# Decision Markets: From Intelligence to Action

## The thesis

**Capital is building the infrastructure for increasingly capable AI. As that intelligence becomes cheaper to deploy, choosing what it should do—and allocating resources to those choices—could become a more important bottleneck. Decision markets are a candidate for that missing coordination layer.**

The argument runs from capital to intelligence to collective action. Funding conditions shape which projects can attract capital and how much liquidity investors want to retain. AI changes what those investments can produce. Decision markets could connect competing forecasts to funded decisions.

## I. Where capital could flow, 2026–2036

The capital-flow thesis links three uses of funding: **building intelligence**, **deploying it in productive businesses**, and **coordinating the decisions those businesses and agents make**. All three already exist and can grow together. The argument is a chain of dependencies, not a timetable in which one stage must end before the next begins. Sustained investment ultimately needs useful returns.

Infrastructure investment is already tangible. The IEA reports capital expenditure above $400B by five large technology companies in 2025, with a 75% increase expected in 2026. These company totals are not pure-AI spending. Power, chips and financing constrain expansion, while dependence on capital markets makes the pace sensitive to expected returns and funding conditions. [5]

Capital can follow those constraints into compute, power, cooling and networking. Demand alone does not establish an attractive investment return; utilization, competition and financing costs matter.

Financing is part of the constraint. As data-centre projects grow beyond what company balance sheets can fund, their pace becomes more sensitive to capital-market conditions and expectations for AI returns. A useful project can still be delayed if it cannot obtain financing on workable terms. [5]

Our application to the next decade is that **AI investment and demand for liquidity can rise at the same time**. Some investors finance long-duration projects; others hold liquid reserves against uncertainty. If returns disappoint or funding tightens, refinancing becomes harder and investors can shorten commitments, sell risk exposures and seek settlement assets. These responses can overlap with continued investment in projects whose expected returns still justify the commitment.

| Use of capital across 2026–2036 | Where incremental funding could go | What would sustain the flow | What could interrupt it |
|---|---|---|---|
| Build capacity | Compute, electricity access, data centres, cooling and network equipment. | Demand that supports utilization and financing costs. | Overbuilding, power delays, expensive funding or weaker returns. |
| Deploy capability | Businesses using AI to deliver products, research, software and operational work; the data and payment infrastructure they require. | Measurable productivity, customer spending and retained margins. | Deployment costs, unreliable performance or competition absorbing the gains. |
| Coordinate activity | Allocation mechanisms, evaluation, execution controls and evidence services for organizations and agents. | Enough consequential decisions to justify specialized infrastructure. | Centralized systems solving the problem adequately, poor metrics or markets that cannot support informed participation. |
| Preserve liquidity | Cash and short-duration instruments; selected non-debtor stores of value depending on the risk being hedged. | Need for settlement, optionality and resilience to funding stress. | Inflation, issuer risk or an opportunity cost that makes longer commitments more attractive. |

This table describes overlapping uses of capital and possible spending destinations. It does not measure transfers between asset buckets or predict their timing. Primary financing can fund new activity; buying an existing asset usually pays its seller. Bank credit can create deposits, while market capitalization can change without an equal cash inflow. [1]

The accompanying diagram contrasts two funding conditions. Easier financing can support longer commitments and more investment; tighter financing can increase demand for cash and make new commitments more selective. These are conceptual tendencies, not a forecast of a synchronized global cycle. Funding liquidity means access to cash or credit; market liquidity means being able to trade without a large price impact. They interact, but are not interchangeable.

The decade growth chart uses global data-centre electricity demand as a proxy for one physical requirement of the buildout, including non-AI computing. Its source anchors are 485 TWh in 2025 and the IEA’s projection of 950 TWh in 2030. [5] We interpolate a 2026 baseline and explore 0%, 5% or 10% annual growth after 2030 through 2036. These extensions are our assumptions, not IEA forecasts. The chart shows electricity-use scenarios. It does not measure installed computing capacity, capital flows, intelligence or investment returns.

