Who Should Give Up Computing Power for AI Safety?

Two ways to fund safety

Li Gan’s arXiv preprint, in its second version revised on 5 November 2026, studies a regulated AI laboratory dividing computing power among capability research, safety research and customer use. It is a theoretical allocation exercise with numerical scenarios, not an empirical estimate of AI catastrophe risk.

The regulator minimizes catastrophe risk while limiting lost customer services and the delay allowed at every capability milestone. The laboratory’s installed computing capacity grows along a predetermined path; regulation changes its allocation, not investment.

Taking compute from customers funds safety without changing capability dates in the baseline model. Taking it from capability research delays stronger models, giving safety knowledge more time to accumulate. But waiting also extends exposure to dangers associated with using models at a given capability. Slower development is therefore not automatically safer.

Both routes cost customers something: either less service now or better models and additional serving capacity later. The comparison measures those losses in a discounted service index, not money. It does not establish who pays the financial bill.

Why finishing matters

The paper’s central mechanism is the value of completing the modeled development period. Once the endpoint is reached, capability research is no longer needed. Can its computing power immediately serve customers?

Under the baseline allocation, capability research receives 50 percent of compute, safety 5 percent and customers 45 percent. If all released research compute can serve customers, their share rises to 95 percent at completion. Delaying that expansion is costly. If customers cannot use the released capacity, they remain at 45 percent, making delay less expensive.

At the baseline constant customer-funded allocation, Gan calculates a break-even post-completion customer share of 63.5 percent. Above it, customers are the cheaper marginal source of safety compute; below it, capability research should also contribute. This is a local comparison at a particular allocation under a common post-completion arrangement, not a universal regulatory target.

Earlier safety buys longer protection

Timing matters because the model treats safety research as building knowledge that reduces subsequent danger. Early knowledge protects more of the remaining development period. Meanwhile, growing compute and improving models make later customer services more valuable.

Under the baseline assumptions, the optimal safety-to-customer compute ratio falls on intervals where neither allocation is constrained. This does not prove that safety’s share must fall everywhere: research allocations can change too, and floors or capacity limits can bind. Sufficiently fast growth in safety productivity can reverse the timing result.

The direction also depends on modeling choices. Baseline learning depends on compute shares rather than absolute computing capacity, and danger ends at the development endpoint. The paper examines extensions to both assumptions.

What the numerical results mean

The baseline assumes an 18-month development period, a 12-month maximum milestone delay and a service-loss budget equal to 11.5 percent of reference services during the development period. That budget is not 11.5 percent of revenue or five-year output.

Its risk assumptions come from an AI Futures Project worked example. Full-speed catastrophe probability is set at 75 percent, and safety productivity is fitted to a judgment that an initial 12-month pause would reduce it to 40 percent. These inputs are conditional judgments, not measured probabilities.

A constant customer-funded safety allocation reduces modeled catastrophe probability to 57.68 percent. Allowing safety shares to change across twelve capability segments reduces it to 51.01 percent without delaying capability. Under the alternative where released compute cannot serve customers, also permitting research cuts lowers risk to 50.10 percent—an additional 0.91 percentage points.

With a 25 percent starting probability and a proportionally matched hazard reduction from the calibration pause, front-loaded customer-funded safety produces 13.76 percent risk. The funding threshold remains unchanged. These results describe the assumed model; they are not forecasts of actual policy effects.

Conditional guidance, not a settled prescription

Larger service budgets favor research funding in the calibration: as customers surrender more compute, each additional unit becomes costlier to remove. But research responsiveness, safety learning and the valuation of services can reverse the funding answer.

Gan also finds small costs of committing before safety technology is known in tested constant-share comparisons without capability spillovers. When safety work also advances capability, uncertainty matters more because it changes which rules are affordable. That finding is relative to the same restricted share grid, not unrestricted policy choices.

The analysis regulates one lab, substitutes a deadline for rivals’ behavior and assumes observable, enforceable allocations. Its numerical candidates are not certified global optima. Its contribution is a conditional accounting framework: compare safety gains with all customer services sacrificed, including those lost when completion arrives later.