since the bill is in committee after a blocked unanimous-consent request. The objection reported by Implicator targeted authority, not oversight itself, which makes A3 the assumption that most directly decides whether A1 can be met. A2 has no support on the record; the only enacted state statute reviewed here is California's, and nothing reviewed addresses preemption. A4 depends on drafting that does not yet exist in law.
What would it cost, and who would fund it?
The defining feature of this regime is that the cost is single and fixed. Developer H satisfies one regime. Assume an assurance cost of US$0.60 million per release and an incremental delay of 15 days beyond the existing plan. Fifteen days at US$33,333 a day is US$0.50 million, so the release costs US$1.10 million, or 9.2% of annual burn. These are calculated figures on invented inputs.
The funding consequence is structural. A single national gate rewards developers that already carry assurance staff, because the fixed cost is spread over more revenue. That is the incumbency dynamic examined in the contrarian case that a Commerce-housed review board favours incumbents, and this scenario is the one in which it would be most pronounced. The benefit to capital is predictability: an investor underwrites one gate, one timeline and one penalty schedule.
The verdict: the federal board is the simplest regime to model and the least evidenced as enacted law, and its fixed cost falls hardest on the developer with the least compliance capacity.
Which signposts show it forming?
- Committee action on S. 5576, or a successor bill, with changes that answer the authority objection.
- Bipartisan sponsorship of any frontier-model statute, including text that addresses judicial review.
- Preemption language appearing in a bill that has a realistic path to a floor vote.
- The voluntary accord's auditor and board-committee commitments being referenced in draft legislation, which would show the private framework feeding the public one.
Scenario B: A State Patchwork Fills the Gap
Hypothetical. In this scenario, Congress does not enact a frontier-model statute, and states continue to legislate with differing definitions, thresholds, duties and penalties.
What assumptions must hold?
- Assumption B1: Congress enacts no preemptive frontier-model statute through 2036.
- Assumption B2: Several additional states enact frontier-model statutes, and their definitions and duties differ from California's and from one another's.
- Assumption B3: Courts and enforcers do not collapse the differences, for example through a dominant-state effect or a widely adopted model statute.
- Assumption B4: Developers cannot satisfy all regimes with a single common process.
What the record says: B1 is consistent with today, since the federal bill has not advanced. B2 is partly supported: California's SB 53 shows that a state can define a frontier developer by compute and revenue thresholds and attach a per-violation penalty. Nothing reviewed for this piece shows how many other states have acted or how their texts differ, so the count of regimes in B2 is an open question, not a finding. B3 and B4 are the real hinge: if one state's statute becomes a template that others copy, the patchwork collapses toward a single standard in practice.
What would it cost, and who would fund it?
The cost in this regime scales with the number of jurisdictions. Developer H must satisfy five regimes at an assurance cost of US$0.25 million each, for US$1.25 million, plus a small incremental delay of 5 days, which at US$33,333 a day is US$0.17 million. The total is US$1.42 million, or 11.8% of annual burn. The per-regime cost is lower than in Scenario A because each statute asks for less, yet the total is higher because the work is repeated.
The sensitivity matters more than the point estimate. If the per-regime cost were US$0.40 million instead of US$0.25 million, the five-regime total would be US$2.17 million. And the patchwork exceeds Scenario A's US$1.10 million as soon as Developer H faces four regimes: four regimes cost US$1.00 million in assurance plus US$0.17 million in delay, or US$1.17 million, while three cost US$0.92 million. The crossover point is a derived result of the invented inputs, not a prediction.
The funding consequence is geographic. Capital will price jurisdictional exposure, and a developer may restrict where a model is offered to limit the number of regimes it must satisfy. Because the repeated cost is mostly legal and procedural, it falls proportionally harder on developers that lack in-house counsel for each jurisdiction.
The verdict: the patchwork is the regime in which no single rule is burdensome and the sum is, so the cost shows up as fragmentation, not as a gate.
Which signposts show it forming?
- A second and third state enacting frontier-model statutes with definitions that differ from California's thresholds.
- Federal preemption language being removed from, or failing to appear in, bills that reach committee.
- A model statute adopted by a body that proposes uniform state laws, which would be a signal that the patchwork is converging.
- Developers publishing state-specific availability terms, which would show regime count becoming a product decision.
Scenario C: An International Standard Makes Assessment Portable
Hypothetical. In this scenario, major jurisdictions converge on a shared technical or treaty standard for frontier-model assessment, and an assessment completed under it is recognised across borders.
What assumptions must hold?
- Assumption C1: A shared standard is agreed among major jurisdictions, including the European Union and the United States.
- Assumption C2: Mutual recognition holds, so an assessment under the standard satisfies each participating regime without duplication.
- Assumption C3: Qualified independent assessors exist in sufficient number to carry the volume.
- Assumption C4: The standard is stable enough to be built against for several release cycles.
