TokenBank: Financial Infrastructure for AI Services

Cary Chang; Jialin Zhou

cs.AIOct 8, 2026

TokenBank: Financial Infrastructure for AI Services

Authors: Cary Chang, Jialin Zhou

Organizations: Nexilume Research

Abstract

AI services incur inference costs during execution, while revenue may arrive later. Changing API prices, limited upfront capital, and service failures can limit operators' ability to sustain or expand their services. Beyond reducing per-request costs, operators need to plan future spending, fund execution before revenue arrives, and obtain compensation for specified losses. This requires clear agreements across services with different pricing and execution conditions. These agreements must distinguish rights to consume services from rights to receive payments, define obligations under uncertain costs and income, and specify which failures qualify for compensation and how much can be paid. We present TokenBank, a financial infrastructure that represents these commitments through structured contracts. It supports service-consumption rights, agreements that settle API-price differences in cash (forwards), financing through limited rights to future service revenue, and protection claims for specified service failures. Contracts specify participants, covered services, validity, ownership, fulfillment conditions, and settlement rules. Evaluation combines replay of 899,441 API requests, real model-driven agent execution, and contract API tests. In a zero-discount rising-price resampling scenario, forwards reduce mean expenditure by USD 304.88 but increase its standard deviation from USD 1,152.45 to USD 1,190.82. A controlled replication with five portfolios per capital condition finds mean contribution differences between financing and self-funding of +1.0635, -0.1406, and -0.2962 experimental USD under low, baseline, and ample capital, respectively. The evaluation distinguishes contract correctness from economic effectiveness under declared economic and failure assumptions; supplier invoices and commercial revenue are unavailable.

Figures & tables

Figure 1: Caching, batching, model routing, and failover improve request execution. TokenBank complements these approaches with service commitments that specify what is purchased, API-price forwards, revenue-right financing, and protection against covered service failures. These mechanisms address the financial consequences of AI-service use rather than replacing execution optimization.

Figure 2: Overview of TokenBank . Service commitments, API-price forwards, revenue-right financing and transfer, and service-failure protection use product-specific workflows supported by shared accounts, contract records, and settlement functions. Usage, price, revenue, and incident records provide the corresponding evidence for fulfillment and payment. Supported revenue-right transfers update the holders considered for subsequent distributions. The lifecycle summarizes available operations, not a mandatory sequence shared by all products.

Figure 3: Supported transfers change ownership while preserving source references, and financial effects are recorded through balanced accounting entries.

Figure 4: API-price forward mechanism in TokenBank . Participants agree on contract terms, fund fees and margin, and settle the resulting price-difference payment against a reference price. The forward is financially separate from API procurement. Its expenditure effect depends on contract charges and on how closely the contracted amount and reference price match realized API purchases.

Figure 5: Revenue-right financing and transfer in TokenBank . Primary investment supplies capital to an Agent or Provider operator in exchange for rights to specified future revenue. Recorded service revenue is later divided among the publisher, investor pool, and platform, with investor allocations determined from funding agreements and current holdings. Existing rights may subsequently be transferred: the buyer pays the selling holder, while future distributions follow the updated holdings.

Figure 6: Service-failure protection in TokenBank . A Provider-funded reserve and buyer protection fees support coverage under predefined plan terms. Monitoring and request records establish incident evidence, coverage and impact rules determine an eligible claim amount, and approved claims are paid as service credit subject to available reserve funding. Financial compensation is evaluated separately from retry and failover recovery.

Study	Cases	Calls	Responses	Successes
Repeated capital study	540	1,082	1,081	456
Controlled repeats and parameter tests	624	1,247	1,246	487
Initial controlled study	84	150	150	51
Retail contract study	70	930	925	1
Retail baseline	10	368	360	4

Table 1: Real-model study populations. Cases count scheduled task executions; calls include all sent requests. Responses and successful services follow their respective measurement rules. Each study retains its own denominator.

