Paying for Safety Compute

Interactive companion to the paper

Safety Research and the Allocation of Frontier Compute

During a frontier takeoff period, an AI lab splits a committed stock of compute among capability research, safety research and customers, and a regulator limits both the services customers lose and how long any capability can be delayed. Extra safety compute can then come from customers or from capability research. Customer funding leaves every capability date unchanged, while funding from capability research gives safety knowledge more time at the cost of later capability and later customer services. Because research funding delays the end of the takeoff period, the cheaper source depends on how much finishing raises customer services: customers should pay when this finishing premium is above a threshold, and capability research should pay too below it. Three quantities move the threshold: the delay a point of research costs, how much extra knowledge safety compute buys, and how fast the value of customer services diminishes. The paper also asks when safety compute should be spent and what a regulator can commit to before it knows how safety research works.

Ask what finishing is worthFunding safety from capability research delays the end of the takeoff period, by 0.11 months for each point taken. The delay costs more the more customer services rise once the frontier is finished. Customers should pay when that rise exceeds 34 percent. It is 64 percent if freed compute can serve them at once and 13 percent if it cannot.Sections 2.1 and 3, Propositions 1 and 2
Spend safety compute earlyIn the baseline, front-loaded customer-funded safety cuts the probability of catastrophe from 75.00% to 51.01%, or from 25% to 13.76% if full speed carries 25%. It comes within 0.91 points of the best tested rule, 0.33 at the lower probability.Proposition 4, Section 5.1, Table 3
Slowing capability research pays when delay is cheapIn the baseline, capability research should pay even when all freed compute can serve customers if labs put more than 64% of compute into capability research or the research pace responds little to compute. Estimated lab shares lie on both sides; one lab’s 2026 accounts put at least 73% of its compute spending outside paying customers, above the flip point. Takeoff periods of 9 to 36 months leave the answer unchanged.Section 5.4, Figure 4

Each step below follows a section of the paper. The figures run the paper’s model in your browser and open at the paper’s settings, where every number equals the one the paper prints. Move a slider to see what drives a result; “Paper’s settings” puts it back. Constant rules are solved in your browser. The two-stage and segment rules are the paper’s saved multistart solutions, which the page evaluates but does not re-solve. All numbers come from one conditional scenario built on the AI Futures Project’s worked example; they are not estimates. The paper treats its 75 percent catastrophe probability as the upper end of a range that runs down to 25 percent (Step 11).

Contents
  1. Four findings at a glance
  2. The pieces
  3. The research clock
  4. Knowledge and risk
  5. Services and the budget
  6. Who should pay
  7. What slowing does to hazard
  8. Who should pay
  9. The threshold and the budget
  10. When to spend
  11. Safety early
  12. The numbers
  13. Known technology
  14. Larger budgets
  15. What moves the threshold
  16. How services are valued
  17. Commitment and limits
  18. Commitment under uncertainty
  19. Predictions and limits

At a glance · Paper, Introduction

Four findings, and the one comparison behind them

Extra safety compute has to come from customers or from capability research; below, “research” means capability research. The paper compares the two sources by what they cost customers and what they buy in safety, and the diagram below shows that comparison. The thirteen steps that follow build it piece by piece; here is where they lead.

A way to picture it: a race team with one budget. The engine is capability research, the brakes are safety research, and part of the budget also serves paying customers. Extra brakes can be paid for in two ways: out of the customers’ share, which leaves the car as fast, or out of the engine budget, which slows the car and delays what a finished car would earn. Whether to slow the engine depends on how much a finished car can earn, and the earlier the brakes are fitted, above all before the car reaches the dangerous high-speed stretch, the more they protect.

  1. The value of finishing decides who should fund the next unit of safety compute. Both sources buy the same safety compute, but funding from capability research also delays completion, and with it everything a finished frontier provides. Slowing capability research thus costs customers in two ways: better models arrive later, and so does the compute that capability research releases at completion. The second cost depends on how much of that compute can serve customers, and it decides the comparison. So the cheaper source depends on the finishing premium, the proportional step up in customer services at completion: in the baseline, customers should pay when it is above 34 percent, and capability research should also pay when it is below. If the compute freed by capability research can serve customers at once, finishing raises their services by 64 percent and customers should pay; if it cannot, finishing adds only 13 percent and capability research should pay too. Stated as the customer share after completion, the funding threshold is 63.5 percent. Steps 4–5
  2. Two kinds of case make capability research the cheaper source whatever customers can use. Either delay is cheap: in the baseline, when labs put more than 64 percent of compute into research or the research pace responds little to compute, a point taken from research delays completion by little (0.09 months at a 65 percent research share and 0.08 with ε = 0.3, against 0.11 at baseline). One frontier lab’s 2026 accounts put at least 73 percent of its compute spending outside paying customers, above that flip point. The length of the takeoff period matters little: a longer period makes each cut cost more months, but customer funding then gives up services over a longer period as well. Or customers’ share is already small: above a budget of 30.2 percent, each further unit taken from customers costs more, while the cost of a research cut stays bounded. Steps 6 and 9 Steps 10–11
  3. Customer-funded safety comes early. Knowledge bought early protects more of the remaining hazard, while customer compute grows more valuable as compute and quality grow. In the baseline, front-loaded customer-funded safety cuts the probability of catastrophe from 75.00 to 51.01 percent, within 0.91 points of the best tested rule that also cuts capability research. At a full-speed probability of 25 percent the risk falls to 13.76 percent, and the rule that also cuts research does 0.33 points better. Steps 7–8
  4. A rule fixed before the safety technology is known costs little, unless safety work advances capability. Without spillover the chosen rule is within 0.28 points of an informed choice in the tested technologies; spillover changes what a rule costs, so it matters more. Step 12

How to read the diagram. Both sources put the same compute into safety, so they differ only in their side effects. A customer-funded unit costs services now and delays nothing. A research-funded unit keeps customers whole during the takeoff period, gives knowledge more time and adds exposure, and delays the finished frontier and the services it brings. Steps 4 and 5 compute each side.

