Pith. sign in

REVIEW 5 major objections 5 minor 25 references

Quantifying Energy and Cost Benefits of Hybrid Edge Cloud: Analysis of Traditional and Agentic Workloads

T0 review · 5 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Hybrid Edge Cloud can cut energy by up to 75% and cost by more than 80% by processing most data locally; the benefit scales linearly with the edge split.

desk verdict The paper's arithmetic is transparent but its headline savings numbers require swapping in different constants halfway through, so the quantitative claims don't follow from the stated model. read the letter →

arxiv 2501.14823 v2 pith:WHZUVDOD submitted 2025-01-21 cs.DC cs.AI

Siavash Alamouti This is my paper · ORCID
classification cs.DCcs.AI
keywords hybridedgecloudcomputingenergyefficiencycostreductionParetoworkloaddistributionagenticworkloadsMonteCarlosimulationdevice-first
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a hybrid edge-cloud architecture, processing most data locally on devices and sending only demanding tasks to cloud gateways, can cut both energy and cost by large margins relative to today's centralized-cloud model. Its headline claim is energy savings up to 75% and cost reductions exceeding 80%, even for data-hungry agentic workloads. The argument reduces the benefit to a pair of linear savings formulas driven by the fraction of data processed at the edge and by per-gigabyte energy and cost constants; the paper's concrete example at an 80% edge split shows about 65% energy and 75% cost savings for traditional 2.4 GB/day devices. The conclusion is that device-first processing is practical and essential for the next generation of AI agents, robotics, and autonomous systems.

What carries the argument

The load-bearing object is the pair of linear savings identities $S_{Energy}=((E_t+E_c-E_l)/(E_t+E_c)) P_{edge}$ and $S_{Cost}=((C_b+C_h-C_s)/(C_b+C_h)) P_{edge}$, together with the assumed per-GB constants. These identities convert a workload-allocation policy, the edge split $P_{edge}$, directly into percentage savings, so the benefit computation is a single multiplication rather than a simulation outcome. The Pareto workload distribution with $\alpha=2$ and $x_m=1$ supplies the justification that 70–90% of workloads are lightweight enough to run locally, and Monte Carlo draws over that distribution (with $\alpha=2$ and $\alpha=3$) are used to confirm that the closed-form savings hold across edge splits from 50% to 90%.

What would settle it

Measure the real per-gigabyte energy for an end-to-end workload: transmitting data to a public cloud region, running inference on a typical cloud instance, and running the same inference on a current smartphone or edge NPU. If measured cloud-plus-transmission energy is not roughly three times local-processing energy, the claimed 62–75% energy savings at an 80% edge split will not reproduce.

Watch

Extended reading notes

Core claim

The central claim is that Hybrid Edge Cloud has a closed-form benefit law. If $P_{edge}$ is the fraction of workload data processed locally, energy savings are $((E_t+E_c-E_l)/(E_t+E_c)) P_{edge}$ and cost savings are $((C_b+C_h-C_s)/(C_b+C_h)) P_{edge}$, where $E_t$, $E_c$, and $E_l$ are the per-GB energy for transmission, cloud processing, and local processing, and $C_b$, $C_h$, and $C_s$ are the per-GB bandwidth, hosting, and software costs. Plugging in the paper's values, $E_t=0.7$, $E_c=1.5$, $E_l=0.5$ kWh/GB and $C_b=0.10$, $C_h=0.20$, $C_s=0.02$ dollars/GB, an 80% edge split produces about 62–65% energy savings and about 75% cost savings for traditional workloads. For agentic workloads generating 20 GB/day, the paper reports per-device savings around 10,000 kWh and $1,500 per year, with a separate agentic accounting reaching up to 75% energy savings.

Load-bearing premise

The entire savings calculation rests on the assumed per-gigabyte energy figures—0.7 kWh for transmission, 1.5 kWh for cloud processing, and 0.5 kWh for local processing—and on those figures holding across workload types, even though the paper's agentic section adopts a different transmission value (5 kWh/GB).