Stablecoins sit in the payment and settlement part of this picture. Issuer minting and redemption affect supply; secondary trading need not. Reserve assets and access to redemption matter, particularly under stress. [2][3] Our frozen September 2026 series measures roughly $310.4B of USD-pegged stablecoin value. That is a stock, unlike annual capital expenditure, and the two must not be added together. The optional decade chart provides settlement context, not a forecast of AI investment or futarchy adoption. [4]

## II. AI changes what capital can do

AI first appears in the capital story as something expensive to build. It then becomes a tool that can lower the cost of producing software, research and operational work. If more capable systems can reliably plan and execute longer tasks, their role expands again: they become participants in economic activity, acting within budgets and mandates.

That transition adds an allocation problem to the existing resource constraints. An organization might be able to generate thousands of technically plausible projects while having enough capital, power, time or access to pursue only a few. More intelligence expands the choice set. It does not eliminate the need to choose.

A capable model can recommend an investment. That leaves several institutional questions unresolved: whose objective it is optimizing, which information it lacks, how another party can challenge its forecast, who bears the downside and whether the selected plan is carried out. A more persuasive answer does not itself settle disagreements among parties with different information or incentives.

**The next bottleneck could therefore be coordination: selecting among competing uses of intelligence and capital.** Conventional firms, managers, contracts and investment committees already perform this function. Decision markets would have to improve on those arrangements in specific settings. They are especially interesting where knowledge is dispersed, outcomes can be measured and outside participants can contribute information that the decision maker does not possess.

AI could also make those markets more practical. In his 2024 crypto-and-AI essay, Vitalik Buterin argues that AI participants could reduce the cost of researching small questions that do not attract human traders. He distinguishes agents participating under a mechanism's rules from using an AI as the rule-setting judge, which creates additional attack risks. These are proposed opportunities and cautions, not demonstrations of reliable autonomous governance. [9]

## III. Decision markets could connect intelligence to action

A prediction market prices claims about an outcome. A decision market uses conditional markets to compare expected outcomes under different actions. Futarchy uses a specified market comparison to help determine which action is taken, while the objective and measurement rules remain institutional choices. Hanson's work develops that separation and its design problems. [6]

Consider an illustrative organization with a $10M budget and two research programmes. AI systems can propose experiments, estimate costs and forecast results. A decision market would organize disagreement around a more precise question: which programme is expected to achieve the higher research-evaluation score at an agreed date if funded? Choosing the evaluator and the score is part of the governance problem; a market cannot make an inadequate measure meaningful.

Participants take positions tied to the alternatives under the mechanism's settlement rules. A governing process can use the resulting comparison to allocate the budget. Execution and outcome records then make it possible to examine what followed. Only the chosen programme produces an observed result in this example. Contracts must specify what happens to positions on the unchosen alternative; Hanson’s proposal uses called-off trades for unrealized conditions. This does not reveal the missing counterfactual. [6]

The attraction is a loop: **propose → compare → allocate → execute → evaluate**. AI can help generate and investigate the options. A market can give informed disagreement an economic channel. Contracts and operating systems implement the authorized choice. Evidence supports evaluation and subsequent decisions.

Buterin's later essay on information finance describes markets designed around information that a user needs, including comparisons between decisions. It argues that cheaper AI participation could make smaller questions economical, while retaining a role for trustworthy non-market judgment. [10] Our inference is that this combination could support an economy with far more delegated decisions than people can review individually.

This is the sense in which decision markets could become a **keystone**: a connecting institution between abundant analysis and scarce resources. Their value would come from helping organizations act on useful disagreement, not simply from producing another forecast or a more active token market.

Decision markets do not create the funding required to act, remove liquidity risk or guarantee a good investment. Their proposed contribution is better use of information when choosing among feasible alternatives.

A research allocator, grant programme, treasury or agent operator could adopt this loop for bounded decisions. None has to delegate the whole organization at once. Nor does every decision justify a market: the improvement must outweigh research, liquidity, delay and operating costs.

## IV. Institutions for an economy of intelligence

The important distinction is between **creating a capable intelligence** and **organizing an economy in which capable intelligences act**. AGI refers here to broadly general capabilities; ASI to capabilities exceeding human performance across a wide range of important tasks. We use these as conceptual terms, not claims that a threshold has been reached.

