Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?

Amirreza Zeinali; Jon Saad-Falcon; Avanika Narayan; Mersad Abbasi; Jonathan S. Hartley; John Hennessy; Christopher Re; Azalia Mirhoseini

cs.LGOct 22, 2026

Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?

Authors: Amirreza Zeinali, Jon Saad-Falcon, Avanika Narayan, Mersad Abbasi, Jonathan S. Hartley, John Hennessy, Christopher Re, Azalia Mirhoseini

Organizations: Stanford University · University of California, Berkeley · The University of Texas at Austin

Abstract

An AI system can produce more economically valuable work while delivering less value for the energy it consumes. We therefore ask: How much economic value does an AI system create per unit of energy consumed? We formalize economic value in two ways: (1) net dollar value, the output's dollar value minus AI-system execution cost, and (2) time saving, the expert's completion time minus the AI system's completion time. Dividing each by measured end-to-end energy yields Net Value per Joule (NVJ) and Time Saving per Joule (TSJ), reported in USD/MJ and hours/MJ. We study 220 public GDPval tasks across 44 occupations, serving five open-weight models locally and evaluating four accelerator types. First, the occupation groups where AI systems succeed most often, deliver the greatest dollar value, or save the most time do not necessarily provide the greatest economic return per joule. For example, among the tested tasks, engineering work delivers about 10% more net dollar value per task than software development. Yet software development delivers about 27% more net value per joule. Second, GLM 5.2 delivers the highest average net dollar value, at $72.27 per task using 0.430 MJ, equivalent in energy to about four hours of active laptop use. By contrast, Gemma 4 31B delivers 56% of that net dollar value using just 23% of the energy, achieving approximately 2.5 times the NVJ and 2.8 times the TSJ. Third, across four single-accelerator deployments, Qwen3.6-27B's NVJ varies by a factor of 4.32 and TSJ by 9.27, with B200 leading both. Additionally, one and eight H100s deliver similar net dollar value, but eight reduce completion time by 22% at 3.24 times the energy. Our framework makes output value and expert time saving under an energy budget explicit objectives for improving models, accelerators, and software, while offering a new basis for assessing AI's potential economic effect on occupational work.

Figures & tables

Figure 1: Overview of Joules-to-GDP. (a) NVJ and TSJ express dollar and expert time benefits per joule. (b) The study compares occupations, models, and accelerators. (c) Occupation returns vary widely, with different NVJ and TSJ leaders. (d) More valuable output can require more energy; the frontier uses unweighted totals on log scales. The bottom panel uses 2.52 MJ, equivalent to one H200 drawing 700 W for an hour or about 23 hours of active laptop use.

Figure 2: An illustrative GDPval example shows the evaluation workflow. An Accountants and Auditors task, Fall Music Tour P&L, compares the expert-only output with the AI-agent output through a blinded LLM-as-a-Judge comparison. A win or tie earns output dollar value and time saving; every attempt incurs AI-system execution cost and energy.

Figure 3: Energy changes occupation rankings, and each return metric favors different work. GLM 5.2 against the expert-only reference, with category-normalized GDP weights. Top panels show win–tie rate, net dollar value, and time saving per task; bottom panels normalize each by energy. Categories follow the win–tie-per-MJ order; whiskers show 95% task-composition intervals.

Figure 4: An expert-with-AI reference reduces GLM 5.2’s NVJ by 70% and TSJ by 65%. Only the expert deliverable and its completion time change; AI runs, costs, energy, and GDP weights stay fixed. Whiskers show 95% task-composition intervals (Appendix K.1 ).

Figure 5: Higher total economic benefit can come with lower return per joule. For example, GLM 5.2 delivers more total value and time saving than Gemma 4 31B, but has lower NVJ and TSJ. The top row compares GDP-weighted NVJ (a) and TSJ (b); the bottom shows unweighted total output dollar value (c) and time saving (d) versus total energy. Dashed steps connect models on each point-estimate efficient frontier. Whiskers show 95% task-composition intervals.

