JouleFlex: A Model-Based Decision and Control Framework for Maximizing Useful AI Work under Dynamic Power Limits

2.5.2 Abstract / תקציר

English (300/300 words)

In July 2026 Israel's Electricity Authority froze new data-center connection requests after the queue reached 27,000 megawatts. Under electricity limits, data centers switch GPUs (AI processors) off or cap their power, deciding in watts (the momentary draw), not joules (the energy consumed). A capped GPU computes slower and pays its fixed power floor longer, so energy per unit of useful work can rise even as the meter shows fewer watts: less AI per kilowatt-hour. A similar challenge exists in aviation, where it is solved: fuel flow rises far faster than speed, yet slow flight burns fuel longer, so airlines cruise at the speed minimizing fuel per kilometer.

The PI's recent study measured the time-power tradeoff for AI workloads (5,500 measurements, 20 models, four GPUs). Energy per task is U-shaped in the power cap. On the large GPUs its minimum, the Joule Point (JP), sits at 43 to 46 percent of rated power; holding a card there cut energy per task 29 to 31 percent while running 1.2 times slower. Even under a strict deadline, capping saved 28 percent of energy while keeping 95 percent of workload value. A fleet simulation on the measured curves (only scheduling simulated) served equal work with 18 to 45 percent less energy.

These insights reshape how AI data centers are planned, managed, and priced, raising decisions no current tool supports: how to design a flexible-connection-ready fleet whose cards run near their JP; how to measure and predict the JP for new cards and jobs; and how to price service to motivate customers toward efficiency, not per-hour billing rewarding the energy-worst point. Each is a modeling problem. This project delivers the decision-support toolkit answering all three, released with an open measurement dataset. Run near its JP, a fleet frees dispatchable power headroom for Israel’s flexible-connection scheme.

עברית (300/300 מילים)

2.5.3 Glossary

2.5.4 Objectives of the program

2.5.4.1 The objective

The objective is to construct JouleFlex: a toolkit of models, software tools and a simulator for the energy management of GPU-rich AI data centers, built on the JP concept, in three layers. Models: the power-response law validated on current hardware and whole servers; a job model calibrating the energy-optimal cap per workload; a transfer model for unmeasured GPUs; and a joint job-and-cap assignment method. Tools: a characterization kit qualifying a new GPU within the validated hardware domain in hours, and a fleet control engine holding a facility inside its power envelope while meeting deadlines. Simulation: a digital twin replaying measured curves, evaluating any policy or connection scenario before deployment. All layers are released openly with the calibrating corpus (ELF-2, the extended measurement dataset), serving two users: the operator who runs it, and the Ministry that plans, regulates and sizes connections with its numbers. The twin answers the fleet-design question directly: how many cards, of which type, run near the JP behind a given connection.

2.5.4.2 The research question

Can AI data centers operate as controllable electrical loads, each GPU near its energy-optimal cap and the facility inside a changing power limit, so that energy per unit of AI work falls and the power capacity required to serve it shrinks, while admitted work keeps its service promises, under Israeli grid constraints?

Six sub-questions structure the work plan:

1. Does the law generalize? The power law and JP were measured on 20 models and 4 GPUs; do they hold on H100-class and, as access allows, B200-class hardware, production serving software, and shared GPUs?
2. When is one cap per card enough? The measured JP was nearly constant per card in the loaded regime; the project maps where that constant breaks and fits a job model there.
3. Where is the true optimum at the wall? Board-power measurements miss host overhead, and the grid connection is sized at the wall: what is the JP of the whole node, and of the facility?
4. Can we skip the measurement? Can an unmeasured GPU's JP be predicted from published electrical characteristics plus at most one short test, characterizing a new card model in hours?
5. Can allocation and power be optimized together, in motion? Choose at once which job runs on which card and at what cap, maximizing served work under a time-varying power limit (price, carbon, grid instruction) with every deadline met; the measured result covers fixed caps and a fixed placement rule; the joint problem is open.
6. How much flexibility can a facility sell, and who has the incentive? How much power can be shed on instruction, how fast, for how long, at what service cost, the numbers behind conditional connections; and which pricing rules make offering flexibility profitable, when per-hour billing rewards the energy-worst point?

2.5.5 The proposed research

2.5.5.1 Scientific and technological background

The general problem. AI data centers are the world's fastest-growing electricity consumers, with demand projected to approach 945 TWh by 2030 [3].