The implication is that who pays turns on a number a regulator can look up, how much of the compute freed by capability research could serve customers once the frontier is finished, and on three quantities that set the threshold: the delay a point of research costs, how much extra knowledge safety compute buys, and how fast the value of services diminishes. Timing is a separate choice: whoever pays, safety compute is worth most early in the takeoff period.

Terms in plain words

Hazard
the probability per month that a catastrophe, an irreversible loss of human control over AI, happens if it has not happened yet
Takeoff period
the stretch from the Automated Coder to the finished frontier; 18 months at full speed in the AI Futures Project’s worked example
State and step hazard
the hazard of running a model, which grows with the months spent at each capability, and of training a stronger one, which depends only on how far the frontier moves
Safety knowledge
what safety research builds; it scales down every later hazard, so knowledge bought early protects more
Customer-funded safety
safety compute taken from customers’ share; no capability arrives later
Research-funded safety
safety compute taken from capability research; every later milestone arrives later
Serving capacity
the largest share of installed compute that can serve customers; what matters is how much of the compute freed by capability research can serve them once the frontier is finished
Finishing premium and threshold
the proportional step up in customer services when the frontier is finished; customers are the cheaper source of extra safety compute when it is above a threshold, 34 percent in the baseline, a customer share after completion of 63.5 percent
Service budget
the cap on the services customers give up over 60 months, against full speed; the baseline cap is 11.5 percent of full speed’s services over its takeoff period, about 2.1 months of them
Front-loading
spending more of the safety compute early in the takeoff period, when compute and model quality are low

Part 1The pieces: compute, knowledge and services

Step 1 · Paper, Sections 2.1 and 2.2

Three uses of compute, and a clock that keeps time in research capacity

A frontier lab uses its compute in three ways. It serves customers, the use called inference. It does capability research, which trains stronger models. And it does safety research, which makes models less likely to cause harm. The three uses compete for the same chips. The paper’s reference allocation puts 50 percent of compute into capability research, 5 percent into safety and 45 percent into customers. Call it full speed. At full speed the frontier runs from an Automated Coder milestone to a top-expert-dominating endpoint in 18 months, as in the worked example of the AI Futures Project.

More safety compute must come from customers or from research. When customers pay, every capability arrives on its reference date and customers give up services. When research pays, the frontier slows. Slower research gives knowledge more time to accumulate before each capability level, but it also prolongs exposure to state risk at that level and delays everything a finished frontier provides. Who should pay is the question of which source removes more hazard per unit of services customers give up.

Where the baseline numbers come from

NumberValueSource
Full-speed takeoff period, Automated Coder to the top-expert-dominating endpoint18 monthsAI Futures Project (2026), worked example: 1.5 years.
Probability of catastrophe at full speed75% (range 25 to 75%)AI Futures Project (2026), worked example; the same 75% for almost no safety effort in Greenblatt (2025). The paper also reports 25%, the probability Samotsvety Forecasting (2022) give for misaligned AI takeover by 2100; surveys of AI researchers report medians of 5 to 10% (Grace et al. 2025).
Probability of catastrophe under a safety-case regime40% (10% at 25%)AI Futures Project (2026). Treating it as a 12-month pause at the start, which fixes safety productivity, is the paper’s modeling choice. When the full-speed probability changes, the regime’s probability moves so that it cuts cumulative hazard in the same proportion.
Longest milestone delay allowed12 monthsAI Futures Project (2026): a rival would need about a year to catch up if the leader halted.
Growth of installed computedoubles every 6 monthsAn illustrative path, not an estimate. Anthropic’s compute and infrastructure expense rose from $1.39 billion in the second quarter of 2025 to $5.56 billion a year later, a doubling every six months (Anthropic, PBC 2026b); the paper treats this as a fast case, because the lab grew fast and spending is not physical compute. The trend in training compute, about six months (Sevilla et al. 2022), is a different object.
Full-speed compute shares50% research, 5% safety, 45% customersThe paper’s reference allocation; assumed. Anthropic put about 6% of its research compute into safety in one week of 2026 (Anthropic 2026a), and compute for paying customers was at most 27 to 38% of its compute spending (Anthropic, PBC 2026b).
Research speed and service quality at the endpoint10 times the startAssumed.
Service budget; horizon; discount rate11.5% of full-speed services over the takeoff period; 60 months; 3% a yearThe paper’s policy settings; Steps 6, 9 and 12 also use budgets of 2.3 to 69.1%.

Capability over time

Delay at each capability level

Left panel, capability. Capability, from 0 at the Automated Coder to 1 at the endpoint, against months. The gray dashed curve is full speed. The blue curve holds research at the share on the slider. The orange dotted curve is the same share without compute growth. The slowed lab moves more slowly than the reference at first; later its pace comes closer to the reference, because it reaches each level when compute is more plentiful.

Right panel, delay. Months behind the reference at each capability level, with and without compute growth. The first level is never delayed. With growth the delay rises more and more slowly; without growth it rises in proportion to the reference date.

The implication is that slowing research does not shift the whole path by a fixed number of months. Growing compute makes a research cut cheaper in time than the cut in research compute suggests.