Model	Execution cost (USD)	Completion time (min)	Energy (MJ)	dollar of AI cost (USD/USD)	AI hour (USD/hour)	megajoule (USD/MJ)
	GDP-weighted task averages			Net value per
GLM 5.2	1.92	22.7	0.430	37.69	191.3	168.1
Qwen3.6-27B	4.12	51.0	0.516	15.46	75.0	123.4
Gemma 4 31B	0.86	7.7	0.097	47.38	316.9	417.5
GPT-OSS 120B	1.98	22.2	0.137	19.09	102.1	275.7
GPT-OSS 20B	1.12	15.3	0.103	31.03	136.3	336.0

Table 1: GLM 5.2 and GPT-OSS 120B have similar cost and completion time but different energy returns. AI-system execution cost, AI-system completion time, and energy are GDP-weighted task averages; the three efficiency ratios share each system’s net-dollar-value numerator. Colors identify systems. Bars show point estimates on a separate zero-to-maximum scale within each efficiency column; longer bars mean greater net value per unit of resource.

Figure 6: Accelerator choice changes return per joule. Qwen3.6-27B, one accelerator, expert-only reference, and GDP weights. Whiskers show 95% task-composition intervals.

Figure 7: Eight H100s deliver similar value but lower return per joule than one. Qwen3.6-27B, expert-only reference, and GDP weights. Whiskers show 95% task-composition intervals. Appendix M gives the full comparison.

Appendix figures & tables25 assets

Supplementary material from the paper’s appendix.

Appendix

Pair	Exact agreement	Binary win–tie agreement
Gemini–Haiku	77.2%	79.2%
Gemini–GPT-5 mini	74.0%	76.4%
Haiku–GPT-5 mini	72.3%	72.8%

Appendix

Table 2: The three judges agree on 72–77% of exact outcomes. An AI win or tie receives task credit. Exact agreement retains win/tie/loss; binary win–tie agreement collapses wins and ties into one category. All comparisons use GLM 5.2 outputs from four attempts per task.

Model and accelerator	Inference	Tensor parallelism	Concurrent requests	Context limit (tokens)	Output limit (tokens)
GLM 5.2, 8 $\times$ H200	vLLM	TP8	20	131,072	32,768
Qwen3.6-27B, 8 $\times$ H100	vLLM	TP8	20	131,072	32,768
Gemma 4 31B, 8 $\times$ H100	vLLM	TP8	20	131,072	32,768
GPT-OSS 120B, 8 $\times$ H100	vLLM	TP8	20	131,072	32,768
GPT-OSS 20B, 8 $\times$ H100	vLLM	TP8	20	131,072	32,768
Qwen3.6-27B, 1 $\times$ B200	vLLM	TP1	20	131,072	32,768

Appendix

Table 3: All evaluated deployments use vLLM, 20 concurrent requests, a 131,072-token context limit, and a 32,768-token output limit. Tensor parallelism is eight for eight-accelerator deployments and one for single-accelerator deployments.

Figure 8: Request-completion share does not track batch energy proportionally. P20, P50, P80, and P90 mark completion of 20%, 50%, 80%, and 90% of requests. Each point pools batch totals after estimating energy consumed at that milestone. Intermediate points use each batch’s average power and recorded completion times; the 100% endpoint is measured batch energy. The dotted line shows proportional completion for reference.

Category	Energy (MJ)	NVJ (USD/MJ)	TSJ (hours/MJ)
Computer & Mathematical	0.355 / 0.337	523.8 / 551.6	7.29 / 7.67
Architecture & Engineering	0.495 / 0.597	413.9 / 343.2	7.89 / 6.54
Management	0.360 / 0.349	374.6 / 387.0	6.75 / 6.97
Protective Service	0.225 / 0.229	278.1 / 272.9	4.96 / 4.86
Production	0.550 / 0.786	221.1 / 154.6	6.12 / 4.28
Community & Social Service	0.282 / 0.363	188.5 / 146.3	6.16 / 4.78

Appendix

Table 4: Equal batch shares change the TSJ leader but leave the NVJ leader unchanged. Each cell shows active-request allocation / equal batch share. Active-request allocation is the paper’s primary method. AI outputs, judgments, expert times, wages, costs, and GDP weights are fixed; only energy assignment changes. Values are point estimates.