Editorial extensions

If this is right

  • An 80% edge split delivers roughly 65% energy savings and 75% cost savings for traditional workloads, matching the abstract's headline benefit range.
  • Agentic workloads at 20 GB/day yield similar percentage savings but roughly an order of magnitude larger absolute savings: about 10,000 kWh and $1,500 per device per year.
  • Even a 30% edge split gives 25–30% savings, so partial or incremental adoption is economically worthwhile.
  • At billions of devices, aggregate savings would reach tens of trillions of kWh and trillions of dollars annually, according to the paper's extrapolation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Extension: Because the savings formulas are linear in $P_{edge}$ and independent of total data volume, the same percentage benefit applies across device classes, with only the absolute savings scaling.
  • Extension: The headline numbers can be tested directly by measuring the three per-GB energy constants on a real fleet and recomputing the formulas; any difference between measured and assumed constants changes the claimed savings proportionally.
  • Extension: The multi-trillion-dollar aggregate estimate assumes per-device savings multiply across tens of billions of devices with no added fleet-level coordination, cooling, or network costs, so a full-system accounting would need to add those terms.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper develops a closed-form analytical model to compare the energy and cost of centralized cloud computing against a hybrid edge-cloud (HEC) architecture, where a fraction Pedge of each device's data volume is processed locally and the remainder is sent to the cloud. The model includes per-gigabyte energy costs for transmission, cloud processing, and local processing, as well as bandwidth, hosting, and software costs. The paper applies this model to traditional IoT workloads (2.4 GB/device/day) and to hypothetical agentic workloads (20 GB/device/day), and it reports Monte Carlo simulations over Pareto-distributed workload sizes. The central claims are that HEC achieves "energy savings of up to 75% and cost reductions exceeding 80%, even in resource-intensive agentic scenarios," with per-device annual savings on the order of 10,000 kWh and $1,500.

Significance. The question of where to place computation is practically important, and the closed-form expressions in §3.2 (S_Energy and S_Cost) give a transparent way to compare architectures if the underlying parameters are known. The paper is also commendable for explicitly modeling agentic workloads, which are likely to grow. However, the manuscript's own numbers do not follow from its model: the 75% agentic energy saving is obtained only by changing the transmission-energy constant in §6.1, and the traditional-workload example in §4.1 is inconsistent with the formula. No sensitivity analysis or empirical data is provided for the per-GB energy and cost constants, which are the very parameters that determine the headline results. The Monte Carlo simulation varies workload sizes but not these constants, so it cannot be considered an independent validation. As presented, the paper demonstrates a simple algebraic relationship and then attaches unsupported numeric values to it.

