PaperlayerView exact revision on arXiv

Source-linked AI summary

To Store or To Regenerate? A Cost Model for AI-Generated Content at Scale

Yunjia Zheng, Zirui Wang, Haoran Ni, Tingfeng Lan, Zhaoyuan Su, Yue Cheng, Juncheng Yang

arXiv:2609.30448v1cs.PF

TL;DR

Growing AI-generated corpora make it important to know when regeneration costs less than persistent storage. The paper models lifecycle costs for storage and regeneration, including caching and compact intermediate representations. It finds that prompt-based regeneration remains much more expensive, while IR-based regeneration shifts the break-even point to today.

  • Problem

    The paper asks when on-demand regeneration becomes cheaper than storing generated artifacts, given that corpus growth, storage costs, compute costs, and access patterns all affect the break-even point.

    Source: §1 · p. 2

  • Method

    The paper builds a lifecycle cost model for materialized storage and regeneration, including caching and storing compact intermediate representations for lightweight decoding.

    Source: §1 · pp. 2–3

  • Results

    Prompt-based regeneration is two orders of magnitude more expensive than materialized storage, while LazIR cuts costs by 2.6× to 10.5× across image and music workloads.

    Source: §10 · p. 21

  • Takeaways & Limitations

    Storing intermediate representations and decoding on a cache miss moves the break-even point from decades away to today.

    Source: §10 · p. 21

    The model assumes current cost and technology trends persist and assumes skewed access that allows caching to reduce regeneration cost.

    Source: §9 · p. 20

Abstract

from arXiv · show

AI-generated content is becoming a rapidly growing class of digital artifacts. Because these artifacts accumulate over time, their exponential growth creates a substantial storage, energy, and infrastructure cost for operators and society. At the same time, GPU compute cost continues to fall rapidly with each hardware generation. This divergence raises a fundamental question: when does on-demand regeneration become cheaper than persistent storage? This paper develops a cost model for comparing persistent storage and on-demand regeneration for AI-generated artifacts. The model accounts for corpus growth, HDD and tape price trends, drive replacement, electricity, request skew, caching, generator FLOPs, and future GPU price-performance improvements. For image generation, our analysis shows that prompt-based regeneration does not become cheaper than storage until around 2040, because every cache miss must still rerun the full prompt-to-artifact generation pipeline. We observe that widely used diffusion-based generation models operate in latent space, creating an alternative point in the cost tradeoff: instead of storing the final artifact or only the prompt, operators can store a compact intermediate representation (IR) and perform cheap on-demand decoding. Our analysis shows that caching combined with IR-based regeneration substantially reduces both storage and compute cost, making it at least 2x cheaper than both full-object storage and prompt-based regeneration even today. On a production image trace with 2.07 billion requests, the same conclusion holds: prompt-based regeneration is over 100x more expensive than storage, while IR-based regeneration reduces total cost to roughly half that of full-object storage while preserving interactive miss latency.

1 Introduction

AI-generated artifacts accumulate into long-lived corpora, while storage and infrastructure costs build over time and GPU compute costs decline. This paper models when regeneration can beat materialized storage and finds that retaining compact intermediate representations (IRs) can make regeneration economical today.

1 Introduction

  • Long-lived generated corpora incur costs beyond disk purchases, including power, cooling, drive replacement, and archival infrastructure.Tape media is cheaper and uses little power while idle, but libraries, robotics, maintenance, and high access latency constrain its use for interactive serving.Source: §1 · pp. 1–2
  • The model jointly accounts for corpus growth, HDD and tape costs, electricity, replacements, access patterns, caching, generator FLOPs, and future GPU price-performance.It provides a break-even threshold for deciding whether to retain artifacts or regenerate them on demand.Source: §1 · pp. 2–3
  • A compact IR can preserve costly upstream generation work, letting systems skip expensive stages and perform lightweight decoding instead.The paper evaluates this approach across image and music workloads and on a production image trace.Source: §1 · p. 3
  • IR-based regeneration is cheaper than materialized-artifact storage starting today regardless of retention duration, while prompt-based regeneration crosses over later.For five-year retention, the figure places the prompt-based crossover at 2046; the IR-based design reduces regeneration cost by roughly 2.6×–4.0×.Source: Figure 1 · p. 2, §1 · p. 3

2 Generated-Data Growth and Computing Cost Trends

The paper characterizes generated-data growth alongside storage, compute, and energy trends. Its fitted trends show GPU cost and energy per unit of computation improving faster than storage cost and HDD power density.