	AI execution cost (USD)			NVJ (USD/MJ)
Category	CoreWeave	AWS	GCP	CoreWeave	GCP
Architecture & Eng.	1.677	2.104	2.819	414.8	412.5
Arts & Media	1.582	1.986	2.660	104.7	102.3
Business & Financial	1.731	2.172	2.911	146.4	144.1
Community & Social	0.983	1.233	1.652	189.4	187.0
Computer & Math	1.245	1.562	2.092	524.6	522.3

Appendix

Table 5: Published node tariffs change calculated costs without changing category NVJ order. For the same recorded GLM 5.2 tasks, only the node-hour price is changed; task outcomes, time, energy, and non-node charges remain fixed. Values are hypothetical costs per category-normalized task and the resulting NVJ, not results of executions on other providers. The AWS column reproduces the evaluated cost boundary.

Figure 9: Wages determine output dollar value without also increasing benchmark weights. Sector GDP shares and within-sector employment shares determine occupation weights. Category comparisons renormalize those weights to one within each category, while the GDP-weighted benchmark metric retains category weights. Expert completion time and occupation-specific median wages determine output dollar value.

Category	Win–tie (%)	Win–tie pp/MJ
Architecture & Eng.	69.00 [38.01, 91.00]	139.43 [75.84, 229.50]
Arts & Media	46.09 [28.27, 66.09]	102.44 [74.15, 136.56]
Business & Financial	56.81 [37.93, 74.97]	112.42 [72.81, 168.62]
Community & Social	60.00 [20.00, 100.00]	212.94 [60.19, 390.26]
Computer & Math	80.00 [40.00, 100.00]	225.08 [122.22, 312.77]
Healthcare	28.23 [8.02, 63.12]	63.48 [21.77, 107.89]

Appendix

Table 8: Protective Service leads win–tie rate per MJ, while Personal Care leads win–tie rate. GLM 5.2, expert-only reference, one attempt per task; occupation outcomes are category-normalized with GDP-based weights. Brackets show 95% task-composition intervals.

Category	Net dollar value (USD/task)	Time saving (hours/task)
Architecture & Eng.	204.84 [108.93, 288.94]	3.905 [2.055, 5.569]
Arts & Media	46.71 [20.18, 79.15]	1.219 [0.554, 2.028]
Business & Financial	73.57 [44.46, 106.85]	1.677 [0.975, 2.508]
Community & Social	53.12 [12.78, 93.45]	1.735 [0.423, 3.047]
Computer & Math	186.16 [68.11, 337.13]	2.589 [0.924, 4.844]
Healthcare	43.50 [14.21, 86.93]	0.742 [0.230, 1.541]

Appendix

Table 9: Architecture & Engineering leads net dollar value and time saving per task. GLM 5.2, expert-only reference, one attempt per task; occupation outcomes are category-normalized with GDP-based weights. Brackets show 95% task-composition intervals.

Category	NVJ (USD/MJ)	TSJ (hours/MJ)
Architecture & Eng.	413.90 [196.34, 744.61]	7.890 [3.660, 14.397]
Arts & Media	103.82 [55.09, 167.63]	2.710 [1.459, 4.508]
Business & Financial	145.57 [83.22, 239.58]	3.318 [1.820, 5.627]
Community & Social	188.50 [43.10, 382.42]	6.158 [1.426, 12.522]
Computer & Math	523.76 [235.93, 770.14]	7.285 [3.182, 11.080]
Healthcare	97.84 [38.77, 196.72]	1.669 [0.643, 3.049]

Appendix

Table 10: Computer & Mathematical leads NVJ; Architecture & Engineering leads TSJ. GLM 5.2, expert-only reference, one attempt per task; occupation outcomes are category-normalized with GDP-based weights. Brackets show 95% task-composition intervals.

Category	Output dollar value (USD)	Expert time (h)	AI time (h)	Time saving (h)
Architecture & Eng.	299.90	6.11	0.45	5.66
Arts & Media	105.66	3.02	0.37	2.65
Business & Financial	133.31	3.34	0.39	2.95
Community & Social	90.58	3.22	0.32	2.89
Computer & Math	234.65	3.67	0.43	3.24
Healthcare	162.29	3.04	0.41	2.63

Appendix

Table 11: Win–tie values and expert times explain why the category rankings differ. Values condition on outputs judged a win or tie for GLM 5.2 against the expert-only reference. Occupation outcomes are category-normalized with GDP-based weights.

Category	Win–tie pp/MJ rank	NVJ rank	TSJ rank
Architecture & Eng.	8	2	1
Arts & Media	10	10	9
Business & Financial	9	8	8
Community & Social	5	6	5
Computer & Math	3	1	2
Healthcare	12	11	12

Appendix

Table 12: Different categories lead win–tie rate per MJ, NVJ, and TSJ. Point-estimate ranks for GLM 5.2 against the expert-only reference, with one attempt per task and category-normalized GDP-based weights; rank 1 is highest.