The math

Step 2 · Paper, Sections 2.3 and 2.7

Knowledge scales hazard down, and customers can pay for it without delay

Safety research builds a stock of knowledge S. Knowledge is measured in units of log-hazard reduction: adding log 2 to S halves every later hazard. It grows at Ṡ = σ y(s), where the learning multiplier y(s) scales learning with the safety share. Half of reference-share learning needs no dedicated compute; the other half has square-root returns. Safety productivity σ = 0.086116 per month is fitted so that a 12-month pause at the start would cut the probability of catastrophe from 75 to 40 percent.

The catastrophe is an irreversible loss of human control over AI. The paper reads this danger as misalignment. A more capable model is more likely to pursue goals its developers did not intend and better able to escape control, so the danger rises with capability. Conditional on no catastrophe so far, hazard has two parts:

λ(t) = [h(F) + k(F)Ḟ] e−S.

State risk h accrues per month spent at a capability level. Step risk k accrues per unit of capability gained. Slower research spends more months at each level, so it adds state risk; step risk does not depend on speed. Knowledge scales both parts down in the same proportion. Full speed carries a 75.00 percent probability of catastrophe. The paper treats it as the upper end of a range and also reports 25 percent (Step 11). Half of the hazard is state risk, and three quarters of it falls in the second half of capability development. The baseline counts hazard only during the takeoff period, as the AI Futures Project’s 75 percent does. So it assumes that the danger ends at completion, for example because a top-expert-dominating model still under control can help align its successors.

The figure opens at customer-funded safety. Research keeps 50 percent, and safety takes the largest share customers can fund within the 11.5 percent service budget of Step 3, 14.77 percent. The probability of catastrophe falls to 57.68 percent, and no capability arrives later.

Probability of catastrophe over time

Hazard met at each capability level

Left panel, probability. The cumulative probability of catastrophe against months. The gray dashed curve is full speed, ending at 75.00 percent. The blue curve uses the shares on the sliders, with customers taking the rest. Most of the risk comes late, near the endpoint.

Right panel, hazard by capability. The hazard met per unit of capability, after the protection that knowledge gives by the date each level arrives. The gray dashed curve is full speed. The blue curve is the total on the sliders’ path, and the green dotted curve its state part. More safety lowers the curve at every later level.

The implication is that safety bought early protects all of the later hazard. Cutting research also lowers the curve at late levels, because knowledge has more months to grow, but it raises the state part, because each level is crossed more slowly.

The math

Step 3 · Paper, Sections 2.4 and 2.5

Customers value services, and a service budget prices both funding sources

Customers value inference capacity and model quality. Their discounted service flow is

b = e−rt√Q = e(G/2 − r)t + μF/2 √min{i, ī}.

Model quality eμF rises tenfold over the takeoff period. The square root gives diminishing value: the marginal unit of customer compute is worth half the average unit. At a fixed level and share the flow grows at κ = G/2 − r, 5.5 percent a month, because compute grows much faster than the 3 percent annual discount rate.

After completion, capability research is no longer needed. What customers receive then is the continuation, and the question is how much of the compute research frees can serve customers right away. Serving capacity ī is the largest share of installed compute that can serve customers. Under full redeployment the freed compute serves customers at once, so a finished frontier gives them 95 percent. Anthropic describes its compute as “a single, fungible pool” that it directs among research, model development, serving customers and internal use, and states that demand for compute has historically exceeded available capacity (Anthropic, PBC 2026b). Under no redeployment the freed compute cannot serve customers, so they keep the reference 45 percent after completion. The same filing warns that if the lab’s compute needs change, it may hold leased or dedicated infrastructure poorly suited to its actual needs (Anthropic, PBC 2026b), as compute built for training might be once training ends. The value of finishing is what completing a month earlier adds to customer services. It is large under full redeployment and small under no redeployment.

Rules are compared over the takeoff period and the 12 months after it. The deadline makes every rule finish within that window, by month 30. Once a rule and the reference have both finished, they give customers the same services, and in the baseline danger has ended, so every difference in services and hazard between rules falls inside the window. Installed compute over the window is taken as given, because compute is bought well before it is used: Anthropic reports that most of its compute is procured under multi-year commitments made well in advance of deployment (Anthropic, PBC 2026b). A lab that cannot put freed compute to use would eventually build less, but only after both paths have finished, so the comparison is unchanged. Service costs are counted over 60 months, and any horizon from month 30 on gives the same cost.

The regulator minimizes cumulative hazard subject to two commitments. The service budget caps the discounted services customers give up, against the reference, over 60 months. The deadline caps the delay of every capability milestone at 12 months. Budgets are shares of a fixed unit J̄, the reference’s services over its takeoff period, months 0 to 18, the period in which customer funding costs anything: J̄ = 31.317 index units, so the 11.5 percent budget of the main comparisons is 3.61 units, about 2.1 months of the reference’s average takeoff-period services.

A customer-funded rule keeps every capability date, so after month 18 it serves customers exactly as the reference does. A constant customer share i therefore costs (1 − √(i/i0)) J̄, and a budget b buys the share with √(i/i0) = 1 − b, whatever the length of the takeoff period or the growth of compute. With research at 50 percent, the 11.5 percent budget lets customers fund a safety share of 14.77 percent, and they keep 35.2 percent. The reference’s 60-month value is 984.785, 31.45 times J̄, so the same budget equals 0.366 percent of five-year reference services.

Discounted service flow

What the budget buys when customers pay

Left panel, service flows. The discounted service flow against months, in the continuation on the buttons. The gray dashed curve is full speed; the blue curve is customer-funded safety at the budget on the slider. Under full redeployment the flow jumps at completion, month 18, when customers receive the compute that capability research no longer needs. Under no redeployment it does not jump. The gap between the curves before month 18 is what customers give up.

Right panel, what the budget buys. With research held at 50 percent, the largest safety share customers can fund at a constant rate (blue) and the probability of catastrophe it gives (green), against the budget. Both are in percent. The first points of budget buy the most.