The Israeli challenge is specific and pressing. Israel's grid is an isolated island system with no cross-border interconnections, so new demand must be met from domestic capacity; in July 2026 the Electricity Authority halted new data-center connection requests for 140 days, after the queue reached about 27,000 MW, roughly three times average national consumption [2]. The binding limit is grid capacity in megawatts: the national need is more AI work from every connected megawatt.

Today's operating decisions think in watts, not joules. Under electrical stress a data center either switches GPUs off, discarding service, or caps GPU power energy-blind: a capped GPU computes slower, paying its fixed power floor for more seconds, so capping too deep lowers the dashboard watts while raising the joules on the bill. Unconstrained, the default is full rated power, where the last speed carries a steep power surcharge. A further choice goes unweighed: for a fixed workload, whether to concentrate it on fewer cards at full power or spread it across more cards at partial power; our measurements show the second is sometimes more energy-efficient.

Our discovery. The applicant's study [1] mapped, over 20 AI models and four GPU types (~5,500 instrumented measurements, released as the ELF dataset), how energy per task depends on the power cap: board power follows a simple physical law in the computing rate θ,

P(θ) = P₀ + a·θ^β,   median R² = 0.99,

a fixed floor P₀ plus a superlinear term. Energy per task, E = P/θ, is therefore U-shaped in the cap; its minimum, the JP, sits at 43 to 46 percent of rated power on the large GPUs, cutting energy per task by 29 to 31 percent at an exactly priced cost: tasks run about 1.2 times slower, and holding throughput takes that factor more cards. The cap is one software command, settling in about 0.2 seconds, fast enough to follow grid signals; behind a fixed feed the measured A100 case yields about 50 percent more AI tasks per megawatt [1].

2.5.5.2 Review of existing knowledge

Four research lines meet here. Energy-proportional computing [4] identified the fixed, work-independent part of server power as the central inefficiency; our floor P₀ measures it for modern GPUs, and the JP is where paying it stops being worthwhile. Speed-scaling theory proved that energy is minimized at a sub-maximal speed [5,6,7]; the ELF dataset gives the measured form for GPU inference, where the law holds per operating point and breaks when pooled across batch sizes. GPU energy systems, notably Zeus [8] and the ML.ENERGY benchmarks [9,10], search for a good setting per job at runtime; the per-card-constant JP replaces that search with a one-time characterization. Production context: cluster studies [11] and public traces [12] supply realistic job arrivals; energy-aware GPU schedulers [13,14] leave each device's power draw at its default; methodology reviews [15] motivate our fixed protocol. Among the cited literature we found no system that keeps an AI fleet inside a time-varying, grid-facing envelope with auditable compliance, or that quantifies AI facilities as providers of firm and flexible capacity.

2.5.5.3 Knowledge gaps

• The law's reach is unmapped. The law and JP are established for single jobs at fixed settings on three inference GPUs; their persistence on H100-class and, as access allows, B200-class hardware, production serving software and shared cards is unknown, and placing the JP on a new GPU still requires a full sweep, since prediction from published characteristics is the study's stated next step [1].
• The job-dependence structure is unmapped. The per-card constant holds in the measured loaded regime; outside it the optimum moves, and no model predicts a job's operating point from its features.
• All published numbers are board numbers. Energy studies measure board power, but the grid meters the wall; host overhead (CPU, memory, cooling, power-supply losses) shifts the true optimum upward by an unmeasured amount; until it is placed at node level, no megawatt claim is grounded.
• Allocation and power control are not optimized together in the cited work. The measured 18–45 percent fleet saving used a fixed budget, a fixed placement rule and caps chosen afterwards; the joint problem, assigning jobs and setting every cap at once under hourly-changing limits, couples discrete scheduling to continuous control through the shared power budget and is unformulated.
• Flexibility is unquantified. How much power an AI facility can shed, how fast, for how long, at what cost has never been measured or standardized; Israel's regulator is rationing connections without the numbers to credit flexible ones.
• The incentive problem is identified but unsolved. Per-hour billing makes the renter's cheapest choice the energy-worst point, and a per-joule surcharge is too weak to fix it [1]; which structural instruments work is open.