Category	Win–tie (%)	Credited value $G$ (USD)	AI cost $C$ (USD)	Time saving $S$ (hours)	Energy $E$ (MJ)	AI time $L$ (hours)
Architecture & Eng.	69.0	206.94	2.104	3.905	0.495	0.467
Arts & Media	46.1	48.70	1.986	1.219	0.450	0.384
Business & Financial	56.8	75.74	2.172	1.677	0.505	0.443
Community & Social	60.0	54.35	1.233	1.735	0.282	0.321
Computer & Math	80.0	187.72	1.562	2.589	0.355	0.420
Healthcare	28.2	45.81	2.306	0.742	0.445	0.299

Appendix

Table 13: A high win–tie rate alone does not determine credited value or time saving. Category-normalized GDP-based task averages, expert-only reference. Credited value and time saving are zero on losses; AI-system execution cost and energy include every attempt. Applying

(G-C)/E

and

S/E

to the underlying unrounded components reproduces the NVJ and TSJ point estimates in Table 10 .

Rank	Category	Net value/AI dollar (USD/USD)	Net value/AI hour (USD/hour)
1	Computer & Math	119.2	442.9
2	Architecture & Eng.	97.4	438.4
3	Management	85.6	423.6
4	Protective Service	64.0	225.8
5	Production	52.1	199.3
6	Community & Social	43.1	165.7

Appendix

Table 14: Computer & Mathematical leads net value per AI dollar and per AI hour across occupation categories. GLM 5.2 against the expert-only reference, one attempt per task. Category-normalized GDP-based averages are combined before forming

(G-C)/C

and

(G-C)/L

, where

G

is credited output dollar value,

C

is AI-system execution cost, and

L

is AI-system completion time in hours. All attempts contribute cost and time. Point estimates use the underlying unrounded components summarized in Table 13 ; categories are ordered by net value per AI dollar.

Occupation	Tasks	Win–tie (%)	NVJ (USD/MJ)	TSJ (h/MJ)
Accountants and Auditors	5	40.0	31.5	0.53
Administrative Services Managers	5	100.0	369.8	6.60
Audio and Video Technicians	5	40.0	22.7	0.53
Buyers and Purchasing Agents	5	40.0	249.3	6.69
Child, Family, and School Social Workers	5	60.0	188.5	6.16
Compliance Officers	5	80.0	333.9	8.45

Appendix

Table 15: Occupation-level economic returns vary widely across the evaluated work. Expert-only reference and one evaluated attempt per task. Win–tie is the fraction of an occupation’s evaluated tasks judged a win or tie. Each occupation’s task components are averaged before ratios are formed. These are point estimates; all attempts contribute cost and energy.

Ranking	Median $\rho$	5th–95th $\rho$	Leader retained
Win–tie rate per MJ	0.951	0.896–0.989	51.0%
NVJ	0.962	0.824–0.995	76.5%
TSJ	0.918	0.813–0.973	40.4%

Appendix

Table 16: The NVJ leader survives one-task deletions more often than the TSJ leader. Across 10,000 independently drawn portfolios, one task is chosen uniformly from each category and the remaining tasks are re-averaged within their occupations. GDP occupation weights and recorded outcomes and resource use stay fixed. Percentile ranges describe these perturbations, not confidence intervals.

Judgment reference	Time reference	NVJ (USD/MJ)	TSJ (hours/MJ)
Expert-only	Expert-only	168.11 [126.00, 222.43]	3.823 [2.870, 5.021]
Expert-with-AI	Expert-only	69.07 [46.04, 95.84]	1.818 [1.256, 2.486]
Expert-only	Expert-with-AI	84.90 [64.79, 107.49]	1.990 [1.432, 2.639]
Expert-with-AI	Expert-with-AI	50.04 [30.77, 70.41]	1.340 [0.814, 1.888]

Appendix

Table 17: Expert-with-AI judgments and completion times each lower reported NVJ and TSJ. For GLM 5.2 under

\mathrm{pass@}1

, the mixed-reference rows separate judgment and valuation effects. Brackets show percentile 95% task-composition intervals.