The implication is that both funding sources are priced in the same unit, the services customers lose. Customer funding costs services before completion. Research funding costs quality that arrives later and a finished frontier that arrives later.

The math

Part 2Who should pay: the funding comparison

Step 4 · Paper, Section 3, Proposition 1

Both sources buy the same safety; they differ in what slowing does to hazard

Take constant shares. Moving one point of share into safety changes cumulative hazard by Hs when customers pay and by Hs − Hx when research pays. Both buy the same direct gain in knowledge. They differ through Hx, the hazard effect of more research, which has two parts (Proposition 1):

Hx = knowledge term − exposure term.

The first part is knowledge. A slower frontier reaches each level later, after more knowledge has accumulated; ℓ(F) is the delay at level F per unit of research share. The second part is exposure. A slower frontier spends more months at each level, which adds state risk. Research gives the larger hazard cut exactly when Hx > 0.

At customer-funded safety with the 11.5 percent budget, the knowledge term is 1.002 and the exposure term 0.183, so Hx = 0.819 and Hs = −3.285. A unit of safety share funded by research cuts 25 percent more hazard than one funded by customers.

The advantage needs productive safety research. Scale safety productivity down with the hazard functions held fixed. At a tenth of baseline productivity the knowledge term is 0.555 and the exposure term 0.900, so Hx = −0.345: customer funding cuts more hazard. Hx changes sign at 16.3 percent of baseline productivity.

Hazard cut by one unit of safety share

The research effect against safety productivity

Top panel, the two sources. The fall in cumulative hazard from one unit of safety share, at the productivity on the slider, when customers pay and when research pays. The allocation is customer-funded safety at the 11.5 percent budget: research 50 percent, safety 14.77 percent.

Bottom panel, the research effect. The knowledge term (blue), the exposure term with its minus sign (orange) and their difference Hx (black), against safety productivity as a multiple of the baseline, on a log scale. The exposure term shrinks as productivity rises, because knowledge also protects the extra months. The knowledge term first grows and then falls, because strong knowledge leaves less hazard to protect. The dot marks the slider.

The implication is that the hazard side favors slowing research only when safety research is productive enough. Hazard is half of the decision; Step 5 adds what each source costs customers.

The math

Step 5 · Paper, Section 3, Proposition 2 and Corollary 1

Who pays depends on how much finishing raises customer services

Services are the other half. When customers pay, one more unit of safety share costs cC = Jpre/(2i), half the average value of their compute before completion. Customers keep i = 0.352 and Jpre = 27.7, so cC ≈ 39.33 index units per unit of share. When research pays, the cost is the quality that arrives later, Qx = 29.66, plus an endpoint term: the completion delay times the value of finishing, net of what the slowed frontier already provides (Proposition 2).

The comparison is made at a given allocation, and a finished frontier gives customers all the compute that can serve them, in the reference and in every policy. With the reference held fixed, choosing this continuation for a policy loses nothing, because after completion hazard no longer depends on how compute is split. A finished frontier then gives customers the continuation share ic, so finishing raises their service flow by the premium √(ic/i) − 1. Capability research is the cheaper marginal source when this premium is below a threshold π*, that is, when the customer share after completion lies below the funding threshold (Corollary 1)

ī* = [max{0, √i (1 + π*)}]2.

Here capability research is cheaper when the premium is below 34 percent. So the customer share after completion must be below (0.594 × 1.343)2 ≈ 0.635, or 63.5 percent. Under full redeployment finishing raises the service flow by 64 percent. Funding from research then costs 66.13 units and cuts hazard by 0.062 per unit of cost, against 0.084 for customer funding: customers should pay. Under no redeployment finishing adds only 13 percent. Research costs 37.05 units and cuts 0.111 per unit of cost: capability research should pay too.

Danger that continues after the takeoff period barely moves the threshold. By completion a unit of share moved into safety adds 5.82 units of knowledge when research pays and 4.51 when customers pay. That ratio, 1.29, is close to the ratio of takeoff hazard reductions, 1.25. So the threshold rises only to 65.8 percent at ξ = 1, when 82 percent of hazard falls after completion, and it never exceeds 66.2 percent.

Hazard cut per unit of service cost

The threshold as danger continues

Left panel, the two sources. Hazard reduction per index unit of service cost when one unit of share moves into safety at customer-funded safety, against the continuation share ic. The green dashed line is customer funding, which does not depend on the continuation. The blue line is research funding, which costs more as finishing becomes more valuable. They cross at the threshold, the dotted line. The dot marks the slider.

Right panel, continuing danger. The threshold against the share of hazard that falls after completion, as continuing danger ξ grows. The dotted line is its limit. Full redeployment, at 95 percent, stays above it in every case.

The implication is that the continuation selects the funding source. When the freed compute can serve customers at once, delay forgoes a large step up in services, and customers should pay. Under no redeployment, delay costs only quality, and capability research should pay too.

The math

Step 6 · Paper, Section 3, Proposition 3, and Figure 3a

The threshold rises with the service budget

A larger service budget changes the marginal comparison, because the customer-funded optimum then leaves customers less compute. Corollary 1 evaluates the threshold locally, at that allocation. Along the customer-funded optimum the threshold takes a simple form (Proposition 3):

√ī*(i) = (1 − θQ)√i + θC ρ(i)/√i, ρ = RR/RC.

The weights θC and θQ do not depend on i. The first term is quality and shrinks like √i. The second is the customer term. Each further unit taken from customers costs more as their share shrinks, so it grows like 1/√i. Whatever customers can use, research becomes the cheaper source once their share is small enough. This needs research funding to keep cutting hazard by a positive fraction of the cut from customer funding, not a hazard advantage.

