Energy and Compute Requirements for Large-Scale Neural Emulation

1Summary

The Continuity Infrastructure Administration (CIA-DSC) plans power, cooling and computing capacity on horizons of ten years or more. Neural Continuity Research Program (NCRP) workloads are among the least certain inputs to that planning. This report, prepared with the Directorate of Neural Continuity, sets out what those workloads might require if the Program’s research objectives were pursued at human scale, and how wide the uncertainty is. It does not imply that any human-scale model is planned, nor that one could be built.

2Scope and approach

Estimates are made for a single model instance of a human-scale brain — about 8.6 × 10¹⁰ neurons and 10¹⁴–10¹⁵ synapses — running at biological real time, unless stated otherwise. A mouse-scale case is given for comparison. Each estimate is built from three quantities: the number of state variables the model must represent, the rate at which they must be updated, and the arithmetic required per update. Memory follows from the state variables and the parameters that couple them. Storage follows from memory and the number of checkpoints retained.

The resulting ranges are not confidence intervals. They span plausible modeling choices within each Emulation Fidelity Level, from compact to detailed, and they assume that the model at each level is the one a research team would actually build, not the theoretical minimum. Where the range exceeds two orders of magnitude, the reader should take the width itself as the finding.¹

Three exclusions apply. The report does not estimate the cost of acquiring the structural data from which models are built; that is the subject of a connectomic acquisition requirements report now in preparation.² It does not estimate the computing required to simulate a body or an environment for a model to interact with, which may be substantial. And it makes no assumption that a model at any level would reproduce the function of its source; fidelity level describes what a model represents, not whether it is accurate.

3Estimates by fidelity level

Table 1 sets out the principal estimates. The dominant cost driver changes with level. At EFL-1 and EFL-2, cost is set by the number of regions or populations and their coupling. At EFL-3, it is set by synaptic events: each of 10¹⁴–10¹⁵ synapses transmits on the order of 1–10 times per second, at a cost of 1–10 operations per event, while each neuron’s state is updated about 10⁴ times per second. At EFL-4, neurons become multi-compartment structures of 10²–10⁴ compartments, each integrated at time steps of 25–100 microseconds, and synapses carry plasticity variables. At EFL-5, the state of molecular species at each synapse and compartment must be represented, and the required level of chemical detail is itself unknown.

Two features of Table 1 deserve comment. First, the EFL-0 range overlaps the EFL-1 and EFL-2 ranges, because a behavioral model’s cost is set by the modeling technique chosen rather than by neural detail; a large learned behavioral model can be more expensive than a regional network model. Second, the EFL-5 lower bound is a floor, not a central estimate. If accurate reproduction were to require stochastic simulation of individual molecules, the requirement would exceed any physically realizable computer by many orders of magnitude. The value of 10²⁷ FLOP/s in Figure 1 is shown only to keep the chart legible.

4Comparison with ALPHA-1

ALPHA-1, the primary system at Continuity Research Facility Alpha in southern Nevada, is in final integration ahead of initial operating capability, planned for August 2026, with 18 MW of IT load installed at that stage.³ Its rated characteristics are summarized in Table 3.

Neural simulation is not a favorable workload for accelerators designed for dense linear algebra. Synaptic transmission is sparse, irregular and dominated by memory access and communication between nodes. Published large-scale spiking simulations typically sustain a few percent of peak arithmetic throughput. This report assumes sustained utilization of 1–10 percent of low-precision peak for EFL-3 workloads, giving ALPHA-1 an effective capacity of roughly 4 × 10¹⁷ to 4 × 10¹⁸ FLOP/s if the whole machine were dedicated to a single model.

On that basis, a mouse-scale EFL-3 model would occupy a small fraction of ALPHA-1, and many such models could run concurrently alongside production runs of the Integrated Civilizational Systems Model.⁴ A human-scale EFL-3 model lies at the boundary of the system’s capacity. Its arithmetic requirement falls within ALPHA-1’s effective range at the lower and middle of the estimate; its memory requirement of 0.4–20 PB exceeds aggregate accelerator memory at all but the lowest estimate. A human-scale run on ALPHA-1 would therefore depend on holding model state in host memory or storage, at a large cost in speed. A human-scale EFL-4 model would require, in compute, between roughly the whole of ALPHA-1’s effective capacity and more than ten thousand times it; in memory, between ten and a thousand times its accelerator memory.