2.5.5.4 Description of the proposed research

JouleFlex builds one measured foundation and three tools on it: first work out the energy-optimal way to run each job on each card, then control power to achieve it. WP1 and WP2 measure and predict: they extend the dataset to current hardware, production software, and whole-server power, then turn those measurements into models that place any card's or job's efficient point without a full sweep. WP3 controls: a real-time engine that decides which job runs on which card and at what power cap, tested in a digital twin before any deployment. WP4 and WP5 connect to the grid: they define and measure the flexibility a controlled facility can offer, and study the pricing and connection rules that make efficient operation worthwhile. Everything is released openly.

2.5.5.5 Methods (work packages)

Success metrics are set in advance: law fit and JP location per card and engine; transfer-model error on held-out cards; energy per served job and deadline compliance versus baselines; envelope violation, ramp and recovery in the demonstration; sensitivity of every economic figure to its price inputs.

2.5.5.6 Novelty and originality

• AI load as a grid resource. Sub-second actuation plus deadline-aware control turns an AI facility into a provider of measurable firm and flexible capacity, a category we have not found in practice or literature, immediately relevant to Israel’s connection challenge.
• From per-job search to one-time measurement, with an exact price. In the loaded regime, one per-card-model measurement replaces earlier per-job runtime search at under one percent penalty [1], and the extra cards to hold throughput equal the latency ratio, so every recommendation carries its exact capital cost.
• Energy first, power as the means. We measured the cost of universal full-power operation (about a third of inference energy) and located the alternative; existing capping satisfies watts while possibly raising joules, whereas JouleFlex sets power to the computed energy optimum per (card, job).
• From the board to the wall. Published GPU energy results are board-power results; measuring the node’s JP at server-input power moves the claim toward what the grid actually meters.
• An open dataset with no equivalent. To our knowledge no open dataset releases dense power-cap sweeps of energy per inference across many AI models and GPUs at the granularity to locate the energy-optimal point; existing open power-cap sweeps cover training or HPC workloads, and inference energy leaderboards report a single operating point, not a sweep. ELF fills that gap, and the project extends an asset that already exists rather than one it must first build.
• The joint allocation-and-capping problem. Assignment and power control were solved separately; coupled through the shared power budget they define a new scheduling-control problem, and the controller is a first measured solution.

2.5.5.7 Fit to the Ministry's R&D goals

The roles divide cleanly: operators run the hardware, the Ministry sets the rules. Just as a regulator sets fuel-economy standards without driving the cars, JouleFlex gives the Ministry the measured basis for connection terms, planning coefficients and efficiency standards, without operating any facility itself. The proposal addresses the Ministry's 2026 R&D policy [17] under call 34/2026 [18], in particular high-priority clauses of Chapter 2:

2.5.5.8 Preliminary results

The preliminary results are the applicant's study The Joule Point [1] and its ELF dataset: about 5,500 instrumented measurements over 20 models and four GPUs establish the response law (median R² = 0.99), locate the JP at 43 to 46 percent of TDP with a 29 to 31 percent energy saving per task at 1.2 times longer runs, show one static cap per card serving all twenty workloads at under one percent mean penalty, and demonstrate an 18 to 45 percent fleet-level saving at equal-or-better deadline compliance in a measured-curve simulation, with about 50 percent more work per megawatt behind a fixed GPU power budget. Cap actuation settles in about 191 milliseconds. Under the per-instance-hour billing model analyzed in [1], the renter’s cost optimum is the energy-worst operating point. ELF spans four GPU types; the fine sweeps behind the law and the JP cover three (A100, A10G, L4), with the T4 swept coarsely for coverage, and R² = 0.99 is the median per fixed-setting sweep while pooled per-card fits across workloads are 0.88 to 0.92 (Fig. A1). The benefit persists under tight service: in the strict-deadline experiment of [1], capping saved 28 percent of energy while keeping 95 percent of workload value. The harness, simulator and analysis pipeline ran end to end; the project scales them out rather than building from zero.

2.5.6 Application: expected benefit and impact

2.5.6.1 What the project delivers, concretely

The project delivers five artifacts: the ELF-2 open dataset, the characterization kit, the control engine, the digital twin, and the grid interfaces and policy report (open data formats for the facility's power-envelope commitment and delivery record, plus a regulator-facing report of measured coefficients and pricing). Appendix Table A1 states what each is and who uses it. Each artifact is released openly on delivery, every release crediting the Ministry’s R&D program under call 34/2026 [18].

2.5.6.2 Benefit and impact

For operators: the preliminary results indicate 29–31 percent lower energy per task (measured) and 18–45 percent lower per served job (fleet simulation), through software control on installed hardware.