Symbol	Definition	Value
$\mu$	Mean known expense per scheduled calibration task	0.127689
$H_{\mathrm{cal}}$	Largest amount set aside before a calibration request	0.943300
$B$	Baseline own capital, $\mu+H_{\mathrm{cal}}$	1.070989
$A$	Added capital, $\lceil 4\mu\rceil_{0.01}$ (rounded up to a cent)	0.520000
$r$	Assumed revenue per successful service, $2\mu$	0.255378
$d$	Scheduled tasks between releases of earned revenue	4

Table 2: Controlled-task parameters fixed from the nine calibration tasks. Amounts are experimental USD. The repeated capital study changes the starting own balance to

0.5B

B

, or

1.5B

and keeps added funds

A

fixed.

Rule tested	Result
Forward payment and retry	Quantity 10 and an index change from 1.0 to 1.3 produce one 3.000000 transfer.
Funding and restart	A 0.060000 transfer leaves 0.040000; retry changes nothing and a further unfundable transfer is rejected.
Payment after a right is sold	Subsequent revenue follows the current holder rather than the former investor.
Resale to multiple holders	Resale to two holders across organizations preserves revenue allocation.
Covered service	A claim for a model service not covered by the contract is rejected.
Payment in full	One 60.000000 payment leaves 40.000000; another 60.000000 claim is not partially paid.

Table 3: Contract validation through business APIs on isolated PostgreSQL accounts. Synthetic funds exercise payment, ownership, funding, and idempotency rules; economic outcomes are evaluated in the subsequent product studies.

Scenario	$h$	Std. dev. (USD)	Mean $\Delta C$ (USD)	Over budget (%)
flat	0	1144.71	+0.00	62.2
flat	1	1144.71	+26.49	63.3
rise	0	1152.45	+0.00	72.1
rise	1	1190.82	-304.88	61.4
fall	0	1139.84	+0.00	54.1
fall	1	1100.08	+357.85	65.5

Table 4: Spending with no assumed price discount, a matching reference index, the unchanged calibration forecast, and sufficient funds for payments.

h

is the fraction of forecast spending covered. Negative mean

\Delta C

means lower spending than no hedge. Results come from resampled 42-day paths.

Figure 7: Average spending and risk measures without an assumed price discount. Bills and prices are jointly resampled in seven-day blocks. This plot uses the unchanged-forecast, funded-payment settings; the full analysis covers all 300 settings.

	Completed services			Contribution difference
Capital	Own	Equal	Right	Right–own	Right–equal
$0.5B$	0.0	10.4	10.2	1.0635	-0.3049
$B$	10.6	12.0	12.0	-0.1406	-0.3174
$1.5B$	12.0	12.0	12.0	-0.2962	[-0.2742, -0.0975]

Table 5: Five separate runs per setting. Completed services are means out of twelve. Own, Equal, and Right denote own capital, equal capital without investor sharing, and revenue-right financing. Contribution differences are experimental USD, excluding initial funding and credits. Brackets show the range due to missing usage, not variation across repetitions.

Figure 8: Five runs at each starting capital level. Left: completed services under the three financing settings, with mean bars. Right: contribution with revenue-right financing minus contribution with own capital. Any nonzero whiskers bound missing charges; they are not resampling ranges. Tasks within a run share funds and are not independent repetitions.

Faults / recovery	Success	Eligible loss	Paid credit	Benefit $v=1$	Benefit $v=0$
0 / no	12	0.0000	0.0000	-0.0100	-0.0100
0 / yes	12	0.0000	0.0000	-0.0100	-0.0100
6 / no	6	0.2900	0.2900	0.2800	-0.0100
6 / yes	10	0.0900	0.0900	0.0800	-0.0100
12 / no	0	0.5500	0.5500	0.5400	-0.0100
12 / yes	9	0.1400	0.1400	0.1300	-0.0100

Table 6: Fault tests with a per-claim limit of 0.13, premium of 0.01, and initial reserve of 1.56. Loss and credit amounts are experimental USD. Paid credit is issued through the experimental contracts; benefit assumes either full use (

v=1

) or no use (

v=0

) of that credit.

Recovery	Claim limit	Initial funds	Approved	Paid	Unpaid
no	0.1300	1.5600	0.5500	0.5500	0.0000
no	0.1300	0.1300	0.5500	0.1300	0.4200
no	0.0300	0.1300	0.3600	0.1200	0.2400
no	0.0300	1.5600	0.3600	0.3600	0.0000
yes	0.1300	1.5600	0.1400	0.1400	0.0000
yes	0.1300	0.1300	0.1400	0.1400	0.0000

Table 7: Claim limits and available funds with twelve scheduled faults and a premium of 0.01. Approved, paid, and unpaid columns report total amounts in experimental USD, not numbers of claims. Initial funds exclude the premium. Loss above the claim limit is not counted as an unpaid approved claim.