What the record says: the EU has a statute in place, and its timetable has shifted rather than disappeared. The voluntary accord's commitment to an independent external auditor or evaluator, as reported, is a private-sector building block for C3, because it implies demand for assessors. Nothing reviewed shows an agreed cross-border standard or any mutual-recognition arrangement, so C1 and C2 have no direct evidence today. The reach question for operators outside the United States, which bears on whether a US rule would ever bind a foreign deployer, is examined in the analysis of US rules and operators in Lagos and London.
What would it cost, and who would fund it?
This regime front-loads the cost. Developer H completes one assessment at an assurance cost of US$0.80 million, higher than in Scenario A because the assessment must satisfy several authorities at once, with an incremental delay of 10 days worth US$0.33 million. The total is US$1.13 million, or 9.4% of annual burn.
The result is close to Scenario A, but the cause differs. In Scenario A the cost is low because there is one national gate. Here it is low because the same assessment is reused. That makes this scenario the most sensitive to assumption C2: if mutual recognition fails for one extra jurisdiction and Developer H must add a US$0.25 million national filing, the total rises to US$1.38 million, which is above Scenario A.
The funding consequence favours developers that sell into several markets, since the portable assessment is amortised across all of them, and it favours investors who finance across borders. It also creates a new class of beneficiary, the assessor, whose capacity becomes the constraint on how many models can be released.
The verdict: the international standard has the best long-run cost profile and the weakest evidence today, because it depends on agreements that no reviewed source shows being negotiated.
Which signposts show it forming?
- A published bilateral or multilateral instrument that recognises assessments across the United States and the European Union.
- The first deferred EU high-risk obligations arriving in December 2027 and the AI Office using its new investigation powers, which would clarify what a standard would have to be compatible with.
- An accredited assessor market forming around the accord's independent-auditor commitment.
- Regulators in more than one jurisdiction citing the same technical benchmark in their guidance.
The Framework: Three Regimes, Three Cost Shapes
The table sets the three scenarios side by side on the same illustrative developer. Every figure is a calculation from the stated inputs, not a measurement.
| Measure (Developer H, hypothetical) | A: Federal board | B: State patchwork | C: International standard |
|---|---|---|---|
| Regimes to satisfy | 1 | 5 | 1 (portable) |
| Assurance cost per regime | US$0.60m | US$0.25m | US$0.80m |
| Assurance subtotal | US$0.60m | US$1.25m | US$0.80m |
| Incremental delay | 15 days | 5 days | 10 days |
| Delay cost at US$33,333 a day | US$0.50m | US$0.17m | US$0.33m |
| Total per release | US$1.10m | US$1.42m | US$1.13m |
| Share of US$12m annual burn | 9.2% | 11.8% | 9.4% |
| Cost driver | Gate length | Number of regimes | Mutual recognition |
| Assumption that decides it | A3, A4 | B3, B4 | C2, C3 |
Read across the rows, the lesson is that the three structures do not differ much in total cost under these inputs. They differ in what the cost depends on. Scenario A depends on how long a single gate holds. Scenario B depends on how many different regimes exist. Scenario C depends on whether one assessment is accepted everywhere. An investor underwriting a frontier developer is therefore underwriting a different variable in each world, and the diligence questions change with it.
The cost is always paid in the same place, which is the release milestone. Every scenario converts an oversight design into cash burned or spent before the first dollar of revenue, and all three produce the same contractual consequence: the duties that reach a developer flow down to its customers through the same vendor-contract mechanism examined in the analysis of incident clocks and vendor contracts. That common thread is why the contract layer is the most stable thing to build on across all three futures.
The verdict: the regime changes who pays, not whether someone pays.
Where These Scenarios Fit: Hybrids and Boundaries
The three scenarios are pure types, and a hybrid is a plausible case. A federal statute that preempts only part of the state field, a patchwork that converges on one state's template, and an international standard that sits beside national filings are all combinations of the assumptions above. The framework handles a hybrid by asking which assumption has failed. If A1 holds and A2 fails, the result is Scenario A with a Scenario B residue. If C1 holds and C2 fails, the result is Scenario C with a Scenario B tail.
Two boundaries limit what the scenarios claim. First, they do not say which regime will form, and no probability is offered because none comes from a named source. Second, they say nothing about any named company's prospects. Developer H is invented, and the voluntary accord's signatories are named only because their participation was reported, not because any outcome is predicted for them.
The voluntary accord deserves one further note. As reported, it describes an assurance model, with an external auditor and a board committee, that could be absorbed into any of the three regimes. It could become the template for a federal board's audit requirements, the baseline that state statutes reference, or the seed of an international assessor market. That makes it the one development on the current record that is compatible with all three scenarios, and for that reason the most useful signpost to follow.
The verdict: scenario analysis earns its value by showing which assumption to watch, not by choosing a winner.