There is no evidence in the sources reviewed that decision markets are a necessary or sufficient condition for AGI or ASI. A system could become highly capable within a centrally directed organization. Markets do not supply the learning architecture, define human values or guarantee that an agent follows its mandate.

The stronger and more defensible thesis is institutional: **decision markets could be part of the missing infrastructure for turning advanced AI into accountable economic agency**. They could provide a way to compare competing plans, allocate bounded resources and expose forecasts to challenge, while other institutions define objectives and enforce constraints.

| Function | What decision markets could contribute | What must still come from elsewhere |
|---|---|---|
| Choosing among plans | Conditional estimates that make disagreement actionable. | Clear alternatives, an appropriate decision rule and disclosure of material terms. |
| Allocating resources | A signal that can inform or trigger a bounded allocation. | Authorization, custody, budgets, contracts and execution controls. |
| Evaluating agents | Incentives and records that may help distinguish useful forecasting or research. | Reliable outcomes, independence checks and resistance to metric gaming. |
| Aligning with human purposes | Information about consequences relative to an agreed objective. | The objective itself, legitimacy, rights, safety constraints and technical alignment. |

This is not only a caution about terminology. It identifies the actual product opportunity. If the goal is a capable model, a decision market is optional. If the goal is a large network of independently controlled agents allocating real resources, a robust coordination mechanism becomes essential—and decision markets are one candidate.

They would fail that role if many agents merely repeat one source, if participants cannot take corrective exposure, if outside benefits reward manipulation more than honest trading, or if the outcome measure can be gamed. A conditional price spread is also not automatically a causal effect: selection into a branch can reveal information. [6] More capable agents can exploit these weaknesses as well as help detect them.

The thesis therefore predicts a need for better specification, participation, evidence and execution around markets. It does not predict that intelligence makes those requirements disappear.

## V. What would make the thesis real this decade?

Existing evidence covers two different things: capital raised by a platform and a bounded market-based allocation experiment. The September 9 research snapshot recorded 21 companies and more than $45M raised on MetaDAO's founder page. A September 10 recheck reports 22 companies and the same $45M+ headline. The letter is dated August 12, 2026, but that publication date does not date the current tally; terms are set per company. That is platform-reported financing adoption, not proof of recurring autonomous decision-making. [8]

Uniswap Foundation's July 2025 conditional-funding-market pilot selected a $100K grant recipient with 29 forecasters and more than $70K traded. Its retrospective also identifies insider influence and forecasting complexity. This establishes a concrete allocation experiment; it does not prove superiority to a committee or generalize to AGI governance. [7]

A stronger test would compare decision quality and total costs with an appropriate alternative, such as a committee or an AI-assisted investment team, on comparable tasks. Trading volume, treasury size and repeat use alone cannot establish superiority.

The next decade would strengthen the thesis if three things become visible. First, AI-assisted participation reduces the cost of obtaining useful, independent information. Second, organizations repeatedly allocate through these mechanisms because the full process improves their decisions. Third, executed outcomes remain inspectable, so failures change future practice rather than disappear behind the next proposal.

One plausible scenario is gradual adoption for bounded choices. An acceleration scenario sees reusable data and agent tooling make smaller decisions economical. A stalled scenario sees conventional organizations absorb AI successfully while decision markets remain a niche. A stress scenario sees AI investment disappoint, funding tighten and corrective participation withdraw. These broad scenarios are not themselves probabilistic forecasts. The accompanying September 10 forecast register instead proposes three narrower, dated tests with explicit subjective probabilities and resolution rules. They test reported financing adoption, one grant programme’s repeat use and physical electricity demand separately. The register is labeled draft until its publication is recorded; no result has resolved.

The financing and governance questions meet here, but remain distinct. Financing conditions affect which commitments are feasible; decision markets concern how choices among them are informed. Capital can fund intelligence rapidly while leaving the institutions around it fragile. The scale of an AI buildout or a treasury does not establish the ability to govern it. The decisive evidence will be repeat use, decision quality after costs and credible execution—not the amount of capital that initially entered.

## VI. Investing in an economy of intelligence

**The ambition is to make the path from an investment thesis to a funded decision—and from that decision to its outcome—open to examination.** If intelligence expands what we can build, the institutions around capital must help us choose what is worth building.