In the calibration θC = 0.244 and 1 − θQ = 0.478. At the 11.5 percent optimum the formula reads 0.284 + 0.513 = 0.797. The ratio ρ rises from 1.13 at a budget of 2.3 percent of J̄ to 1.69 at 69.1 percent. The threshold is 63.5 percent at an 11.5 percent budget and 106.5 percent at 34.5 percent. It reaches 95 percent at a budget of 30.2 percent, when customers keep 22.0 percent of compute. Above that budget research is the cheaper source even when customers can use all freed compute.

The rise does not rest on a hazard advantage. At a tenth of baseline productivity slowing research cuts less hazard than customer funding (ρ between 0.81 and 0.92). Yet the threshold still rises, from 38.2 percent at an 11.5 percent budget to 139.9 percent at 69.1 percent. The threshold rises throughout the baseline budget range; in general, a rise also requires the relative hazard benefit of research funding not to fall too quickly as safety grows.

Threshold by budget

The two terms of the threshold

Left panel, the threshold. The funding threshold at the customer-funded optimum for each budget, in percent of compute. Below it, research is the cheaper marginal source. The dotted lines mark the continuation shares of full redeployment (95 percent) and no redeployment (45 percent). The orange dashed curve uses a tenth of baseline safety productivity. A threshold above 100 percent means research is cheaper whatever customers can use.

Right panel, the two terms. The quality term (1 − θQ)√i and the customer term θC ρ/√i against the budget. Their sum, in black, is the square root of the threshold. The customer term drives the rise. Where the sum crosses the dotted line at √0.95, the threshold passes the full-redeployment share.

The implication is that the test of who pays matters most at larger budgets. A small budget favors customer funding under full redeployment; a large one favors slowing research in every continuation.

The math

Part 3When to spend: front-loading

Step 7 · Paper, Section 4, Proposition 4, and Section 5.1

Customer-funded safety comes early relative to customer use

On interior stretches the baseline optimum front-loads safety relative to customer use. Along such a stretch the marginal values of safety and customer compute are equal. The shadow value of knowledge is the hazard still ahead, A(t). Equating the two margins gives the ratio of safety to customer compute and its rate of change (Proposition 4):

d/dt log(s/i) = 2 d/dt log σ − 2λ/A − G + 2r − μḞ.

Three forces push safety early relative to customer use. Knowledge bought early protects more of the remaining hazard. A unit of customer share serves more as installed compute grows and as quality rises. Only discounting pushes the other way. With flat productivity the ratio falls by at least 10.46 percent a month, from compute growth net of discounting alone, κ = 5.5 percent a month. Rising productivity reverses the decline only if it rises more than eμ/2 = 3.16-fold across the takeoff period (ζ > μ/2 = 1.15). The threefold rise of the uncertainty exercise (ζ = 1.10) cannot.

Customer-funded timing buys much more than a constant share. Research stays at 50 percent, so every capability arrives on its reference date and the rule costs the same in every continuation. Customer-funded rules lower the probability of catastrophe from 57.68 percent at a constant share to 52.14 percent with two stages, and to 51.14 and 51.01 percent with six and twelve capability segments. The twelve-segment rule puts more than a third of compute into safety at the start. It lowers the share in every segment until the share reaches its floor at F = 0.5.

At this budget customer-funded timing captures nearly all the gains of the tested joint rules. Under no redeployment the best twelve-segment rule that may also cut research gains 0.91 points more; under full redeployment it gains 0.002 points. This comparison uses the AI Futures Project’s worked example and the paper’s square-root valuation of services.

Shares by capability

Safety over customer compute

Left panel, shares. The safety (blue) and customer (green) shares of the customer-funded rule on the buttons, against capability; research stays at 50 percent. The gray dashed lines are the reference shares. The orange dotted line is the safety share of the best twelve-segment rule under no redeployment, which may also cut research.

Right panel, the ratio. Safety over customer compute along the same rule, against months, on a log scale. The gray dashed line falls at 10.46 percent a month, the least decline Proposition 4 allows on an interior stretch with flat productivity. The rule’s ratio falls faster until safety reaches its floor, and then stays there.

The implication is that early safety is cheap in services, because compute and quality are still low, while it protects almost all of the takeoff period’s hazard. The multistart rules are the paper’s saved solutions, evaluated here. They come from a restricted class, so they illustrate the timing result rather than test its interior condition.

The math

Part 4The numbers: whole rules, budgets and what moves the answer

Step 8 · Paper, Section 5.2, Table 4, Figure 1c

With known technology the continuation decides who pays

The continuation decides who pays, as Corollary 1 predicts. Under full redeployment the best constant rule is customer-funded safety itself. Delay would forgo what a finished frontier serving 95 percent of compute adds, so funding from research costs more per unit of hazard reduced.

Under no redeployment research pays. The best constant rule keeps customers at 45 percent and moves research to 40.7 percent and safety to 14.3 percent. That cuts the probability of catastrophe by 2.85 points, to 54.83 percent. Customers could not use more compute, so delay costs only quality.

Dynamic rules use both sources. Under no redeployment the best two-stage rule switches at F = 0.2. Before the switch customers give up most of the extra safety compute and research the rest; after it, safety falls to 7.8 percent. Under full redeployment the best two-stage rule is customer-funded: research stays at 50 percent, and safety compute moves toward the start of the takeoff period (Table 4 below). Across serving capacity the best two-stage rule beats the best constant rule by 3.77 to 6.66 points. From a serving capacity of 85 percent it never cuts research.

The reference reset, a third continuation, returns customers to the reference share after completion. Its best two-stage rule lowers the probability of catastrophe to 38.60 percent. But it nearly stops research after F = 0.5, gives customers 93.0 percent of compute and ends the takeoff period at month 27.60. The paper traces this gain to the continuation, in which a finished frontier serves customers less than a slowed one.