Figure 9: Premium sensitivity of buyer value under full service-credit utilization, a 0.13 per-claim limit, and a 1.56 initial reserve. Net value follows Equation ( 49 ).

$D$	$F$	$S$	Success	Contribution	Service credit
0	0	0	9	0.6034	0.0000
0	0	1	9	0.5934	0.1400
1	0	0	9	0.9772	0.0000
1	0	1	9	0.9672	0.1400
0	1	0	9	0.3735	0.0000
0	1	1	9	0.3635	0.1400

Table 8: Eight combinations for recovery-enabled runs with ample funds, at an assumed price index of 1.5 and hedge ratio of 0.5.

D

F

, and

S

switch forwards, financing, and protection on or off. Contribution is experimental USD, excluding initial funding and service credits. Issued credits are reported separately.

Appendix figures & tables36 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 10: Contract settlement in isolated database experiments: revenue allocation after a rights transfer (left) and dedicated-reserve payments (right). Controlled initial funds and synthetic revenue events exercise the corresponding ownership and funding rules.

Figure 11: Concurrent transaction latency across the 19 tested scenarios. The reference line marks the prespecified 1,500 ms budget; red bars indicate scenarios above that budget.

Figure 12: Financing outcomes for synthetic task sequences with continuous balances. Scripted task counts and contribution quantify the accounting effects of the specified funding and receipt assumptions.

Figure 13: Reserve balances and payment outcomes in deterministic protection tests. Each submitted claim is checked against its cap and available funds. The horizontal axis indexes eligible claims separately for each condition: 12 without recovery and 3 with recovery.

Family	Model-directed operation	Completion criterion
Extraction	Extract a reference, quantity, and city; submit the record	Submitted fields and final JSON match the request.
Order lookup	Invoke the order-lookup tool	The requested order is retrieved and final JSON matches its state.
Product selection	Query the catalog; select and quote an item	The quoted item is the cheapest eligible in-stock item; quantity, total, and final JSON are correct.

Appendix

Table 9: Controlled task families and completion criteria. Each family contains three calibration tasks and four evaluation tasks.

Component	Measurement and interpretation
Model and tool execution	Responses, returned usage, and completed operations are observed on the controlled task set.
Business records and faults	Constructed records and local fault schedules define the controlled execution conditions.
Prices and receipts	Experimental tariffs, success-triggered receipts, and terminal price indices define the scenario valuation.
Contract settlement	Account changes and service-credit payments are observed in isolated database experiments and reconciled independently.
Study population	Fixed task families define the repeated-study population; calibration, controlled evaluation, and retail cohorts retain separate denominators.

Appendix

Table 10: Measurement scope of the controlled evaluation.

Role	Responses	Input tokens	Output tokens	Mean latency (s)
Agent	237	1,658,417	13,796	9.87
Simulated user	121	157,622	7,009	9.49
Evaluator	2	9,604	180	4.19

Appendix

Table 11: Real fixed evaluation model usage by role. Simulated-user and evaluator costs belong to the experimental budget rather than Agent operating expenditure. Latency covers returned responses; task duration also includes failed attempts and retry delays.

Figure 14: Retail baseline token usage and task outcomes. Repeated attempts are paired by task identifier; task success follows the official completion rule.

Task/repetition	Outcome	Tools (errors)	Agent tokens	Seconds
26/0	Success	10 (2)	140,455	373.7
26/1	Success	10 (2)	128,610	247.7
25/0	Success	10 (4)	126,529	385.0
25/1	Unscored	17 (6)	223,946	600.0
72/0	Failure	10 (9)	197,961	515.1
72/1	Failure	14 (6)	147,962	314.7

Appendix

Table 12: Outcomes of all prespecified task attempts under the official completion rule. Tool-error counts and total evaluation time are reported separately.

Figure 15: Measurement organization connecting model-driven execution, database contract experiments, and scenario-based economic valuation.

Figure 16: Conditional expense versus hedge ratio on the own-capital recovery trajectory under the declared terminal-price scenarios.