major comments (5)
  1. [§6.1 vs §3.2] The abstract's headline of 'energy savings of up to 75%' in agentic scenarios is not a consequence of the model presented. With the constants stated in §2.2 and §4.1 (Et = 0.7, Ec = 1.5, El = 0.5 kWh/GB) and the 80% edge split used throughout, Eq. (3.2) gives S_Energy = ((0.7+1.5-0.5)/(0.7+1.5)) × 0.8 = 61.8%. The 75% figure appears only in §6.1, where agentic workloads are assigned a transmission energy of 5 kWh/GB, a value that is not part of the model, the numerical tables, or the Monte Carlo simulation. The claim is therefore a parameter substitution, not a derived result.
  2. [§4.1] The numerical example for traditional workloads is internally inconsistent with the model. Eq. (3.2) with 876 GB/year, Pedge=0.8, and the stated constants gives EHEC = 876 × (0.8×0.5 + 0.2×2.2) = 735.8 kWh/year and S_Energy = 61.8%, yet the text reports 674 kWh/year and 'approximately 65%.' No derivation is given for these alternative numbers. Similarly, §6.1 asserts that traditional workloads show energy savings 'as much as 80%,' which cannot occur under Eq. (3.2) even at Pedge=1, where the saving is bounded by (1.7/2.2) = 77.3%. These contradictions undermine confidence in the reported results.
  3. [§3.2 and Abstract] The claim of 'cost reductions exceeding 80%' is not established by the model. With Cb=0.10, Ch=0.20, Cs=0.02 $/GB, the cost saving formula gives S_Cost = (0.10+0.20-0.02)/(0.10+0.20) × Pedge = 0.9333 × Pedge. Achieving S_Cost > 0.80 therefore requires Pedge > 85.7%. The paper does not justify such a high edge split for agentic workloads; the Pareto argument in §3.1 addresses workload counts (70-90% lightweight), not data-volume fractions, and §6.2 provides only qualitative statements. The abstract's numeric cost claim is thus not supported by the stated model and assumptions.
  4. [§5] The Monte Carlo simulation does not validate the model's key parameters. It varies workload sizes according to Pareto distributions and varies the edge split, but it holds the per-GB energy and cost constants fixed at the §4.1 values. It therefore cannot test the 5 kWh/GB transmission energy that underlies the §6.1 agentic claim, nor can it reveal how sensitive the reported savings are to the base constants. The observation that the two Pareto shape parameters give 'almost identical' results is a consequence of normalizing to a fixed annual workload; it does not establish robustness. As reported, the simulation is a recapitulation of the algebraic formula, not an independent check.
  5. [§2.2] The three per-GB energy constants are load-bearing but unsupported. The text states that the values are 'based on studies published in' [11]-[15], but those references are general surveys (Masanet et al., Cisco Global Cloud Index, Akamai) and do not directly supply 0.7 kWh/GB for transmission, 1.5 kWh/GB for cloud computation, or 0.5 kWh/GB for local computation. The paper provides no measurement, no dataset, and no sensitivity analysis for these constants. Since the savings formula is linear in these parameters and the headline numbers scale directly with them, the manuscript's central quantitative conclusions rest on unverified inputs.
minor comments (5)
  1. [§3.2] The variables Pedge and Pcloud are described as 'probabilities' but are used as fractions of data volume; state explicitly that they are deterministic split fractions satisfying Pedge + Pcloud = 1.
  2. [References] Reference [15] is missing its URL ('Retrieved from , 2023'); several other references (e.g., [13]) mix nonstandard formats and would benefit from a consistent style.
  3. [Figures] Figures 1 and 2 are referenced but not included in the submitted text; please ensure all figures are embedded and legible.
  4. [§6.1] Section 6.1 uses the phrase 'exponential scaling of data transmission costs,' but the model in §3.2 is linear in data volume; this should be reworded to avoid implying nonlinearity.
  5. [Discussion] The paper would benefit from an explicit limitations paragraph acknowledging that the quantitative results depend on assumed constants and that the Monte Carlo simulation does not vary those constants.

Circularity Check

2 steps flagged · score 6.0 of 10

Monte Carlo 'validation' recomputes the same inputs, and the 75% agentic saving is produced by swapping in a new transmission-energy constant rather than by the stated model.

  1. other [Section 5, Simulation Results]
    "To validate the numerical results and model the real world more accurately, we can use Monte Carlo simulations. ... The results for the two Pareto shape parameters were almost identical. While Pareto distributions differ in workload skewness, the normalization to a total annual workload of 7300 GB minimizes the effect of the skewness on overall energy and cost savings."

    The simulation is presented as validation, but it uses the same per-GB energy and cost constants from Sections 2.2 and 4.1, the same savings formulas from Section 3.2, and the same edge-split percentages as the numerical analysis. The only varied input is the workload-size distribution, and the paper itself concedes that normalization to 7300 GB 'minimizes the effect' of that variation. Hence the simulation's agreement with the numerical results is guaranteed by construction; it is a recomputation of the same inputs, not an independent check of the savings claim.