Best constant rule by serving capacity

Probability of catastrophe by rule

Table 4: known technology, who pays for safety compute, and when

Top panel, serving capacity. The research share of the best constant rule against serving capacity, for continuing danger ξ = 0, 0.25 and 1. At 50 percent research, customers pay for all extra safety. The rule cuts research below the threshold and keeps it at 50 percent from a capacity of 65 percent, or 70 percent when ξ = 1. The dot is the rule at the slider’s capacity.

Bottom panel, the rules. The probability of catastrophe of customer-funded safety, the best constant rule and the best two-stage rule in the two consistent continuations. All meet the 11.5 percent budget and the 12-month deadline. The table adds the shares, the service cost, the largest milestone delay and the reference reset.

The implication is that the funding answer of Corollary 1 survives the move from marginal comparisons to whole rules. The constant rules are solved in your browser; the two-stage rules are the paper’s saved multistart solutions, evaluated here.

The math

Step 9 · Paper, Section 5.3, Table 3, Figure 3

At larger budgets capability research pays too, mostly early in the takeoff period

Under full redeployment the best constant rule keeps research at 50 percent up to a budget of 28.8 percent. From 34.5 percent on it cuts research, to 48.3 percent at a 34.5 percent budget and 37.7 percent at 69.1 percent. Under no redeployment it cuts research at every budget, from 45.0 percent at a 5.8 percent budget to 20.1 percent at 69.1 percent.

Richer timing enlarges the gain from cutting research and moves the cut to the start of the takeoff period. At a 34.5 percent budget under full redeployment, the best twelve-segment rule cuts research most at the start and restores it to 50 percent from F = 0.15. Its largest milestone delay is 1.04 months. It lowers the probability of catastrophe 2.03 points below the best customer-funded twelve-segment rule, against 0.07 points among constant rules. At 69.1 percent the gain is 8.92 points; under no redeployment it is 7.78 and 15.53 points at the two budgets. An early cut delays every later capability, giving knowledge more time before all of them, while it prolongs exposure only where state risk is still low.

Lower safety productivity moves the threshold the other way. At the 11.5 percent customer-funded optimum it is 54.3 percent at a quarter of baseline productivity and 38.2 percent at a tenth. It falls below the no-redeployment share of 45 percent at 13.6 percent of baseline productivity. At a tenth of baseline productivity, the best constant rule under no redeployment keeps research at 50.0 percent: customers pay even though they cannot use more compute. Twice the baseline productivity raises the threshold only to 65.2 percent.

Best constant rule by budget

Threshold by safety productivity

Table 3: customer-funded rules and rules that may also cut research

Left panel, the budget. The research share of the best constant rule against the service budget, on the paper’s budget grid, under full redeployment (blue) and no redeployment (green). The circles mark the budget on the slider.

Right panel, productivity. The threshold at the 11.5 percent customer-funded optimum when safety productivity is scaled with the hazard functions held fixed, on a log scale. Below the dotted line at the no-redeployment share of 45 percent, customers should pay even when they cannot use more compute. The dot marks the slider.

The implication is that customer-funded safety still provides most of the reduction from full speed at larger budgets, but no longer nearly all of it. The table’s constant rules are solved in your browser; its two-stage and segment rules are the paper’s saved multistart solutions, evaluated here.

The math

Step 10 · Paper, Section 5.4, Figure 4

Three quantities move the threshold, and the answer flips inside the evidence

The first quantity is the delay a point of research costs. With research at x0, taking one unit of share from research delays completion by

ℓ(1) = (1 − e−γT0)/(G x0), γ = εG.

The elasticity ε and the share x0 set how much a unit of share slows the pace. A point costs 0.11 months at baseline, 0.09 with a 65 percent research share and 0.08 with ε = 0.3. The cheaper delay raises the threshold to 97.4 and 98.2 percent, above the full-redeployment share, so research should pay whatever customers can use. Costlier delay reverses this. With ε = 1 the threshold is 41.1 percent, and with a 35 percent research share it is 46.6 percent against a no-redeployment share of 60 percent; in both, customers should pay even when they cannot use more compute.

The length of the takeoff period barely moves the threshold. The delay grows with the length of the takeoff period, damped by compute growth: a point costs 0.07 months with a 9-month takeoff period and 0.15 with a 36-month one. But the budget is a share of the services over the takeoff period, so a longer period also raises what customer funding gives up, and every length buys the same customer-funded safety share. The threshold is 60.6 percent at 9 months, 63.5 at 18 and 60.9 at 36.

The second quantity is how much extra knowledge safety compute buys. When most safety learning needs no dedicated compute, customer-funded safety adds little knowledge, while slowing research still gives all learning more time. The threshold is 75.2 percent when three quarters of learning needs no compute and 116.0 percent at nine tenths. Weak returns to safety compute and a larger reference safety share work the same way: returns exponent α = 0.2 gives 95.0 percent, and an 18 percent safety share gives 91.1 percent, above that variant’s full-redeployment share of 82 percent.

The third quantity is the curvature of service value, which Step 11 re-solves in your browser. Short-run price elasticities of demand for language-model tokens are just above one (Demirer et al. 2025). With service value Qη priced at its margin the elasticity is 1/(1 − η), so an elasticity near one means η near zero, and Q0.1 puts the threshold at 59989 percent. These are short-run estimates; longer-run responses may be larger. Other parameters move the threshold less. A research speedup of 3 rather than 10 raises it to 78.6 percent and one of 30 lowers it to 53.9 percent; the pause target, the state share and timing of hazard, compute growth and the length of the takeoff period keep it between 58.9 and 67.3 percent, and the full-speed probability does not move it (Step 11).