Change and fixed quantity	NVJ reduction (%)	TSJ reduction (%)
Judgments; expert-only reference times fixed	58.9 [43.0, 72.7]	52.5 [37.0, 66.5]
Times; expert-only reference decisions fixed	49.5 [33.1, 63.0]	47.9 [29.1, 63.2]
Times; expert-with-AI reference decisions fixed	27.5 [-1.7, 50.8]	26.3 [-6.2, 51.8]
Judgments; expert-with-AI reference times fixed	41.1 [20.9, 60.4]	32.7 [7.2, 54.9]
Both judgments and times	70.2 [54.4, 82.5]	65.0 [46.3, 79.4]

Appendix

Table 18: Changing both judgments and times reduces NVJ by 70.2% and TSJ by 65.0%. Every change substitutes the expert-with-AI reference for the expert-only reference. Values are percentage reductions [95% interval]; a negative endpoint permits an increase.

Figure 10: More runs and higher four-run win–tie thresholds both lower measured NVJ and TSJ, but for different reasons. GLM 5.2 uses the expert-only reference. Left: pass@1, pass@2, and pass@4 include the energy of one, two, and four runs per task. Right: the end-to-end energy of four runs is fixed while the four-run win–tie threshold increases. Whiskers are percentile 95% task-composition bootstrap intervals.

Figure 11: Automation adds time saving beyond expert assistance, but the balance shifts with the success threshold. At lower thresholds, most saved hours come from automation; at stricter thresholds, more come from expert assistance. GLM 5.2 tasks meet each stated win–tie threshold against the expert-with-AI reference. The left panel reports GDP-weighted hours; the right divides each transition’s saved hours by completion time at its starting expert reference. All four runs’ energy is retained. Percentages across transitions use different denominators and should not be added.

Policy	Output value (USD)	AI cost (USD)	Energy (MJ)	NVJ [95% CI]	TSJ [95% CI]	Net value/ AI cost
$\mathrm{pass@}1$	74.19	1.92	0.430	168.1 [126.0, 222.4]	3.82 [2.87, 5.02]	37.69
$\mathrm{pass@}2$	85.75	4.00	0.891	91.8 [71.0, 118.6]	2.13 [1.65, 2.72]	20.45
$\mathrm{pass@}3$	89.41	5.98	1.332	62.7 [48.9, 80.0]	1.46 [1.16, 1.85]	13.96
$\mathrm{pass@}4$	93.42	7.92	1.760	48.6 [38.3, 61.7]	1.16 [0.93, 1.45]	10.79

Appendix

Table 19: Four attempts increase output dollar value but reduce NVJ and TSJ. GLM 5.2 uses the expert-only reference. Every included run incurs cost and energy; output dollar value is counted once. NVJ is USD/MJ and TSJ is hours/MJ. Brackets give conditional 95% task-composition intervals.

Policy	Output value (USD)	AI cost (USD)	Energy (MJ)	NVJ [95% CI]	TSJ [95% CI]	Net value/ AI cost
$\geq 1/4$	93.42	7.92	1.760	48.6 [38.3, 61.7]	1.16 [0.93, 1.45]	10.79
$\geq 2/4$	81.58	7.92	1.760	41.9 [31.7, 54.9]	1.02 [0.79, 1.31]	9.30
$\geq 3/4$	71.53	7.92	1.760	36.1 [26.3, 48.8]	0.89 [0.67, 1.17]	8.03
$4/4$	57.12	7.92	1.760	28.0 [18.4, 40.5]	0.72 [0.50, 0.99]	6.21

Appendix

Table 20: Stricter four-run thresholds reduce credited output dollar value and returns at fixed energy. GLM 5.2 uses the expert-only reference. Every included run incurs cost and energy; output dollar value is counted once. NVJ is USD/MJ and TSJ is hours/MJ. Brackets give conditional 95% task-composition intervals.

Policy	Output value (USD)	AI cost (USD)	Energy (MJ)	NVJ [95% CI]	TSJ [95% CI]	Net value/ AI cost
$\geq 1/4$	35.11	7.92	1.760	15.4 [10.5, 20.7]	0.46 [0.32, 0.61]	3.43
$\geq 2/4$	27.70	7.92	1.760	11.2 [6.1, 16.7]	0.37 [0.23, 0.52]	2.50
$\geq 3/4$	7.71	7.92	1.760	-0.1 [-1.9, 2.1]	0.10 [0.05, 0.16]	-0.03
$4/4$	3.48	7.92	1.760	-2.5 [-3.7, -1.1]	0.05 [0.01, 0.09]	-0.56

Appendix

Table 21: Stricter four-run thresholds turn NVJ negative against the expert-with-AI reference. GLM 5.2 incurs cost and energy for every included run; output dollar value is counted once. NVJ is USD/MJ and TSJ is hours/MJ. Brackets give conditional 95% task-composition intervals.