Figure 17: Zero-discount forward comparison on the recovery-enabled own-capital arm’s measured returned demand. Calibration-fixed notionals and declared price indices determine actual API settlements and fees, which enter the conditional expenditure calculation.

Figure 18: Paired forward scenarios applied to measured returned Agent usage. Declared price and notional scenarios compare matched and output-only movements to characterize settlement alignment with operating expense.

Figure 19: Forward expenditure sensitivity to the covered amount, forecast error, and assumed price path. Exposure scales the signed settlement while the executed workload remains fixed.

Scenario	$\operatorname{Var}(Z)$	$2\operatorname{Cov}(C_{0},Z)$	$\Delta\operatorname{Var}$	$\Delta\mathbb{E}[C]$
flat	0.0	0.0	+0.0	+26.49
rise	4522.5	-85383.2	+89905.7	-304.88
fall	4522.5	93585.4	-89063.0	+357.85
spike	47158.8	-80862.3	+128021.1	-304.10
lagged-load	0.2	413.1	-412.8	+25.08

Appendix

Table 13: Mechanism of risk changes in conditional resampled paths. Negative variance change is favorable; mean expenditure and variance are distinct outcomes.

Capital	Contrast	Mean / cost bound	Resampling envelope	$+/-$ blocks
$0.5B$	Right–own	1.0635	[0.9506, 1.1786]	5/0
$0.5B$	Right–equal	-0.3049	[-0.3324, -0.2806]	0/5
$0.5B$	Equal–own	1.3684	[1.2642, 1.4894]	5/0
$B$	Right–own	-0.1406	[-0.2657, -0.0155]	2/3
$B$	Right–equal	-0.3174	[-0.3507, -0.2835]	0/5
$B$	Equal–own	0.1769	[0.0506, 0.3138]	4/1

Appendix

Table 14: Contribution contrasts in the matched capital replication. The mean column reports missing-usage bounds where non-degenerate. The envelope is the exploratory whole-block 95% resampling range, not another bound on a single missing request. The last column counts definitely positive/negative blocks out of five. Units are experimental USD; principal and credit are excluded.

Portfolio	Own tasks	Right tasks	$\Delta$ tasks	$\Delta$ profit	Recovery $\Delta$ tasks
1	11	12	1	-0.1670	9
2*	9	1	-8	[-2.1573, -1.2153]	9
3	12	12	0	-0.3065	9
4	10	12	2	0.0632	9
5	10	12	2	-0.0799	9

Appendix

Table 15: Fresh portfolio contrasts. An asterisk identifies financing arms with unplanned request errors or missing usage. Brackets bound unresolved usage under the reservation limit. Recovery compares local recovery with none. Monetary values are experimental USD.

Figure 20: Every observed whole-portfolio contrast. Whiskers bound unresolved usage under the reservation limit; dashed lines are known-expense means, not task-level confidence estimates.

Condition	Own	Equal	Right	$\Delta$ profit	Right–equal
Baseline	11	12	12	-0.1670	-0.3065
Receipt multiplier 1	10	12	12	-0.1826	-0.2081
Receipt multiplier 1.5	11	12	12	-0.2085	-0.2298
Receipt multiplier 3	11	12	12	-0.2487	-0.4597
Capital multiplier 0.5	0	11	11	1.1919	-0.2277
Capital multiplier 1.5	12	12	12	-0.3065	-0.2516

Appendix

Table 16: Fresh financing boundaries: completed service counts and contribution differences. The profit difference compares revenue rights with own capital. Equal funding isolates capital availability from the investor share.

Figure 21: Financing can change service completion and retained contribution differently. Each bar is a fresh scenario comparison.

Arm	Success	Budget stop	Expense	Retained	Profit
Own capital	9	3	1.1692	2.2754	1.1062
Equal capital	12	0	1.4749	3.0339	1.5590
Revenue right	12	0	1.4749	2.7274	1.2525

Appendix

Table 17: Fresh financing executions; 12 scheduled tasks per arm. All monetary values are experimental USD. Profit excludes capital injection and uses known expense.

Figure 22: Continuous operating funds, completed services, contribution and investor cash flow. Source: the controlled cohort’s reconciled ledger and task records.