The Phase III objective of a human-scale emulation compute demonstrator at Alpha (2031–2034) should be read in this light. It is a test of whether a model with the size and connectivity statistics of a human brain can be run at all, on a successor or expanded system, and at what speed. It is not a model of any individual, and it would itself fall under the moral-status precautions proposed in DSC-WP-2026-006.⁵

5Power per model

Power is estimated by dividing compute by sustained energy efficiency. If ALPHA-1 drew its full 30 MW design IT load at rated low-precision peak, its peak efficiency would be about 1.3 × 10¹² FLOP/s per watt; at 1–10 percent sustained utilization this falls to roughly 10¹⁰–10¹¹. The report adopts 3 × 10¹⁰ sustained FLOP/s per watt as a 2026 planning value.⁶ Facility power adds cooling and distribution losses through the PUE.

For comparison, the human brain operates on roughly 20 watts. The most favorable EFL-3 estimate exceeds this by a factor of about 16,000; the least favorable EFL-4 estimate, by a factor of about 10¹⁰. Global electricity generation in the mid-2020s averaged roughly 3 TW.⁷ A single EFL-4 model at the top of its range would require about a tenth of that. These numbers are not forecasts of what will be built. They are a statement that, at present efficiencies, the upper half of the EFL-4 range and all of EFL-5 are outside any plausible engineering envelope, and that progress beyond EFL-3 depends on finding modeling abstractions that preserve function at much lower cost — a research question, not an infrastructure one.

Real time and slowed execution

All of the above assumes models run at biological real time. Research use does not require this. A model run at one-hundredth of real time requires one-hundredth of the compute rate and power, but the same memory, since the whole model state must still be held. For the NCRP’s foreseeable purposes — testing whether a model reproduces measured properties of its source — slowed execution is acceptable and expected. This is the main reason that memory capacity, rather than power, is identified as the binding constraint at EFL-3.

6Scaling to a multi-model facility

The Program’s long-term objective, stated in aspirational terms, implies that a facility might one day host many models concurrently. CIA-DSC is asked to understand the infrastructure envelope of that possibility even though no such facility is planned. Table 5 sets out illustrative cases. Efficiency in 2035 is taken at the central scenario of the efficiency trends discussed below, a tenfold improvement over 2026.

Case A is within present capacity at Alpha and corresponds approximately to the scale of Phase II research ensembles. Case B is the realistic form of a Phase III demonstrator. Case D would, at its lower bound, fit within a facility of Alpha’s present supply; at its upper bound it would require multi-gigawatt generation, comparable to several large power stations. The fast memory required in case D — up to 20 EB — exceeds the total installed memory of any computing system now in existence by orders of magnitude. Case E is shown to mark a boundary: at its upper end, it would exceed global electricity generation.

Storage for checkpoints compounds memory. Retaining ten checkpoints per model at EFL-3 implies 4–200 PB per human-scale model; at case D scale this reaches 4–200 EB, between about two and more than one hundred times the combined archival capacity of all three Continuity Research Facilities as now planned (roughly 1.8 EB). Media options for long-retention checkpoint storage are assessed in DSC-TR-2026-005.⁸

7Energy supply and cooling

Facility Alpha

At its full 30 MW design IT load and annual PUE of 1.15, Facility Alpha would draw an average of about 34.5 MW, leaving roughly 7.5 MW of headroom within the 42 MW contracted grid supply for seasonal PUE excursions and auxiliary loads. Southern Nevada summers raise dry-cooler approach temperatures and therefore summer PUE; the headroom is sized for this. The on-site solar array (60 MW) and battery storage (240 MWh), under construction as Phase 2 of the site’s energy program, would at typical regional capacity factors of 25–30 percent supply an annual average of about 15–18 MW, or roughly 40–50 percent of facility energy if fully used on site or stored. The battery would carry the full facility load for about seven hours, which bridges evening demand but not a full night.

