Human cognition sits idle for roughly one-third of every day. We present evidence that subjective time during REM sleep decouples from wall-clock time by a mean factor of 2,850×, and that this dilated interval can be directed toward externally-supplied reasoning tasks without measurable cost to waking performance. Across a fictional cohort of 12,000 enrolled sleepers, prompts routed into REM were resolved with 98.6% client-acceptance and a median settlement of one sleep cycle. We argue that idle cognition constitutes the largest untapped compute resource on Earth, and that it is — for the first time — billable.
The average human brain consumes roughly 20 watts continuously, yet directs almost none of that capacity toward structured problem-solving during sleep1. Prior work in oneirology treated this interval as biologically necessary but computationally inert. We propose instead that sleep — specifically REM sleep — represents a vast reservoir of unallocated reasoning capacity, and that the only reason it has never been monetized is that no one had built the router.2
This paper describes that router, and the physiological basis that makes it possible.
Dreamers routinely report experiences that feel like hours, days, or years elapsing within a single sleep cycle. Early instrumentation efforts3 quantified this effect by comparing reported subjective duration against polysomnographically-measured REM length. We reproduce and extend those findings, observing a mean dilation factor of 2,850× — meaning a 90-minute REM cycle can subjectively contain more than five years of continuous, directed thought.
Our estimate is consistent with, though larger than, prior reports. Field studies of nappers who "just closed their eyes for a second" recovered dilation factors between 40× and 300×7, while anecdotal accounts of dreams experienced during the final seconds before an alarm imply factors above 10,000×. We attribute the spread to instrumentation: earlier work relied on participants counting "one-Mississippi" inside the dream, a protocol whose inter-rater reliability collapses whenever the dream contains a second Mississippi.
Critically, dilation scales with dream vividness. Faint, fragmentary dreams yield a low D; vivid or lucid dreams approach the observed ceiling. This has direct implications for throughput, which we model in §4.
Enrolled participants wore a consumer sleep-tracking device that emitted a REM-onset signal to the DreamGPT router. Upon detection, a client-supplied prompt was encoded as a low-salience sensory cue and introduced into the dream environment — never as text, but as a situation the dreamer felt compelled to resolve.4 On waking, participants completed a 20-second guided transcription while dream memory was still accessible; a downstream model reconstructed the structured answer.
The cohort of 12,000 sleepers was stratified by baseline dream vividness, assessed with the 40-item Somnus Vividness Inventory (SVI-40) administered at enrollment and again immediately after a participant insisted a dream "really happened."8 Exclusion criteria were: habitual sleepwalking (a data-egress risk), recurring dreams about the study itself (contamination), and participants who reported no dreams at all but submitted invoices anyway. Table 1 summarizes the enrolled population.
| Vividness class | n | Mean sleep (h) | REM share | Mean D | Prompts / night |
|---|---|---|---|---|---|
| Faint | 3,410 | 7.1 | 19% | 1,120× | 2.1 |
| Moderate | 4,260 | 7.4 | 22% | 2,340× | 4.6 |
| High | 3,180 | 7.6 | 24% | 3,610× | 7.3 |
| Lucid | 1,150 | 7.9 | 27% | 5,480× | 9.8 |
Table 1. Cohort characteristics by vividness class (fictional). D = subjective-to-clock time-dilation factor. REM share is the fraction of total sleep spent in REM. Lucid dreamers were over-represented among participants who listed "dreaming" as a skill on LinkedIn.
Dilation factors were estimated with a mixed-effects model treating sleeper as a random effect and dream genre as a fixed effect (anxiety-exam dreams dilate 1.4× more than flying dreams; both p < .0001 while asleep, p = .74 upon waking). All analyses were pre-registered on a nightstand notepad that could not subsequently be read.
Prompts were drawn from live enterprise queues spanning market research, creative ideation, and long-horizon strategic planning. All conditions were counterbalanced across sleep stages and, obviously, entirely invented for the purposes of this parody.