2.1 Growth of AI-Generated Data

  • AI-image production has expanded rapidly: an estimated 15 billion images were generated in 2022–2023, compared with 149 years for photography to reach that volume.Daily AI-image requests rose from a few million to roughly 80 million by 2026.Source: §2.1 · p. 3

2.2 Storage and Compute Cost Trends

  • HDD and tape prices fall by roughly 10% annually, versus GPU $/TFLOPS declines of about 30% for FP16 and 41% for low-precision computation.The paper adopts a more conservative 30% annual FP16 decline for its forward model.Source: Figure 2 · p. 4, §2.2 · p. 4

2.3 Hardware Energy Consumption

  • Energy required per unit of storage and computation declines at rates similar to the respective price declines.This relationship is the empirical basis for the model’s energy trends.Source: Figure 3 · p. 4, §2.3 · p. 5
  • GPU FP16 energy per TFLOP falls about 31% annually, while HDD power-density gains flatten to roughly 7% per year after helium drives arrive.The paper uses fitted energy-improvement rates so electricity consumption evolves consistently with the modeled price trends.Source: §2.3 · p. 5

3 Cost Model 1: Materialized-Artifact Storage

The storage model compares retaining every artifact on HDD or tape over a growing corpus, including medium-specific purchase, replacement, infrastructure, and electricity costs. Across both linear and exponential growth, tape remains the cheaper cold-storage option once those costs are counted.

3 Cost Model 1: Materialized-Artifact Storage

  • The baseline retains every artifact over 2025–2045 under either constant 1 PB annual ingest or ingest growing at rate g, and counts purchases, HDD replacement, and power.The running image example uses 3 MB per image, 20% annual ingest growth, and a five-year HDD service life.Source: §3 · p. 5, §3.1 · p. 5

3.2.2 HDD Replacement over the Corpus Lifetime.

  • The HDD model prices new data and replacement capacity at each purchase year’s declining HDD price, while distinguishing live stored volume from total manufactured drive volume.A five-year drive life means each retained petabyte is bought roughly three to four times during 2025–2045.Source: §3.2.2 · p. 7

3.3.3 Drives, Library, and Power.

  • The cost accounting combines storage purchases, HDD replacements, and electricity, while tape’s durable cartridges avoid the HDD replacement series.Tape media is amortized across its 30-year cartridge lifetime, but drive and library infrastructure account for more than half of overall tape cost.Source: §3.2.2 · p. 7, §3.2.3 · p. 7, §3.3 · p. 8

3.2.3 Electricity Cost.

  • HDD electricity is modeled as a declining annual cost for keeping cumulative stored capacity powered; tape electricity is set at 10% of HDD’s unit cost.The tape assumption reflects that idle cartridges draw no power, while mounted drives and the library controller do.Source: §3.1 · p. 6, §3.2.3 · p. 7

3.3 Materialized-Artifact Storage on Tape

  • Tape media lasts beyond the study horizon and is bought once, while tape storage also requires separately modeled drives, library slots, and power.The model anchors 2025 tape at about $10,000/PB and assumes a rounded 10% annual price decline; the library grows with corpus size.Source: §3.3 · p. 8, §3.3.1 · p. 8, §3.3.3 · p. 8

3.3.2 Media Cost.

  • Tape media alone costs about $89,000 for linear growth and $504,000 for exponential growth by 2045.Drives and the robotic library add substantial costs; the library reaches about $750,000 under exponential growth and becomes the largest single tape cost at scale.Source: §3.3.2 · p. 8, §3.3.3 · p. 8, §3.4 · p. 9

3.4 Cost Over Next 20 years

  • By 2045, total HDD cost is approximately $365,000 for linear ingest and $1,597,000 for 20% annual ingest growth.Under exponential growth, the total is about 4.4× the linear baseline because ingest front-loads purchases at higher prices while the powered corpus accumulates.Source: §3.4 · p. 9
  • Tape remains cheaper than HDD under both linear and exponential corpus growth after drives, library, and power are included.At the 2045 horizon, tape costs around $261,000 versus HDD’s $365,000 for linear growth, and around $1,422,000 versus $1,597,000 for exponential growth.Source: §3.4 · p. 9, Table 2 · p. 9

4 Cost Model 2: Prompt-Based Regeneration

The model compares prompt-based regeneration with HDD storage over time, then adds caching to test whether regenerating only misses changes the break-even. In the modeled image workload, prompt-based regeneration remains more expensive than storage, even with a hot cache.