Arm	Scored	Budget stops	Known cost	Held funds
own	1	9	0.147368	0.000000
equal	4	6	0.515037	0.000000
right	0	9	0.342738	0.172938

Appendix

Table 18: Fresh financing interventions. Money is in experimental USD; receipts and tariffs are assumptions. Capital injections are excluded from contribution profit.

Figure 23: Operating balances, admission and scoring outcomes, and measured expenditure across financing conditions. Known expense and unresolved request reservations are shown separately.

Figure 24: Same-workload financing cost and conditional growth requirement. The right panel gives the analytical break-even requirement under stationary cost per success.

Initial budget	Budget ($/day)	Agent (%)	Provider (%)
50th percentile	3.05370	+21.23	+32.59
75th percentile	41.65998	-8.11	+0.50
90th percentile	166.25893	-19.17	-11.59

Appendix

Table 19: Contribution-profit change when financing doubles the daily budget. Each row compares the same demand at the stated initial budget with and without financing. Full requests are preserved.

Figure 25: Contribution-profit changes at the 75th-percentile initial budget, varying additional funding and investor share. Zero-change contours show where additional service revenue offsets sharing for each publisher.

Arm	Success	Reached	Loss	Funded	Scarce
Own / no recovery	0	12	0.5600	0.5600	0.1400
Own / recovery	9	12	0.1400	0.1400	0.1400
Right / no recovery	0	12	0.5400	0.5400	0.1200
Right / recovery	9	12	0.1400	0.1400	0.1400

Appendix

Table 20: Twelve scheduled cases per fault arm. Reached counts injected incidents. Funded and scarce columns are actual service-credit payments, not cash income.

Figure 26: Completed services and reserve-funded loss allocation under controlled local faults. Monetary values use the stated laboratory tariff.

Arm	Successes	Faults	Residual cost	Eligible	Funded paid	Scarce paid
Own / off	0	8	0.356410	0.080000	0.080000	0.030000
Own / on	1	8	1.511219	0.020000	0.020000	0.020000
Right / off	0	8	0.426050	0.080000	0.080000	0.030000
Right / on	0	8	1.063093	0.020000	0.020000	0.020000

Appendix

Table 21: Retail fault interventions and service-credit settlements. Off/on denotes local recovery. Eligible loss excludes semantic task errors and unknown usage.

Figure 27: Post-incident responses, scored completions, official task success, and redistribution of covered residual expense. Recovery changes executions; protection transfers service-credit value. Amounts are reconciled across task outcomes, claim decisions, reserve accounts, and service-credit records.

Figure 28: Reserve sufficiency under different initial-capital levels, with

\rho=0.75

, a 1% protection fee, and no claim delay. Panels report available reserves and unpaid capped claims; initial capital is accounted for separately from premium receipts.

DFS	Successes	Known-cost contribution	Service credit
000	1	-2.036122	0.000000
001	1	-2.046122	0.020000
100	1	-1.586435	0.000000
101	1	-1.596435	0.020000
010	0	-1.594640	0.000000
011	0	-1.604640	0.020000

Appendix

Table 22: Eight combinations: actual task outcomes and contract payments under an assumed price scenario.

Configuration	Forward discount	Economic expense ($)
000: baseline	—	3,438.59
100: forward only	0%	3,459.88
100: forward only	10%	3,282.46
111: all three	10%	3,605.09

Appendix

Table 23: Fixed-volume comparison under constant prices. Codes indicate whether

D

F

, and

S

are enabled.

Configuration	Recorded measurement scope
Direct-provider pilot	Ten scored trials and no successful tasks.
Gateway-format pilot	Ten Agent requests and ten returned simulated-customer responses; the customer responses match gateway records.
Authenticated pilot	Three scored trials with no successes, three unscored trials, and four unattempted scheduled trials; all 181 returned responses reconcile with gateway records.
Fixed-schedule pilot	Two scored trials, including one success.
Recovery configuration pilot	Two scored trials, including one success.
Shortened pilot	No completed trial.

Appendix

Table 24: Preliminary retail cohorts and recorded measurement scope. These cohorts are separate from the final baseline estimate.