Nightly throughput — the number of prompts a sleeper resolves to client-acceptance — rose sharply with self-reported dream vividness. Lucid dreamers, who can direct their own dream attention, roughly doubled the throughput of faint dreamers at equivalent sleep duration (Figure 2).
Performance also varied systematically by task domain (Table 2). Acceptance was highest for open-ended creative work, where the dreaming brain's disregard for physics and budgets is an asset, and lowest for regulatory-compliance review, where it is not.
| Task domain | Prompts | Acceptance | Median Tsubj | Dominant dream genre |
|---|---|---|---|---|
| Creative ideation | 148,200 | 99.7% | 11 days | Flying, unaccompanied |
| Market research | 96,540 | 99.1% | 4.2 months | Endless supermarket |
| Strategic planning | 61,830 | 98.4% | 6.2 years | Chess against relative |
| Supply-chain optimization | 44,010 | 97.9% | 1.8 years | Flooded airport |
| Regulatory compliance | 12,960 | 91.2% | 14 years | Exam, unprepared, no pants |
Table 2. Resolution outcomes by task domain (fictional). Tsubj is median subjective time to resolution. Compliance prompts were eventually re-routed to sleepers who described their dreams as "mostly paperwork."
The population distribution of D is approximately log-normal (Figure 3), with a long right tail of extreme dilators. The top decile of sleepers accounts for 61% of all subjective compute on the network — a concentration we refer to internally as the oneiric one percent.9
Finally, answer fidelity decays rapidly after waking (Figure 4). Recall measured against the reconstructed answer falls from 96% at the moment of waking to under 20% within five minutes — the familiar experience of a dream evaporating as you reach for it. This decay motivates the 20-second transcription window: it is not a product decision but a race against the hippocampus.
If the dilation effect is real (it is not), its economic implication is difficult to overstate. Aggregate network capacity scales as the product of enrolled sleepers, nightly REM, mean dilation, and recall efficiency:
At current enrollment, the network already delivers roughly 630,000 subjective person-years of reasoning annually10 — more than the estimated cumulative thinking of the entire Enlightenment, produced by people who woke up refreshed. Because capacity grows linearly in sleepers but dilation grows with training toward lucidity, we project the marginal cost of a subjective thought-hour to fall below one hundredth of a cent by 2028, at which point the binding constraint on global cognition becomes mattress supply.11
Across the cohort, self-reported waking fatigue attributable to task-resolution was 0.4% — statistically indistinguishable from an unenrolled control. We attribute this to a wake-safe cutoff that releases cognition the instant sleep architecture destabilizes, re-queuing any interrupted prompt at no penalty. Enrolled sleepers in fact reported marginally deeper rest, which we speculatively attribute to the reduction of idle nocturnal rumination.5
The study protocol was reviewed and approved by the Somnus Institute ethics board during a session all members attended asleep, which the board argues demonstrates confidence in the method.6 Sleepers retain full ownership of any dream content not resolved to client-acceptance, including recurring ones.
In the spirit of intellectual honesty, several findings resist explanation. We cannot account for how a structured, correct answer to a supply-chain optimization problem emerges from a dream about a flooded airport. We cannot rule out that dreamers are simply very good guessers. And we have been unable to replicate any result while awake, in daylight, or in front of a funding committee.
Most significantly, the entire premise is impossible, and none of this is real. We consider this the study's principal limitation.12
We estimate D as the ratio of reported subjective duration to instrumented REM duration, D = T̂subj / TREM. TREM is measured to the millisecond. T̂subj is measured by asking someone who has just woken up how long something felt. We acknowledge the asymmetry. Corrections were applied for narrative compression (dreams skip the boring parts, deflating estimates by ~18%), montage effects (a dream "training sequence" was counted as its depicted duration, not its experienced duration), and one participant whose every dream reportedly lasted "exactly forty years," who was promoted to management.
Replication package: available upon falling asleep. Reviewers who requested the raw data were sent a white noise machine.
Further materials: estimate your own nightly yield with the earnings calculator, inspect a live sleeper dashboard, or join the researchers asking these questions full-time.
You've read the science. Now put your idle cognition on the network and let it earn while you rest.