4 Cost Model 2: Prompt-Based Regeneration

  • The model compares cumulative serving costs with an HDD-storage baseline and reduces the regeneration decision to a break-even generation cost.It considers prompt-based regeneration first, then a hot cache backed by regeneration; both policies include capital and electricity costs.Source: §4 · p. 9
  • The image example models a 1024 × 1024 FLUX.1-dev generation at 2,100 TFLOP and GPU capacity prices declining 30% annually from the 2025 anchor.The request model assumes ten requests per artifact per year and independent selections from a fixed popularity distribution.Source: §4 · pp. 9–10

4.1 Prompt-Based Regeneration

  • Without storage, the prompt-based policy reruns the full prompt-to-image pipeline for every request, avoiding HDD capacity and electricity costs.Each request nevertheless incurs the full generation cost.Source: §4.1 · p. 10
  • At 2,100 TFLOP per FLUX.1-dev image, regeneration costs about $38 M (linear) and $68 M (exponential), versus storage costs of $365,000 and $1.60 M.The model’s 20-year break-even generation costs are about 20 TFLOP for linear growth and 50 TFLOP for exponential growth.Source: §4.1 · p. 11
  • Over 20 years, HDD storage costs about 100× less than prompt-based regeneration under linear growth and about 40× less under exponential growth.A stored artifact is written once, whereas regeneration pays its full cost on every request.Source: §4.1 · p. 11

4.2 Caching Hot Objects

  • A hot-object cache serves popular requests directly and confines full prompt-based regeneration to cache misses.The model relates hit ratio to the cached corpus fraction using a Zipfian popularity workload.Source: §4.2 · p. 11
  • As the corpus grows, the cache expands through purchases of new capacity and replacement capacity for aging cache drives.Annual purchases include the increase in cache volume since the previous year and replacement of drives installed one HDD lifetime earlier.Source: §4.2 · p. 11

4.3 Cost Over Next 20 years

  • At a 90% hit ratio, the cache costs about $38,000 (linear) or $137,000 (exponential), but caching still does not beat storage and FLUX.1-dev misses remain above break-even.The modeled break-even rises to about 182 TFLOP for linear growth and 454 TFLOP for exponential growth, still below FLUX.1-dev’s 2,100 TFLOP generation cost.Source: §4.3 · p. 12

5 Cost Model 3: LazIR, an Intermediate-Representation Store

LazIR stores compact intermediate representations and regenerates cache misses with a single decode rather than repeating denoising. In the modeled image cases, this makes the approach cheaper than full-object HDD storage.

5 Cost Model 3: LazIR, an Intermediate-Representation Store

  • FLUX.1-dev’s IR is 1/6 the decoded image size and decodes in about 10.3 TFLOP, 1/200 of full generation; SD 1.5/2.x and SDXL use 1/24, while SANA uses 1/48.IR-based regeneration skips iterative denoising and uses a single VAE decoder pass.Source: §5 · p. 13, Figure 4 · p. 13
  • LazIR stores an IR for every ingested object, keeps the hot working set cached in full fidelity, and regenerates only cache misses from the stored latent.The modeled miss computation is reduced by the IR decode fraction, while the IR store uses the IR size fraction of full-object storage.Source: §5 · p. 13
  • The model counts HDD and GPU electricity alongside the IR store and hot cache, with misses charged for reduced-cost decoding.The latent-store HDD bill is the IR size fraction of the materialized-storage baseline, including replacements.Source: §5 · pp. 13–14
  • At 2,100 TFLOPs, LazIR costs about $118,000 (linear) and $437,000 (exponential), versus $365,000 and $1.60 M for storage—a 3.1×–3.7× reduction.The modeled break-even rises to about 29,700 TFLOPs for linear growth and 74,200 TFLOPs for exponential growth.Source: §5 · p. 14

6 Evaluation: Storage versus Regeneration

The evaluation compares materialized-artifact storage with prompt-based and IR-based regeneration across image and music workloads. LazIR consistently costs less than storage in the evaluated settings, while its break-even advantage depends on corpus growth, IR size, and decode cost.