  2. self definitional [Section 6.1, Agentic Workloads, versus Sections 3.2 and 4.1]
    "These workloads consume far more energy when processed in the cloud due to both high transmission energy (5 kWh/GB) and computational energy (1.5 kWh/GB) for cloud processing. By processing 80% of these workloads locally on end devices, HEC reduces overall energy consumption by approximately 75%."

    The paper's own formula, Eq. (3.2), is S_Energy = ((Et + Ec - El)/(Et + Ec)) * Pedge. With the model's stated Section 2.2/4.1 constants (Et = 0.7, Ec = 1.5, El = 0.5 kWh/GB) and Pedge = 0.8, this gives about 61.8%, not 75%. The 75% figure appears only after Section 6.1 introduces Et = 5 kWh/GB for agentic workloads; then the formula gives about 73.8%, which is rounded to 75%. The headline agentic claim is therefore not derived from the model's parameters but is constructed by substituting a new input at the discussion stage, making the claimed result equivalent to that substituted input by construction.

full rationale

The underlying arithmetic model is not inherently circular: Eq. (3.2) openly defines energy savings as a function of per-GB energy constants and the assumed edge split, and those constants are cited to external literature. However, the paper contains two circular or construction-equivalent moves. First, the Monte Carlo simulation purportedly validates the numerical results, yet it uses the same constants, the same formulas, and the same edge splits; varying only the Pareto workload sizes cannot test the energy or cost parameters, and the paper admits the normalization removes the effect of that variation. Second, the abstract's headline claim of 75% energy savings in agentic scenarios is not what the formal model with the stated constants gives at the cited 80% edge split: it yields 61.8%. The 75% figure is obtained only after Section 6.1 silently swaps the transmission energy to 5 kWh/GB, making the headline an arithmetic consequence of a newly inserted parameter rather than a derived result. Self-citations to the author's HEC concept [1] and to mim-OE software licensing are present, but I do not treat them as load-bearing circularities because the quantitative equations do not logically depend on those references; the main circularity is the validation loop and the parameter-substituted headline.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

Every headline number follows from assumed constants and the edge split. The paper provides no measurements and no sensitivity analysis over the parameters that determine the result. The only simulation varies workload size distribution, which the model itself shows is nearly irrelevant after volume normalization.

free parameters (8)
  • Transmission energy Et = 0.7 kWh/GB
    Used in the energy model; cited to references [11]-[12] with no uncertainty or measurement.
  • Cloud processing energy Ec = 1.5 kWh/GB
    Used in the energy model; cited to references [13]-[15], but contradicted by the 5 kWh/GB transmission value in Section 6.1.
  • Local processing energy El = 0.5 kWh/GB
    Central to the savings claim; assumes local processing is efficient and uniform across devices, with no measurement in the paper.
  • Bandwidth cost Cb = $0.10/GB
    Chosen from the cited range $0.01-$0.12/GB; the midpoint is not justified.
  • Hosting cost Ch = $0.20/GB
    Assumed average cloud cost; cited to provider pricing pages without exact evidence.
  • Software cost Cs = $0.02/GB
    Set as 10% of hosting cost for mim-OE, introducing the author's commercial platform into the cost comparison.
  • Edge split Pedge = 0.8 (baseline)
    Assumed Pareto-based split; all savings scale linearly with this value.
  • Data volume DT = 876 GB/yr traditional; 7300 GB/yr agentic
    Derived from 2.4 and 20 GB/day assumptions; the agentic volume has no cited source.
assumptions (5)
  • domain assumption Workload sizes follow a Pareto distribution with alpha=2 and xm=1.
    Invoked in Section 3.1 and supported only by cited workload studies, not by data in this paper.
  • domain assumption 70%-90% of workloads are lightweight and can run on existing end devices.
    Assumed in Sections 2.4 and 3.1 using industry reports [2], [20], [22]-[23]; not demonstrated here.
  • domain assumption Per-GB energy and cost are constant and independent of workload size, device type, and concurrency.
    The linear models in Section 3.2 require this; real efficiency varies with hardware and utilization.
  • domain assumption Tasks are assigned to the edge or cloud with uniform probability based on the edge split.
    Stated in Sections 2.4 and 5; ignores workload-specific routing and scheduling.
  • domain assumption Devices have spare capacity to absorb 80% of workloads without extra hardware.
    Implied by the 80% edge split and the discussion in Section 3.1; not tested for agentic workloads.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantifying Energy and Cost Benefits of Hybrid Edge Cloud: Analysis of Traditional and Agentic Workloads." pith.science (2026). https://pith.science/paper/WHZUVDOD