Transition	Time saving (hours)	Time saving (%)	Energy (MJ)	TSJ (hours/MJ)
Expert-only $\rightarrow$ expert-with-AI	0.21 [-0.04, 0.45]	23.0 [-5.3, 45.9]	–	–
Expert-with-AI $\rightarrow$ AI system	0.58 [0.34, 0.83]	83.7 [77.8, 87.1]	0.430	1.340 [0.814, 1.888]
Expert-only $\rightarrow$ AI system	0.78 [0.55, 1.03]	87.4 [84.4, 89.8]	0.430	1.818 [1.256, 2.486]

Appendix

Table 22: Most measured time saving occurs between the assisted expert reference and autonomous AI output. GDP-weighted task-average decomposition under

\mathrm{pass@}1

. Bracketed values are percentile 95% task-composition bootstrap intervals.

Figure 12: Greater output dollar value need not mean greater economic return per joule. The five models use GDP-weighted aggregates, the expert-only reference, and one attempt per task. NVJ accounts for execution cost as well as energy; TSJ captures expert time saving per unit of energy. Whiskers are percentile 95% task-composition bootstrap intervals.

Model	GPUs	Output value (USD)	AI cost (USD)	Energy (MJ)	NVJ [95% CI]	TSJ [95% CI]	Net value/ AI cost
GLM 5.2	8 $\times$ H200	74.19	1.92	0.430	168.1 [126.0, 222.4]	3.82 [2.87, 5.02]	37.69
Qwen3.6-27B	8 $\times$ H100	67.81	4.12	0.516	123.4 [92.2, 163.2]	2.43 [1.73, 3.28]	15.46
Gemma 4 31B	8 $\times$ H100	41.38	0.86	0.097	417.5 [239.5, 642.9]	10.86 [6.76, 15.72]	47.38
GPT-OSS 120B	8 $\times$ H100	39.73	1.98	0.137	275.7 [152.5, 429.2]	6.38 [3.71, 9.63]	19.09
GPT-OSS 20B	8 $\times$ H100	35.88	1.12	0.103	336.0 [180.7, 536.4]	8.26 [4.67, 12.59]	31.03

Appendix

Table 23: GLM 5.2 leads output dollar value; Gemma 4 31B leads NVJ and TSJ among five models. The models share the evaluated tasks, the expert-only reference, and a one-attempt policy; values are GDP-weighted task averages. NVJ is USD/MJ, TSJ is hours/MJ, and net value per dollar of AI cost divides net dollar value by AI-system execution cost. Brackets give conditional 95% task-composition intervals. Cost inputs and allocation are in Appendix G .

AI and Labor

AI and labor studies how artificial intelligence changes jobs, wages, skills, hiring, and the organization and quality of work.

See all 95 papers

Feb 23, 2026

Photons = Tokens: The Physics of AI and the Economics of Knowledge

Debates about artificial intelligence capabilities and risks are often conducted without quantitative grounding. This paper applies the methodology of MacKay (2009) -- who reframed energy policy as arithmetic -- to the economy of AI computation. We define the token, the elementary unit of large language model input and output, as a physical quantity with measurable thermodynamic cost. Using Landauer's principle, Shannon's channel capacity, and current infrastructure data, we construct a supply-and-demand balance sheet for global token production. We then derive a finite question budget: the number of meaningful queries humanity can direct at AI systems under physical, information-theoretic, and economic constraints. We apply Coase's theory of the firm and the durable-goods monopoly problem to the AI value chain -- from photon to atom to chip to power to token to question -- to identify where economic value concentrates and where regulatory intervention is warranted. We argue that the expansion of the token budget does not resolve a deeper constraint: under structural uncertainty, the decisive variable is not how many questions can be answered but which questions are worth asking -- a problem of agency and direction that computation alone cannot solve. We connect limits of measurement in the token economy to a structural parallel between Goodhart's law and the Heisenberg uncertainty principle, and to Arrow's impossibility result for efficient information pricing. The framework yields order-of-magnitude estimates that discipline policy discussion: at current efficiency, the projected 2028 US AI energy allocation of 326~TWh could support roughly

6.5 \times 10^{17}

tokens per year, or 225,000 tokens per person per day -- more than three orders of magnitude above estimated mid-2024 utilization.