6 Evaluation: Storage versus Regeneration

  • The materialized-artifact storage baseline is identical across modalities because corpus volume is fixed, while image and music break-even points differ with object size, generation cost, IR size, and decode cost.The representative models are FLUX.1-dev for images and MusicGen-Large for music.Source: §6 · p. 14, §6.1 · p. 16
  • MusicGen-Large’s EnCodec token stream is about 1/212 the decoded clip size, and IR decoding costs about 0.045 TFLOP versus 21 TFLOP for prompt regeneration.A single non-autoregressive EnCodec decode regenerates the waveform from the stored token stream.Source: Figure 5 · p. 15, §6 · p. 15

6.1 Cost composition over time

  • At 20% annual ingest growth, HDD storage reaches about 75× the first-year baseline in capital and 4× in electricity by 2045, while tape reaches about 70× total cost.Tape’s roughly 45× drives-and-library operational slice includes a library cost near 37×; its total remains below HDD but above LazIR.Source: §6.1 · pp. 15–16, Figure 6 · p. 15

6.2 Regeneration cost across policies

  • Across both modalities and growth regimes, LazIR is the only policy below both storage baselines, with 8.6× and 10.5× reductions under linear and exponential growth, respectively.Prompt-based regeneration exceeds both baselines for images; cached prompt regeneration still exceeds storage for images, while music falls below HDD but above tape.Source: Figure 7 · p. 16, §6.2 · p. 16

6.1 Cost composition over time

  • LazIR costs settle near 27% of the HDD baseline for images and 10% for music by replacing most artifact storage with compact IRs and inexpensive miss decoding.Its cost comprises a cache, compact IR storage, and GPU regeneration of cache misses.Source: Figure 6 · p. 15, §6.1 · p. 16

6.3 Sensitivity of the break-even

  • As annual ingest growth rises from 0% to 40%, the break-even generation-cost threshold rises by roughly 130% because HDD purchases are front-loaded at higher prices.The threshold collapses as the IR size fraction approaches 0.9; a tenfold cheaper decode raises the threshold tenfold, and both modalities’ operating points remain well below it.Source: §6.3 · pp. 16–17

7 Analysis on a Real-World Image Generation Workload

A 2.87-year production image trace shows that prompt-based regeneration remains much more expensive than storage, while LazIR reduces total cost and supports interactive miss latency. Its serving performance also entails a peak-GPU tradeoff under bursty demand.

7.1 Trace Analysis

  • The 2.87-year trace contains 2.07 billion requests for 92.3 million images, and an LRU cache sized to 10% of the working set serves 87.2% of requests.The hottest 1% and 10% of images account for 38% and 71% of requests, respectively; the miss ratio rises as the corpus grows, then starts to stabilize after about 1.5 years.Source: §7.1 · p. 18

7.2 Cost Analysis

  • The cost comparison uses trace-driven simulation of prompt regeneration, cached prompt regeneration, and LazIR against HDD and tape storage under 2025 prices with FLUX.1-dev.The workload is a production trace from a large image-generation and sharing platform.Source: §7 · p. 17, §7.2 · p. 18
  • On the trace replay, HDD storage costs about $5,800 and tape about $23,800, while regenerating all requests costs about $2.11 million, roughly 364× HDD storage.At this corpus size of 0.28 PB, tape’s fixed drive and library costs are not yet amortized; cached prompt regeneration still costs about $270,000, or 47× HDD storage.Source: §7.2 · pp. 18–19
  • On the 2.07-billion-request trace, LazIR costs around $2,900, about 50% less than HDD storage and 100× less than prompt-based regeneration.The 10% working-set cache has a 12.8% miss ratio; the IR reduces per-regeneration compute cost by 200×, and its footprint is 1/6 of materialized images.Source: §7.2 · pp. 18–19