Training: the project trains one postdoctoral researcher across GPU energy measurement, fleet control and energy-market analysis; the open ELF-2 dataset and twin become teaching material, building Israeli capacity at the AI-energy interface.

For the Israeli grid and regulator: the larger benefit. The binding constraint is connection capacity: about 27,000 MW of pending requests [2] strain planned generation through 2035. The JP converts efficiency into capacity: in the measured A100 case, capping delivered about 50 percent more work per megawatt of GPU power. Power shaping lets facilities accept conditional connections: firm plus flexible capacity curtailed on instruction. With flexibility measured, committed and audited (WP4), the regulator gains a third option and the frozen queue stops being all-or-nothing. The measured capacity gain is a GPU-board-level result; as an illustrative upper bound, every 1,000 MW of the queue connected under JP operation could deliver the AI work of up to 1,500 MW uncapped, and a workload that would consume 1 TWh uncapped would avoid about 0.3 TWh, roughly 150,000 tonnes of CO2 at an assumed grid intensity of 0.5 kg per kilowatt-hour; WP1 and WP4 measure how much of the gain survives at server input and facility level before any connection coefficient is recommended. The regulator-facing report quantifies this option and will be submitted as a formal response to the Authority’s open consultation on the flexible-connection track [19], placing measured numbers in the docket while the rules are written; the open release lets any operator or authority adopt or certify the mechanism.

2.5.7 Work plan

2.5.7.1 Stages, deliverables, duration and success metrics

Two interim briefings to the chief scientist unit are named deliverables: an M8 technical brief and the M21 policy report draft before publication, in Hebrew, circulable to the Authority and Noga while the flexible-connection track [19] is under design. Every milestone is an inspectable artifact tested against pre-stated criteria; a bounding result is itself a deliverable. Minimum-success path: H100/L40S-class validation, one production serving engine, server-input BMC measurement, the job and transfer models, the joint controller, and one Israeli flexibility scenario; B200-class cards, a second engine, fine-tuning, shared-GPU regimes and the formal approximation guarantee are extension objectives pursued as access and time allow.

2.5.7.2 Budget by stage

2.5.7.3 Risks and mitigation

2.5.8 The applicant

Project manager and principal investigator: Alexander Apartsin, Senior Lecturer, Computer Science, HIT; single-PI, responsible for all five work packages.

Fit of expertise to the proposed work. The scientific foundation is the PI's own study [1]: the ELF dataset, the law, the JP, the cost identity, the fleet saving. The work plan maps onto that record: power sweeps (WP1); response models with quantified uncertainty (WP2); simulators and schedulers against measured ground truth (WP3, WP4); techno-economic analysis with its incentive analysis already in [1] (WP5). The harness, models and simulation the work scales are built and validated; open release enables independent scrutiny. The project has no external partners; the Israel scenario runs on public Noga and Electricity Authority data. Research personnel. One postdoctoral researcher at 90% position, both years, carries the measurement, modeling and control work under the PI's supervision, reported within a month of signature (נספח טז' §2.5).

Resources available to the applicant. The project owns no hardware: board sweeps run on AWS EC2 with root access, the RAPL tier on AWS bare metal, and the wall tier and WP4 demonstration on hourly dedicated servers with BMC PSU telemetry. HIT provides laboratory space, computing and grant administration; the project buys no measurement hardware.

2.5.9 Funding

No other funding exists or is applied for; per נספח טז' §1.5.2 and §2.10, no complementary funding will be received from any other body and no student receives parallel state funding.