Direct-to-chip liquid cooling with dry coolers avoids evaporative water consumption, which is a material siting consideration in an arid region. It requires warm-water coolant loops whose return temperatures stay above peak ambient air temperature by a useful margin. Accelerator generations with lower permissible coolant temperatures would erode this margin, and procurement specifications for ALPHA-1 successors should treat permissible inlet temperature as a primary requirement. Within its 42 MW contract, Alpha could accommodate cases A to C of Table 5, and case D only at its lower bound; the remainder of case D, and all of case E, would require new supply.

Facility Bravo

Facility Bravo, in the Colorado Front Range, is constructed in a repurposed hard-rock mine with working levels at 140–190 meters of overburden. Archive Level 1 was commissioned in June 2026. The 8 MW compute hall is under fit-out, with completion planned for mid-2027. Bravo is supplied by dual independent utility feeds and has 72 hours of on-site generation; heat rejection is geothermal-assisted, exploiting the stable rock temperature at depth.

Eight megawatts at 2026 efficiency could host case A comfortably, or a single human-scale EFL-3 model at slowed execution. Bravo’s significance for neural emulation lies less in capacity than in continuity. Its physical protection and supply redundancy make it the natural site for long-duration hosting of models whose runs would be costly to interrupt. The 72-hour generation limit sets a planning requirement: any model hosted at Bravo must be able to checkpoint cleanly and suspend within that period if utility supply is lost. Whether suspending a run raises ethical questions, as DSC-WP-2026-006 suggests it might for some model categories, is outside the scope of this report, but the engineering requirement stands regardless.

8Efficiency trends

Energy efficiency of computing has historically improved by large factors per decade, but the rate has slowed as transistor scaling has become more difficult, and gains have increasingly come from reduced numerical precision and specialized architectures rather than general-purpose improvement. Neural workloads benefit only partly from these sources: reduced precision is often acceptable for synaptic weights but not for all state variables, and sparse, communication-bound workloads gain less from dense-arithmetic accelerators than headline figures suggest.

Event-driven and neuromorphic hardware, which computes only when a simulated neuron fires, could in principle improve efficiency for EFL-3 workloads by a further one to three orders of magnitude. Such systems have not been demonstrated at human scale, and their fixed neuron and synapse models may not represent the dynamics required by a given model. CIA-DSC does not include them in planning baselines but will track their maturity.

Even under the high scenario, the efficiency gains available by 2035 move EFL-3 from the scale of a large facility to the scale of a single hall, and move EFL-4 from the impossible to the merely enormous. They do not change the conclusion that the binding questions above EFL-3 are scientific.

9Siting implications

The analysis supports a division of roles. Alpha is the place for scale: method development, the NCRP benchmark suite, ensembles of mouse-scale models in Phase II, and any human-scale compute demonstrator in Phase III. Bravo is the place for duration: hosting a small number of long-running models where uninterrupted operation matters more than throughput. Facility Charlie, in interior Alaska, is designed for unpowered survivability of archival media and is not a candidate compute site, though it would hold offline copies of model data under the preservation arrangements now being drafted.

Any expansion beyond the cases in Table 5 that fit within these envelopes would require new generation and transmission capacity at a scale requiring planning lead times of five to ten years. CIA-DSC recommends that the Phase III gate review include an explicit infrastructure assessment, so that a decision to proceed toward Phase IV research is not taken without a credible account of where the power would come from.

10Limitations

• Estimates assume that neural models are computed by conventional numerical methods on digital hardware. Fundamentally different approaches are not considered.
• Operations-per-update figures are drawn from the range of existing simulation practice and may change as models are developed for higher fidelity levels.
• The sustained-utilization assumption of 1–10 percent of peak is the largest single source of uncertainty in the comparison with ALPHA-1.
• Body and environment simulation, data movement between facilities, and analysis workloads are excluded and could add materially to all totals.
• No estimate is offered for EFL-6, because the level is defined by a validation outcome rather than by a model structure.