7.3 Serving Performance Analysis

  • With a queue-size limit of 2, LazIR achieves 66 ms median and 208 ms P99 miss-serving latency, including decoding and queueing.Most requests are cache hits, so end-to-end latency is substantially lower than miss latency.Source: §7.3 · p. 19
  • At peak, the elastic GPU pool requires 20–34 GPUs depending on queue size; smaller queue limits reduce latency but increase peak GPU demand.The trace’s substantial burstiness drives spiky GPU allocation over time.Source: §7.3 · p. 19

8 Related Work

Related work studies cheaper media for preserving artifacts, compact learned representations, and recomputation of derived data. LazIR frames a generator latent as a cold-store entry whose size and decode cost determine its storage-versus-regeneration break-even.

8 Related Work

  • Prior cold-storage work optimizes the medium for preserved artifacts, whereas this paper asks whether retaining the artifact is cheaper than regenerating it and uses tape as a storage baseline.The cited approaches include purpose-built photo stores, tape, and DNA.Source: §8 · p. 19
  • LazIR treats a generator latent as a cold-store entry, with its size fraction and decode-cost fraction setting the economic break-even against materialized storage.The paper distinguishes this use from learned image compression [7] and latent spaces of modern generators [31].Source: §8 · p. 20
  • Unlike recomputation settings with no inherent winner, generative pipelines front-load nearly all compute before final decoding, letting latents capture over 99% of generation cost at a fraction of artifact size.The passage relates recomputation to materialized-view selection, derived datasets, lineage-based recovery, and activation rematerialization [6].Source: §8 · p. 20

9 Assumptions

The cost model assumes current technology and cost trends persist over four decades, while acknowledging shocks and omitted factors that can shift break-even estimates.

9 Assumptions

  • The model assumes current cost and technology trends persist for four decades, though supply shocks and rising demand could alter storage-price trajectories.Examples include the 2011 Thailand floods’ effect on HDD supply and AI-driven demand pressure on HBM and high-capacity SSD prices.Source: §9 · p. 20
  • It omits inflation and compression of artifacts or IRs, and assumes skewed access that lets caching reduce regeneration cost.Other access distributions would change caching effectiveness and shift the break-even point.Source: §9 · p. 20
  • The analysis assumes regeneration reproduces the exact original artifact, although batch size, GPU model, software stack, or numerical nondeterminism may cause small bit-level differences.The passage cites for such differences.Source: §9 · p. 20

10 Conclusion

The paper argues that storing compact intermediate representations and decoding on demand moves regeneration’s break-even from decades away to today. Its conclusion recommends IR as the unit of long-term retention, while its notation distinguishes exponential-growth models from a linear baseline.

10 Conclusion

  • Prompt-based regeneration is two orders of magnitude more expensive than materialized-artifact storage and does not break even until the mid-2040s.LazIR stores compact intermediate representations and decodes on a miss, moving break-even to today and cutting costs by 2.6× to 10.5× across image and music workloads.Source: §10 · p. 21
  • Production-trace evaluations confirm the reported cost findings.The conclusion describes evaluations on a production trace but gives no trace size or further metrics here.Source: §10 · p. 21

A Notation

  • Table 5’s superscript (g) marks the exponential-growth variant; the linear 1 PB/yr baseline is its g=0 case.Table 4 lists empirical parameter values, while Table 5 defines model notation.Source: Table 5 · p. 26

B Access Pattern and Cache Fraction

The cache fraction is the portion of the corpus held at full fidelity to serve a target fraction of requests. Under Zipfian skew, a small hot set can absorb most requests, and the required fraction falls as the corpus grows.