Signposts to Watch Before the Next Review
The table consolidates the signposts by scenario and states which direction each one moves the picture. A signpost is evidence about an assumption, not a verdict on a scenario.
| Signpost | Assumption it tests | If it occurs |
|---|---|---|
| Committee action on S. 5576 with the authority objection addressed | A1, A3 | Strengthens Scenario A |
| Preemption language in a bill with a floor-vote path | A2, B1 | Strengthens A, weakens B |
| Additional states enacting frontier statutes with differing thresholds | B2, B3 | Strengthens Scenario B |
| A uniform-law body adopting a model frontier statute | B3 | Weakens B, moves it toward a single standard |
| First deferred EU high-risk obligations applying (2 December 2027) and first AI Office use of its new powers | C1, C4 | Clarifies what any shared standard must match |
| A mutual-recognition instrument covering assessments | C1, C2 | Strengthens Scenario C |
| An accredited assessor market forming around the accord's auditor commitment | C3 | Supports C and eases A |
A reader who tracks only two of these will capture most of the information. The first is whether any federal bill advances with the authority question answered, because that single event moves Scenario A from unmet to plausible. The second is whether a second and third state follow California, because that moves Scenario B from a single data point to a pattern. The EU and mutual-recognition signposts matter more over a longer period, and they are the slowest to resolve.
The verdict: the next twelve months cannot decide the regime, but they can eliminate assumptions, and eliminated assumptions are the useful output of scenario work.
Why This Matters for Operators and Capital in Washington D.C., London and Lagos
For operators and allocators who move between the three markets, the scenarios translate into different working postures. The questions below are the ones the evidence can answer.
What does each scenario mean for Washington D.C.?
For policy-facing operators and government contractors, Scenario A means one counterpart and one rulebook, and Scenario B means a state-by-state engagement model that rewards local presence. The practical step in either case is to build the compliance function around the assumption that decides the regime, not the one that is most visible this month.
What does each scenario mean for London?
For London-based investors and operators who sell into both the EU and the US, Scenario C is the most valuable and the least evidenced. Under Scenarios A and B, a London developer faces its home requirements plus a US layer. The investor question is whether a portable assessment will ever exist, and until it does, diligence should price duplicated filings. UK domestic law was not reviewed for this piece and is not assessed here.
What does each scenario mean for Lagos?
For Lagos-based operators that build on US or European models, the exposure is mostly contractual under any scenario: duties on the developer reach the operator through the terms of access, as the Lagos and London analysis in this network examines. A patchwork raises the chance that availability terms differ by jurisdiction, while an international standard would lower the cost of an operator showing that the model it uses has been assessed. The common action is to treat the vendor contract as the control surface.
What does this mean for capital allocation?
For allocators, the scenarios turn a vague regulatory risk into specific diligence variables. Under Scenario A, the variable is the length of the gate. Under Scenario B, it is the count of regimes. Under Scenario C, it is whether one assessment travels. A fund that records which variable it is assuming can revisit the assumption when a signpost moves, instead of repricing the whole thesis.
The verdict: the regime is uncertain, but the variable each regime puts at risk is known, and capital can be structured around it.
Key Takeaways
- Which oversight regime forms by 2036 will decide where the cost of releasing frontier AI falls, because each of the three scenarios attaches that cost to a different variable.
- The current record supports no scenario fully: the federal bill is in committee, one state statute is in force, a voluntary accord exists as reported, and the EU high-risk timetable has been law since 27 July 2026, with the first deferred obligations due on 2 December 2027.
- In the illustrative model, a state patchwork costs more than a single federal gate once a developer faces four regimes, so the number of regimes matters more than the cost of any one of them.
- An international standard has the best long-run cost profile and the thinnest present evidence, because it rests on mutual recognition that no reviewed source shows being negotiated.
- The voluntary accord's independent-auditor commitment is the signpost compatible with all three scenarios, which makes it the most informative single development to follow.
Conclusion
The three regimes examined here fix three different costs of release: a gate in Scenario A, a repetition in Scenario B, and a portability test in Scenario C. The current record contains enough to name the assumptions that separate them and to list the signposts that will show which is forming, and not enough to rank them. A reader who holds the assumptions as live variables, and revisits them when a committee acts, a state legislates or a recognition instrument is published, will be positioned for any of the three. The next review of the signposts belongs in 2027, when a year of legislative and regulatory events will have tested the assumptions on the page.
External References
- Implicator: Cruz blocks AI bill requiring 45-day model access as Trump backs voluntary accord
- Quiver Quantitative: S. 5576, Artificial Intelligence Risk Management and Security Act of 2026
- Future of Privacy Forum: California's SB 53: The First Frontier AI Law, Explained
- Gibson Dunn: EU AI Act Omnibus Agreement β Postponed High-Risk Deadlines and Other Key Changes
- EUR-Lex: Regulation (EU) 2026/1744 (Digital Omnibus on AI)
- EUR-Lex: Regulation (EU) 2024/1689 (Artificial Intelligence Act)
- Tom's Hardware: Trump signs AI executive order seeking 30-day government access to frontier models before release