The answer flips inside the ranges the evidence allows. At the 11.5 percent budget, research should pay whatever customers can use when the research share exceeds 64 percent or when ε is below 0.31. Customers should pay even under no redeployment when the research share is below 42 percent or ε exceeds 0.85. Estimated research shares lie on both sides of the first flip point: OpenAI put about three quarters of its 2024 compute spending into research and development (You 2025), Anthropic about 60 percent in 2025 (Emberson and Edelman 2026), and OpenAI about half in 2025 (Epoch AI 2026). Anthropic’s registration statement gives a bound from its financial statements: compute for paying customers was at most 38 percent of its compute spending in 2025 and 27 percent in the first half of 2026, so at least 62 and 73 percent went to research, model development, internal use and free users (Anthropic, PBC 2026b). These shares count dollars and include some safety work, and they move fast, while the model holds reference shares fixed. The 2026 bound lies above the first flip point, so on those accounts the rule points to research. Extra safety compute should then come from slowing capability work rather than from customers who are already rationed. The 2025 bound lay just below it. The bound counts free users and internal use with research, so a true research share below the flip point cannot be ruled out. Forecasts put the median takeoff period between 9 and 18 months (Lifland et al. 2026), and no takeoff length from 9 to 36 months flips the answer. So the paper gives a funding rule: the quantities that decide who pays and the values at which each reverses the answer.

Where slowing research is cheap, rules that also cut research gain more. Under no redeployment the best twelve-segment rule that may cut research beats the best customer-funded one by 0.91 points at baseline, 8.5 points with a 65 percent research share and 13.1 points with service value Q1/4.

The threshold as one parameter varies

Delay from one point taken from research

Figure 4 in numbers: each value of the chosen parameter

Left panel, the threshold. The funding threshold at the 11.5 percent customer-funded optimum as the parameter on the buttons varies, one panel of the paper’s Figure 4 at a time; the large dot is the baseline. Where the threshold lies above the blue dotted line, the full-redeployment share, research is the cheaper source whatever customers can use. Where it lies below the green dotted line, the no-redeployment share, customers are cheaper even when they cannot use more compute. In the research-share and takeoff panels a shaded band marks the estimated research shares or the forecast median takeoff periods; the takeoff panel stays between the two dotted lines at every length.

Right panel, the delay. Months of completion delay from one percentage point taken from research, from the closed form above. The takeoff length, the elasticity and the research share move it; the other parameters leave it at its baseline value and move the threshold through what safety compute buys or how services are valued.

The implication is that who should pay depends on numbers a regulator can look up: how much compute labs put into research, how responsive research is to compute, and how much safety progress needs dedicated compute. How long the takeoff period is expected to last matters little. Each variant refits the calibration targets, which takes the paper’s full pipeline, so the thresholds and gains are the paper’s saved results; the page computes the delay itself.

The math

Step 11 · Paper, Sections 5.4 and 5.5, Table 5

How services are valued moves the answer; the robustness checks do not

The third quantity of Step 10, the curvature of service value, can be explored here. Replace √Q by Qη and re-solve, holding the technology and the hazard functions fixed. With η = 1/4 the threshold is 239.9 percent, and the best constant rule cuts research to 37.9 percent even under full redeployment. With η = 3/4 the threshold is 42.9 percent, and the best constant rule keeps research at 50.0 percent even under no redeployment.

Curvature is a choice about how services are valued, not an estimate. The faster the value of services diminishes, the sooner taking compute from customers becomes costly, so the higher the threshold.

The tested changes in gradual redeployment and in how the hazard is built preserve the funding answer. If customers’ use beyond the reference share expands over 6 to 24 months after completion, the budget at which research becomes the cheaper source rises from 30.2 to between 34.2 and 48.4 percent of J̄. Across constructions of the hazard functions the 11.5 percent threshold stays between 61.8 and 65.3 percent. The reference probability moves risk levels only. At 25 percent, the lower end of its range, the threshold stays at 63.5 percent and the twelve-segment customer-funded risk is 13.76 percent, against 51.01 percent at 75 percent. A slower, 36-month takeoff period keeps the funding answer. The budget unit covers each takeoff period’s own services, so the 11.5 percent budget buys the same 14.77 percent customer-funded safety share at 18 and at 36 months; the threshold moves from 63.5 to 60.9 percent, and it reaches 95 percent at a budget of 34.0 percent, against 30.2 at baseline. These checks are the paper’s; the page does not recompute them.

The slider of the figure below moves the full-speed catastrophe probability. It rescales the hazard and leaves the rules and the threshold where they are, so only the risk levels change: a rescaled risk is 1 − (1 − P)a, where a is the new cumulative hazard divided by the 75 percent one. The paper reports 25 percent as the lower end of its range, and values below it are this page’s own rescaling. The paper’s Table 8 also runs 10 and 5 percent, with the twelve-segment customer-funded risk at 5.28 and 2.61 percent and the threshold unchanged, so nothing in this paper breaks at lower levels. What does break is in two companion papers: the Labs calibration of racing at 50 percent research needs a full-speed probability above about 12 percent, and the Agreements game at the scenario’s own risks has no supportable review date at the baseline monitoring cost below about 25 percent.

Threshold by curvature

Best constant rule by curvature

Table 5: the funding answer and the curvature of service value

Left panel, the threshold. The funding threshold at the customer-funded constant rule against the curvature η of service value Qη, each with a budget of 11.5 percent of its own unit, the reference’s takeoff-period services under that curvature. The dotted lines mark the full-redeployment (95 percent) and no-redeployment (45 percent) shares. The dot marks the slider.

Right panel, the rules. The research share of the best constant rule against η, under full redeployment (blue) and no redeployment (green). At 50 percent, customers pay for all extra safety.