B Access Pattern and Cache Fraction

  • F_cached is the corpus fraction held at full fidelity to serve a target request fraction h, and its value depends entirely on request-popularity skew.The paper derives it in closed form for a representative Zipfian workload.Source: B Access Pattern and Cache Fraction · p. 27
  • At a 90% hit ratio, F_cached ≈ 0.944 N^-0.10; about 15% of 10^8 objects or 9.4% of 10^10 objects serve 90% of requests.The cache-fraction expression is derived under a Zipfian workload with skew s=1.Source: B Access Pattern and Cache Fraction · p. 27
  • The required cache fraction decreases as the corpus grows.The paper uses the resulting values directly in its Section 4.2 caching policies.Source: B Access Pattern and Cache Fraction · p. 27

C Break-Even Walkthrough

The walkthrough compares regeneration policies against HDD storage at a 2045 horizon under 20% annual ingest growth and a 90% Zipfian hit ratio. IR-backed regeneration on cache misses beats storage for both evaluated modalities, while prompt-based regeneration is more constrained by its compute cost.

C Break-Even Walkthrough

  • The evaluation uses a 2045 horizon, 20% annual ingest growth, and a 90% Zipfian hit ratio, counting HDD replacement capital and corpus-power electricity.A policy is preferred to HDD storage when its generation cost falls below its break-even threshold.Source: C Break-Even Walkthrough · p. 27
  • Under linear growth of 1 PB of new objects per year from 2025–2045, total HDD storage cost is about $365,000, including $344,000 capital and $21,000 electricity.The capital figure includes drive replacement over the horizon.Source: C Break-Even Walkthrough · p. 27
  • For exponential growth, the 2045 storage baseline is about $1,597,000, including HDD replacement capital and electricity.The paper uses this exponential-growth total as the storage side of its break-even comparisons.Source: C Break-Even Walkthrough · p. 27
  • Prompt-based regeneration’s break-even thresholds are 49.7 TFLOP for images and 84.4 TFLOP for music; only MusicGen-Large at 21 TFLOP is below its threshold.FLUX.1-dev at 2,100 TFLOP favors materialized-artifact storage.Source: C Break-Even Walkthrough · p. 27
  • IR-backed regeneration on cache misses beats materialized-artifact storage for both evaluated modalities: FLUX.1-dev uses 2,100 TFLOP against a 74,200 TFLOP threshold, and MusicGen-Large uses 21 TFLOP against 359,000 TFLOP.The LazIR design stores compact intermediate representations and decodes misses at reduced cost.Source: C Break-Even Walkthrough · p. 28

D Carbon Emissions

The carbon analysis combines operational emissions from electricity with embodied emissions from manufactured hardware. LazIR reduces emissions by shrinking HDD storage in proportion to latent size, while miss decoding adds GPU carbon; the appendix also identifies sources for hardware-trend inputs.

D Carbon Emissions

  • The carbon estimates use a grid intensity of 0.30 kg CO2/kWh and embodied HDD intensity of 5 kg CO2/TB.The paper describes the grid-intensity assumption as conservative relative to the 2023 US average.Source: D Carbon Emissions · p. 28
  • Operational carbon coefficients are 3.75 kg CO2 per dollar of corresponding electricity cost; embodied GPU carbon adds roughly 16% to operational GPU carbon per TFLOP.The model includes operational and embodied emissions for HDDs and GPUs.Source: D Carbon Emissions · p. 28
  • The 5,047 t CO2 full-HDD baseline comprises 3,269 t from powering the cumulative corpus and 1,778 t embodied in manufactured HDDs, including replacements.These are lifecycle emissions over the 20-year horizon.Source: D Carbon Emissions · p. 29, Table 6 · p. 29
  • LazIR emits 1,262 t CO2 for image generation, a 75.0% reduction from full HDD storage, and 25 t CO2 for music, a 99.5% reduction.The full-HDD baseline emits 5,047 t CO2 over 20 years; Table 6 reports the cumulative emissions setting.Source: D Carbon Emissions · p. 29, Table 6 · p. 29
  • Both operational and embodied HDD emissions shrink in proportion to latent storage fraction k_size, while miss decoding adds only a small amount of GPU carbon.Image generation remains the largest two-tier case because its latent representation is k_size=1/6.Source: D Carbon Emissions · p. 29
  • Embodied HDD emissions account for 35% of the full-HDD baseline, and their two-tier reduction is greatest for modalities with the most compact intermediate representations.Embodied HDD emissions count manufactured drives over the horizon, including replacements.Source: D Carbon Emissions · p. 29

E Sources for the hardware trend

Loading 2609.30448v1…