Investors need to understand the opportunity, the terms of their exposure and the conditions under which their thesis fails. A compelling account of technological progress is only the beginning. The work is to connect it to a specific use of capital, a plausible source of returns and evidence that can change the decision.

The tools for this economy should make that work easier. They should help people discover opportunities, compare competing plans, investigate market expectations and follow the commitments made with their capital. Sources, assumptions and uncertainty should travel with the investment thesis, so a new reader can challenge it and its author can revise it.

That record should survive the investment. What was funded? What did the organization promise? What changed? What did it deliver? Human investors and delegated agents need a way to carry those questions across research, allocation and review. Better tools should be judged by whether they improve those decisions after the costs of using them.

Decision markets offer one way to put competing judgments at stake. Their promise depends on informed participation, meaningful outcomes and institutions that carry authorized choices into practice. Building those conditions is part of the investment opportunity itself.

**Capital is building intelligence. We should build the means to invest with judgment, challenge how capital is used and learn from what follows.**

## Forecast audit — September 10, 2026

The published stablecoin simulator was replayed at every eligible historical weekly origin, using only its input prefix, a 52-week lookback, four-week blocks, 5,000 paths and the existing seed. All three regimes had higher mean absolute error than carrying the latest value forward at 1-, 4-, 13- and 26-week horizons. The default 104-week lookback leaves no test outcomes in the 105-reading input. These are descriptive results on a frozen September snapshot; overlapping outcomes are dependent and historical values may have been revised. They do not validate a ten-year projection. Full predictions, outcomes, hashes and evaluation source accompany the site. The test has not established a forecasting advantage, and it says nothing directly about decision-market adoption.

The new forecast register contains initial subjective probabilities, not calibrated model estimates. Its exact questions, source cutoffs, evidence deadlines and missing-evidence rules must be retained across revisions. Publication must be recorded before claiming a prospective record; no future outcome can be scored now. Research costs and comparative decision quality after costs remain unpriced tests. A yes on financing or electricity cannot substitute for evidence on those claims.

## Sources

Evidence reviewed September 9, 2026. Reported observations, source projections and this report's scenarios are distinguished in the text. The capital-allocation categories are an original qualitative outlook, not measured flows or investment recommendations. No source cited establishes decision markets as a prerequisite for AGI/ASI. The historical stablecoin dataset remains frozen; the AI-infrastructure discussion uses the IEA's newer April 2026 report.

1. Bank of England. Money creation in the modern economy. Quarterly Bulletin, 2014 Q1. https://www.bankofengland.co.uk/quarterly-bulletin/2014/q1/money-creation-in-the-modern-economy
2. Bank for International Settlements. Stablecoins: framing the debate. April 20, 2026. https://www.bis.org/speeches/20260420-stablecoins-framing-debate
3. Circle. How minting and redemption works. Issuer documentation. https://developers.circle.com/circle-mint/concepts/how-minting-works
4. DefiLlama. Frozen stablecoin API observations, retrieved September 9, 2026. The downloadable input identifies the field and source hash. https://stablecoins.llama.fi/stablecoincharts/all
5. International Energy Agency. Key Questions on Energy and AI: Executive summary. April 2026; reported 2025 data and subsequent estimates. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary
6. Robin Hanson. Shall We Vote on Values, But Bet on Beliefs? 2013 paper, author's manuscript. https://hanson.gmu.edu/futarchy2013.pdf
7. Porter Geer / Uniswap Foundation. Case Study: Piloting Conditional Funding Markets for Unichain Growth. September 10, 2025. https://www.uniswapfoundation.org/blog/unichain-cfm-pilot-case-study
8. Charlie Dalton / MetaDAO. Letter to Founders. August 12, 2026; September 10 recheck reports 22 companies and $45M+ raised; September 9 research recorded 21. https://www.metadao.fi/founders
9. Vitalik Buterin. The promise and challenges of crypto + AI applications. January 30, 2024. https://vitalik.eth.limo/general/2024/01/30/cryptoai.html
10. Vitalik Buterin. From prediction markets to info finance. November 9, 2024. https://vitalik.eth.limo/general/2024/11/09/infofinance.html