@misc{pith2026250114823,
  author       = {Pith},
  title        = {Pith review of: Quantifying Energy and Cost Benefits of Hybrid Edge Cloud: Analysis of Traditional and Agentic Workloads},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WHZUVDOD}},
  note         = {Machine review of arXiv:2501.14823}
}
read the original abstract

This paper examines the workload distribution challenges in centralized cloud systems and demonstrates how Hybrid Edge Cloud (HEC) [1] mitigates these inefficiencies. Workloads in cloud environments often follow a Pareto distribution, where a small percentage of tasks consume most resources, leading to bottlenecks and energy inefficiencies. By analyzing both traditional workloads reflective of typical IoT and smart device usage and agentic workloads, such as those generated by AI agents, robotics, and autonomous systems, this study quantifies the energy and cost savings enabled by HEC. Our findings reveal that HEC achieves energy savings of up to 75% and cost reductions exceeding 80%, even in resource-intensive agentic scenarios. These results highlight the critical role of HEC in enabling scalable, cost-effective, and sustainable computing for the next generation of intelligent systems.

Figures

Figures reproduced from arXiv: 2501.14823 by the authors.

Figure 1
Figure 1. Energy and Cost Comparison of HEC with cloud-only solutions [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Energy and Cost Savings of HEC as a function of edge split. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 22 canonical work pages

  1. [11]

    Masanet, A

    E. Masanet, A. Shehabi, N. Lei, S. Smith, and J. Koomey. Recalibrating global data center energy-use estimates. Science, 367(6481):984–986, 2020

    work page 2020

  2. [15]

    Aws pricing overview

    Amazon Web Services. Aws pricing overview. Retrieved from , 2023. 12

    work page 2023

  3. [1]

    Alamouti, Fay Arjomandi, and Michel Burger

    Siavash M. Alamouti, Fay Arjomandi, and Michel Burger. Hybrid edge cloud: A pragmatic approach for decentralized cloud computing. IEEE Communications Magazine, 60(9):16–29, 2022

    work page 2022

  4. [2]

    Tensorflow lite

    Google AI. Tensorflow lite. TensorFlow White Paper. [Online]. Available: https://www. tensorflow.org/lite

    work page

  5. [3]

    Jackson and C

    K. Jackson and C. Bunch. Understanding Cloud Workloads: An Overview of A WS Instance Types and Usage Patterns. Unknown Publisher, 2019

    work page 2019

  6. [4]

    Barroso, J

    L. Barroso, J. Clidaras, and U. H”olzle. The Datacenter as a Computer: An Introduction to the Design of Warehouse-Scale Machines. Morgan & Claypool, 2013

    work page 2013

  7. [5]

    Y. Xu, J. Bi, W. Yuan, and W. Yuan. Cloud workload characteristics and resource management approaches: A comprehensive study. IEEE Transactions on Cloud Computing, 5(4):772–785, 2017

    work page 2017

  8. [6]

    Reiss, J

    C. Reiss, J. Wilkes, and J. Hellerstein. Google cluster-usage traces: Format + schema. Un- known Journal, 2011

    work page 2011

Show all 25 references
  1. [7]

    Dean and S

    J. Dean and S. Ghemawat. Mapreduce: Simplified data processing on large clusters. Commu- nications of the ACM, 51(1):107–113, 2008