Alec Litowitz, Nick Polson, Vadim Sokolov

QStar Capital · University of Chicago · George Mason University

AI Governance

Ask PDF

May 20, 2026

Energy per Successful Goal: Goal-Level Energy Accounting for Agentic AI Systems

Current AI energy benchmarks measure consumption per model invocation. For agentic systems, where a single user goal may trigger multistep orchestration, tool calls, retries, and failure recovery, the invocation count is an implementation artifact rather than a task property, and per inference normalization misrepresents the energy cost of completing a goal. We present A-LEMS (Agentic LLM Energy Measurement System), a cross layer framework that redefines the unit of AI energy accounting as Energy per Successful Goal (EpG): total workflow energy across all attempts, including failures and retries, divided by successfully completed goals. A-LEMS combines a temporal boundary model, a five layer attribution pipeline mapping hardware signals (RAPL on x86; SPBM and DCGM on ARM) to workflow energy, and a reproducibility protocol binding every measurement to hardware, environment, and runtime configuration. Building on EpG, we define the Orchestration Overhead Index (OOI), the ratio of agentic to linear EpG on matched goals under identical success criteria. Across nine task families on Intel x86 and NVIDIA Grace GB10 ARM, CPU only OOI ranges 0.35x to 11.07x. Including GPU dynamic energy, which accounts for 83.8% of agentic EpG in local inference, widens this to 0.4x to 26.48x, so CPU only measurement understates overhead about twofold for reasoning tasks. For remote inference, network wait energy shifts OOI from 2.5x to 2.35x. The overhead arises from orchestration structure (planning, retries, recovery) that induces additional inference calls, not from costlier individual calls. The metric is not biased upward: single tool calculator dispatch inverts to OOI = 0.61x (95% CI [0.35, 0.97]) on x86. Goal level accounting via EpG and OOI provides a measurement foundation for benchmarking agentic AI.

Deepak Panigrahy, Aakash Tyagi, Stephen Abkin +2

Independent Researcher, USA · Texas A&M University, Department of Computer Science and Engineering, USA

LLM Agent Orchestration Agentic Workflows AI Agent Benchmarks

Ask PDF

Jul 22, 2026

Google's AI & Economy ATLAS v1.0: Mapping Gemini Usage in the Economy

This paper introduces the AI & Economy ATLAS (Activity, Task, Landscape, and Adoption Study), an ongoing economic research initiative using Google AI usage data. The first iteration of ATLAS is built on 15 million de-identified interactions across the Gemini App, Google AI Mode, and Gemini API. Using privacy-preserving algorithms as well as established and bespoke classification methods, we map AI usage to over 800 occupations, 4000 tasks, 300 household activities, 150 countries, and 140 languages. We then make a number of observations on what the data reveals about AI's diffusion, and its usage at work and in day-to-day life. In the workplace, we show that while AI adoption spans occupations covering just above 88% of US employment, penetration remains shallow and overwhelmingly collaborative in nature, with end-to-end task automation limited in scope. Outside of work, AI spans activities making up about 98% of Americans' non-sleep time, with disproportionately high use in high-friction tasks such as engaging with government and professional service providers, likely delivering economic value that standard national accounts may miss. Globally, adoption scales with national wealth and has broad linguistic distribution, with English queries representing only around a third of volume. As we build upon ATLAS and expand its scope and capabilities, we will continue to provide large-scale empirical evidence to inform the public, policy and academic questions about the ongoing AI transformation.

Zanna Iscenko, Scott Strand, Yiyuan Chen +15

Google · Google DeepMind

Generative AI Adoption Human-AI Collaboration AI and Labor

Ask PDF

Abstract

AI and Labor

Explore similar work

Photons = Tokens: The Physics of AI and the Economics of Knowledge

Energy per Successful Goal: Goal-Level Energy Accounting for Agentic AI Systems

Google's AI & Economy ATLAS v1.0: Mapping Gemini Usage in the Economy