2.5.10 Bibliography

1. Apartsin, A. and Aperstein, Y. (2026). The Joule Point: an Energy-Optimal Operating Point for AI Inference. Preprint, September 2026. DOI: 10.13140/RG.2.2.20385.36964 (CC BY 4.0), with the released ELF dataset.
2. JNS (22 July 2026). Israel halts new data-center grid requests for 140 days amid surge in demand. Press report on the Electricity Authority's temporary order: pending applications of approximately 27,000 MW, about three times Israel's average electricity consumption. https://www.jns.org/news/israel-news/israel-halts-new-data-center-grid-requests-for-140-days-amid-surge-in-demand (the Authority's order is published on gov.il).
3. International Energy Agency (2025). Energy and AI: Energy Demand from AI. https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
4. Barroso, L. A. and Hölzle, U. (2007). The Case for Energy-Proportional Computing. IEEE Computer 40(12), pp. 33–37.
5. Yao, F., Demers, A., and Shenker, S. (1995). A Scheduling Model for Reduced CPU Energy. Proceedings of FOCS 1995, pp. 374–382.
6. Miyoshi, A., Lefurgy, C., Van Hensbergen, E., Rajamony, R., and Rajkumar, R. (2002). Critical Power Slope: Understanding the Runtime Effects of Frequency Scaling. Proceedings of ICS 2002, pp. 35–44.
7. Gandhi, A., Harchol-Balter, M., Das, R., and Lefurgy, C. (2009). Optimal Power Allocation in Server Farms. Proceedings of ACM SIGMETRICS 2009, pp. 157–168.
8. You, J., Chung, J.-W., and Chowdhury, M. (2022). Zeus: Understanding and Optimizing GPU Energy Consumption of DNN Training. arXiv:2208.06102. https://arxiv.org/abs/2208.06102
9. Chung, J.-W., Ma, J. J., Wu, R., et al. (2025). The ML.ENERGY Benchmark: Toward Automated Inference Energy Measurement and Optimization. arXiv:2505.06371. https://arxiv.org/abs/2505.06371
10. Chung, J.-W., Wu, R., Ma, J. J., et al. (2026). Where Do the Joules Go? Diagnosing Inference Energy Consumption. arXiv:2601.22076. https://arxiv.org/abs/2601.22076
11. Hu, Q., Sun, P., Yan, S., Wen, Y., et al. (2021). Characterization and Prediction of Deep Learning Workloads in Large-Scale GPU Datacenters. arXiv:2109.01313. https://arxiv.org/abs/2109.01313
12. Alibaba Group (2020). cluster-trace-gpu-v2020: production GPU cluster traces from Alibaba PAI. https://github.com/alibaba/clusterdata
13. Lipe, E., Karia, N., Espenshade, C., Stein, C., Tantawi, A., and Tardieu, O. (2026). Energy Efficient Scheduling of AI/ML Workloads on Multi-Instance GPUs with Dynamic Repartitioning. arXiv:2606.25082. https://arxiv.org/abs/2606.25082
14. Lu, T. and Reda, S. (2026). Agentic CPU-GPU Scheduling for Heterogeneous AI Workloads. arXiv:2607.22242. https://arxiv.org/abs/2607.22242
15. Rodriguez, C., Degioanni, L., Kameni, L., Vidal, R., et al. (2024). Evaluating the Energy Consumption of Machine Learning: Systematic Literature Review and Experiments. arXiv:2408.15128. https://arxiv.org/abs/2408.15128
16. Kwon, W., Li, Z., Zhuang, S., et al. (2023). Efficient Memory Management for Large Language Model Serving with PagedAttention. Proceedings of SOSP 2023. arXiv:2309.06180. https://arxiv.org/abs/2309.06180
17. Ministry of Energy and Infrastructure (2026). מדיניות התמיכה במחקר ופיתוח לשנת 2026. Chapter 1 clauses 6, 8, 15, 16; Chapter 2 priority list.
18. Ministry of Energy and Infrastructure (2026). קול קורא 34/2026 למימון מחקרים בתחום האנרגיה ומדעי האדמה והים.
19. Calcalist (June 2026). Report on the Electricity Authority's public hearing (שימוע): a dedicated connection track for transmission-grid consumers of 50 MVA and above, including a system-operator right to curtail consumption at peak and payments for holding connection capacity. Under public consultation; cited as regulatory direction, not settled regulation. https://www.calcalist.co.il/calcalistech/article/rylkyxemgx (the hearing documents are published on gov.il).

Appendix A: Figures (within the 5-page allowance of נספח א'1 §2.2)

Appendix B: Gantt, milestones and deliverables (for entry into the online system)

This appendix is entered directly into the Ministry's online Gantt system (נספח א'1 §3), and Appendix C is the data source for the Ministry's budget Excel (נספח טז'); neither forms part of the research-description figures appendix, whose 5-page allowance is used by Appendix A.

Schedule

Milestones with judgeable outputs

הערה להזנה במערכת המקוונת

נספח ג': מפרט תקציבי (בעברית, לפי נספח טז')

סיכום

בדיקת עמידה במגבלות הקול הקורא

פירוט לפי סעיפים

1. כוח אדם

מלגאי הבתר-דוקטורט יבצע את קמפיין המדידות (ELF-2), את מודל ההעברה, את פיתוח בקר עיצוב-ההספק ואת ניסויי הגמישות מול הרשת, ויהיה מעורב בכל אבני הדרך MS1 עד MS4 ובתוצרים D1 עד D3. חוקר ראשי אינו מקבל שכר (§2.6). יש לצרף לבקשה טבלאות שכר ואישור רו"ח חתום המאשר את עלויות השכר הנהוגות במוסד (§2.2).