    2008

  2. [8]

    R. N. Calheiros, R. Ranjan, C. A. F. De Rose, and R. Buyya. Cloudsim: A toolkit for model- ing and simulation of cloud computing environments and evaluation of resource provisioning algorithms. Software: Practice and Experience, 41(1):23–50, 2011

    2011

  3. [9]

    Baliga, R

    J. Baliga, R. Ayre, K. Hinton, and R. S. Tucker. Energy consumption in wired and wireless access networks. IEEE Communications Magazine, 49(6):70–77, 2011

    2011

  4. [10]

    R. S. Tucker. Energy consumption of data networks in the internet. IEEE Journal on Selected Areas in Communications, 34(5):1507–1519, 2018

    2018

  5. [12]

    Venkatesh, V

    V. Venkatesh, V. Sivaraman, and R. Vyas. Energy efficiency in iot and edge computing: A comparative study. IEEE Access, 9:56702–56715, 2021

    2021

  6. [13]

    State of the internet report

    Akamai. State of the internet report. Retrieved from https://www.akamai.com/ state-of-the-internet , 2020

    2020

  7. [14]

    Cisco global cloud index: Forecast and methodology, 2020–2025

    Cisco. Cisco global cloud index: Forecast and methodology, 2020–2025. Retrieved from https: //www.cisco.com/c/en/us/solutions/service-provider/global-cloud-index.html , 2021

    2020

  8. [16]

    Google cloud pricing

    Google Cloud. Google cloud pricing. Retrieved from https://cloud.google.com/pricing, 2023

    2023

  9. [17]

    Azure pricing overview

    Microsoft Azure. Azure pricing overview. Retrieved from https://azure.microsoft.com/ en-us/pricing, 2023

    2023

  10. [18]

    Hinton, O

    M. Hinton, O. Vinyals, and J. Dean. Distilling the knowledge in a neural network. arXiv preprint, arXiv:1503.02531, 2015. [Online]. Available: https://arxiv.org/abs/1503.02531

    2015 arXiv

  11. [19]

    The case for edge ai

    McKinsey & Company. The case for edge ai. McKinsey Tech- nology Insights, [Online]. Available: https://www.mckinsey.com/ industries/technology-media-and-telecommunications/our-insights/ why-edge-ai-is-the-future-of-intelligent-devices , 2022

    2022

  12. [20]

    Arm cortex-a processors for ai workloads

    ARM. Arm cortex-a processors for ai workloads. ARM White Paper, [Online]. Available: https://www.arm.com/solutions/artificial-intelligence, 2023

    2023

  13. [21]

    A. S. G. Andrae and T. Edler. On global electricity usage of communication technology: Trends to 2030. Challenges, 6(1):117–157, 2015

    2015

  14. [22]

    The rise of edge-native ai

    Intel and Dell Technologies. The rise of edge-native ai. Intel Edge AI Report, [Online]. Avail- able: https://www.intel.com/content/www/us/en/artificial-intelligence/overview. html, 2023

    2023

  15. [23]

    Ai workloads: Benchmarking edge vs

    Deloitte Insights. Ai workloads: Benchmarking edge vs. cloud. [Online]. Available: https: //www2.deloitte.com/us/en/pages/technology-media-and-telecommunications/ articles/ai-edge-computing.html, 2023

    2023

  16. [24]

    Z. Turc, M. Chang, Y. Lee, and K. Toutanova. Mobilebert: A compact task-agnostic bert for resource-limited devices. arXiv preprint, arXiv:2004.02984, 2020. [Online]. Available: https: //arxiv.org/abs/2004.02984

    2004 arXiv

  17. [25]

    Efficient processing of deep neural networks

    MIT-IBM Watson AI Lab. Efficient processing of deep neural networks. arXiv preprint, arXiv:1606.09274, 2016. [Online]. Available: https://arxiv.org/abs/1606.09274. 13

    2016 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.