The implication is that the faster the value of services diminishes, the higher the threshold and the more research should pay. With nearly linear services, customers should pay even when they cannot use more compute.

The math

Part 5Commitment, predictions and limits

Step 12 · Paper, Section 6

A rule fixed before the technology is known costs little in the tested cases, unless safety work advances capability

The regulator does not know how safety research works when it commits. The paper considers 13 technologies. They vary safety productivity and the transfer of knowledge to stronger models. They also vary whether safety compute scales with installed compute, whether it also advances capability, and whether danger continues after completion. The regulator commits to a constant rule that meets the budget and the deadline in every technology. It picks the rule with the smallest largest excess probability of catastrophe over an informed choice, the minimax-regret rule.

In the ten technologies where safety compute does not advance capability, the chosen rule is within 0.28 points of the informed choice on the same share grid, in every continuation and budget considered. At the 11.5 percent budget the bound is 0.15 points. Across 12 variants of the parameters that move the funding threshold (Step 10) it is at most 1.13 points, reached with a 9-month takeoff period, while the regret from spillover is 7.2 to 10.7 points in every variant. Without spillover the affordable rules are known in advance, because a rule’s capability path, service cost and delay do not depend on the safety technology. Within that set the ranking of rules barely moves in the tested technologies. The first fact is structural; the second is a numerical finding for constant rules. The bound compares grid rules with each other: a feasible rule off the grid can beat the grid’s informed choice by more than the bound.

The remaining regret comes from spillover. When safety compute also advances capability, a rule delays less and costs fewer services, and an informed regulator would spend the slack on more safety. Under full redeployment, knowing whether spillover is present would cut the 11.5 percent worst regret from 5.90 points to zero; knowing whether productivity is weak would leave it at 5.90. Under no redeployment the fixed rule cuts research to 41.0 percent, with a worst regret of 4.86 points against 8.92 for the best customer-funded rule. At larger budgets the fixed rules cut research in both continuations: under full redeployment the 69.1 percent rule cuts research to 36.5 percent.

Robustness does not mean low risk. Under full redeployment the 11.5 percent rule leaves the probability of catastrophe at 99.25 percent when danger continues after completion, and at 98.06 percent in the combined weak-productivity, transfer and spillover state.

Excess risk over the informed choice, on the 11.5 percent budget frontier

The panel. Constant rules on the 11.5 percent budget frontier: for each research share, the largest affordable safety share. Each colored curve is one technology. It gives each rule’s probability of catastrophe minus that of the best rule on the frontier in the same technology, in percentage points. The black curve is the largest of them, and its lowest point is the rule a regulator who does not know the technology would choose. The page computes six of the paper’s technologies, those that vary safety productivity or continuing danger. It does not compute the other seven: four change how knowledge grows or transfers, and three add spillover.

The implication is that, without spillover, the best rule sits at almost the same research share in every technology, so committing before the technology is known costs little. The paper’s worst regrets, which include spillover, are larger.

The math

Step 13 · Paper, Sections 7 and 8

What the model predicts and what it leaves out

A joint allocation can be implemented as compute shares that change at an observable capability milestone. Enforcement needs a meaningful distinction between inference, safety research and capability research. The White House Accord on Super Intelligence of September 2026 sets up monitoring, internal teams, independent auditors and board oversight, but it sets no compute for safety research and does not say who gives it up. Auditing a rule from this paper would mean checking how compute is split among the three uses and whether safety work also advances capability.

Labs already pay for some safety by slowing research, but they do not yet keep the accounts an audit would need. Anthropic’s registration statement says it has delayed releases and limited the initial availability of models so that safety measures keep pace with capabilities, and that it plans compute as one pool rather than by workload (Anthropic, PBC 2026b); its estimate that about 6 percent of research compute went to safety rests on best-effort labels for one week (Anthropic 2026a). The same filing says that demand for compute has historically exceeded available capacity and that the lab is “first and foremost a research lab” (Anthropic, PBC 2026b). Together these suggest that it sells less inference than customers would buy to keep compute in research, so it values research compute above the revenue that compute would earn, a valuation that includes staying ahead of rivals. Its 2026 accounts put at least 73 percent of its compute spending outside paying customers, and on those accounts the rule points to research, so a regulator who also counts catastrophe risk would fund extra safety by trimming that research share rather than by rationing customers further. The model makes three conditional predictions.

  1. Who pays for safety compute should track how much freed compute can serve customers when the frontier finishes and the size of the service budget, given how services are valued and given the safety technology. With a small budget, a frontier whose freed compute can serve customers at once should fund extra safety from customers. One whose freed compute cannot serve customers should also slow capability research. Either comparison can reverse with the valuation of services, the share of compute in research, how fast the research pace responds to compute, or the productivity of safety research. Large enough budgets should slow research in both, mostly early in the takeoff period.
  2. Where both margins are interior, safety compute relative to inference should fall over a takeoff period, unless safety productivity rises fast.
  3. A rule that frees inference compute is worth what customers can use; idle capacity adds nothing to its service value.

The analysis is limited in six respects. The numbers come from one conditional scenario whose hazard functions are assumed. Services are an index, not revenue, so the model says nothing about who bears the monetary cost. One frontier lab is regulated, the deadline stands for a rival whose response is not modeled, and the lab has no private objective. Installed compute grows exogenously. The rules come from restricted classes and are not certified global optima. And the results do not rank instrument families: the regulator minimizes hazard under a service budget and a deadline and imposes no risk ceiling.

The funding and commitment comparisons need a common continuation. A model in which a finished frontier serves customers less than a slowed one turns the endpoint term into a saving, and it then credits safety research with gains that belong to the continuation.