2. חומרים אזילים וציוד מתכלה

3. ציוד

הפרויקט אינו רוכש ציוד מחשוב או מדידה. מדידת ההספק מתבצעת מטלמטריית היצרן (NVML לכרטיס, RAPL לצומת, BMC/Redfish לשקע) בחומרה שכורה; דיוק המדידה מדווח לפי מפרטי היצרן ומאומת בעקביות בין-שכבתית, ומיקום נקודת הג'ול (JP) אינו תלוי בהטיה כפלית קבועה של החיישן.

4. שונות

תקרות נסיעה לחו"ל: כרטיס טיסה הלוך ושוב עד 8,000 ₪, קצבת שהייה עד 250 ₪ ליום, קצבת לינה עד 650 ₪ ליום. מימון נסיעות סטודנטים הוא בהיקף 75% מעלות הנסיעה, ובשיתופי פעולה בינלאומיים 100% (§5.4.4, §5.4.6). כל נסיעה מחייבת בקשה מראש ובכתב; לא יאושרו בקשות בדיעבד (§5.4.8).

5. קבלני משנה ועבודות חוץ

שירותי מחשוב ענן לצורך הרצות מדידה מבוקרות על משפחות מאיצים שאינן זמינות במוסד. פלטפורמת הביצוע היא AWS: מופעי EC2 עם מאיץ מלא (GPU passthrough) והרשאות root, המאפשרים שליטה עדינה במגבלת ההספק ובמדידת האנרגיה ברמת המאיץ, לפי המחירון הפומבי של AWS (תמהיל on-demand ו-spot); תדפיס מחירון מתוארך יצורף כהצעת המחיר לפי §6.3. הרצות שאינן דורשות שליטה בהספק יכולות לרוץ גם אצל Modal Labs (modal.com), המספק גישה לעשר משפחות מאיצים שונות בחשבון אחד, בחיוב לפי שנייה וללא חיוב על זמן סרק. פירוט השעות והעלויות להלן מחושב מתדפיסי המחירון הפומבי של AWS המצורפים (נאספו 29.08.2026, אזור us-east-1): תעריפי spot לריצות סריקה הניתנות להפסקה, on-demand היכן שמצוין; מופעים מרובי-מאיצים מתומחרים לשעת GPU; כרטיסי דור חזית מתוכננים לפי תעריף H100 ומתומחרים בפועל לפי המחירון בעת הריצה (§6.3):

סעיף זה פטור מתקורה (§7.1). התשלום מבוצע כנגד חשבוניות (§6.4). תעריפי AWS נלקחו מנקודות הקצה הפומביות של מחירון AWS (on-demand ו-spot) ב-29.08.2026, ותדפיסים מתוארכים מצורפים; תעריפי שרתי BMC לפי מחירון ספק מפורסם (תדפיס מצורף). שער החליפין נבדק (כ-2.96 ₪ לדולר) ולצורכי תכנון נעשה שימוש בשער שמרני של 3.20 ₪ לדולר

6. תקורה

תקורה בשיעור 15% מהתמורה עבור הוצאות המחקר, למעט סעיף קבלני משנה, בהתאם ל-§7.1. התקורה מכסה הוצאות עקיפות ובכללן שירותי מזכירות, ראיית חשבון, משאבי אנוש, תחזוקת חשבונות מחקר, ניהול המחקר, שימוש בספריות, גישה לשירותי מחשב, שימוש במתקני מחקר ובמשרדים, ומים, אנרגיה ולוגיסטיקה. הוצאות אלה אינן נרשמות בנפרד במפרט (§7.3).

הצהרות נדרשות

• המוסד לא קיבל ולא יקבל מימון נוסף או משלים לביצוע המחקר או עבור המועסקים בו מכל גוף שאינו המשרד (§1.5.2).
• לא יתקיים כפל מימון לסטודנטים המועסקים במחקר מכל גורם ממשלתי אחר (§2.10).
• עלות הכנת ההצעה חלה על המבקש בלבד (§1.5.3).