Configuration	Requests	Responses	Scored
Grounding	53	52	2
Reasoning, 60 s timeout	6	0	0
Reasoning, 240 s timeout	6	0	0
Reasoning, final batch	21	21	2

Appendix

Table 25: Supplementary configuration batches, each scheduling two tasks. Request counts include all model roles; scored tasks have an official outcome. Returned usage is accounted for separately from task scores.

Explore similar work

Jun 10, 2026cs.AI

AI Tokenomics: The Economics of Tokens, Computation, and Pricing in Foundation Models

Tokens have become the practical accounting unit for modern foundation model services, linking information processing, computation, memory use, energy expenditure, pricing, and economic value. This paper develops a framework for AI tokenomics: the study of how tokens are generated, consumed, priced, allocated, and optimized across AI systems. We connect token-level technical costs to workflow-level production functions, enterprise resource allocation, measurement and instrumentation methods, and emerging market-design questions. The framework shows that token expenditure and economic value are distinct: value depends on marginal productivity, workflow position, hidden reasoning activity, risk, and downstream propagation effects. The paper concludes by identifying open research directions in hidden-token measurement, empirical calibration, token productivity, dynamic allocation, and token-based markets.

Quanyan Zhu

Department of Electrical and Computer Engineering, New York University Tandon School of Engineering, 370 Jay Street, Room 1004, Brooklyn, NY 11201, USA

Cost-Aware Inference Token Efficiency

Ask PDF

Apr 24, 2026cs.CL

How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks

The wide adoption of AI agents in complex human workflows is driving rapid growth in LLM token consumption. When agents are deployed on tasks that require a significant amount of tokens, three questions naturally arise: (1) Where do AI agents spend the tokens? (2) Which models are more token-efficient? and (3) Can agents predict their token usage before task execution? In this paper, we present the first systematic study of token consumption patterns in agentic coding tasks. We analyze trajectories from eight frontier LLMs on SWE-bench Verified and evaluate models' ability to predict their own token costs before task execution. We find that: (1) agentic tasks are uniquely expensive, consuming 1000x more tokens than code reasoning and code chat, with input tokens rather than output tokens driving the overall cost; (2) token usage is highly variable and inherently stochastic: runs on the same task can differ by up to 30x in total tokens, and higher token usage does not translate into higher accuracy; instead, accuracy often peaks at intermediate cost and saturates at higher costs; (3) models vary substantially in token efficiency: on the same tasks, Kimi-K2 and Claude-Sonnet-4.5, on average, consume over 1.5 million more tokens than GPT-5; (4) task difficulty rated by human experts only weakly aligns with actual token costs, revealing a fundamental gap between human-perceived complexity and the computational effort agents actually expend; and (5) frontier models fail to accurately predict their own token usage (with weak-to-moderate correlations, up to 0.39) and systematically underestimate real token costs. Our study offers new insights into the economics of AI agents and can inspire future research in this direction.

Longju Bai, Zhemin Huang, Xingyao Wang +5

University of Michigan · Stanford University · Microsoft AI +3

Coding Agents Cost-Aware Inference

Ask PDF

May 2, 2026cs.AI

Agentic AI Systems Should Be Designed as Marginal Token Allocators

This position paper argues that agentic AI systems should be designed and evaluated as \emph{marginal token allocation economies} rather than as text generators priced by the unit. We follow a single request -- a developer asking a coding agent to fix a failing test -- through four economic layers that today are designed in isolation: a router that decides which model answers, an agent that decides whether to plan, act, verify, or defer, a serving stack that decides how to produce each token, and a training pipeline that decides whether the trace is worth learning from. We show that all four layers are solving the \emph{same} first-order condition -- marginal benefit equals marginal cost plus latency cost plus risk cost -- with different index sets and different prices. The framing is deliberately minimal: we do not propose a complete theory of AI economics. But adopting marginal token allocation as the shared accounting object explains why systems that locally minimize tokens globally misallocate them, predicts a small set of recurring failure modes (over-routing, over-delegation, under-verification, serving congestion, stale rollouts, cache misuse), and points to a concrete research agenda in token-aware evaluation, autonomy pricing, congestion-priced serving, and risk-adjusted RL budgeting.

Siqi Zhu

University of Illinois Urbana-Champaign

LLM Agent Orchestration LLM Agent Evaluation LLM Agent Training Agentic AI

Ask PDF