TheBlockRadar
Explainer· Independently researched

AI Verifiable Execution in Crypto: Proof, Tokens, and Costs

Explore AI verifiable execution in crypto, proof of SQL, token economics, and the hardware costs shaping decentralized AI infrastructure.

AI Verifiable Execution in Crypto: Proof, Tokens, and Costs

AI needs a way to prove what its agents did

On Nov. 5, Amundi put a tokenized share class of its Core STOXX Europe 600 index fund onto public Ethereum. The token gives exposure to a fund unit, but Amundi has not published eligibility, fees, transfer rules, or redemption procedures for that class.

The risk sits in those missing terms. A fund unit on Ethereum does not automatically give you permissionless trading, retail access, or a claim redeemable through DeFi, and the blockchain record cannot replace the fund’s legal ownership and compliance systems.

This is the central issue in the AI-and-crypto overlap: not whether an AI agent can produce an answer, but whether you can verify the work, the authority behind it, and the assets it is allowed to move.

Space and Time Enterprise, launched by Microsoft-backed blockchain data project Space and Time, is built around that problem. It allows a company to run a calculation on private records and provide proof that the calculation was correctly performed without releasing the records themselves.

For a finance team, the example is straightforward. An AI agent might calculate collateral coverage, customer exposure, or whether a payment meets policy limits, while an auditor receives evidence of the query and result instead of unrestricted access to the entire database.

That is a narrower promise than “AI on blockchain.” It is also easier to evaluate because the product has a defined job: prove that a database query was executed correctly against a specified dataset.

Proof of SQL is a receipt for a database calculation

Space and Time calls its system Proof of SQL. The basic mechanics resemble a tamper-evident receipt: a user asks a database a question, the system computes the answer, and it generates cryptographic evidence that the answer came from the stated records and query.

The proof lets another party verify the computation without rerunning the entire database workload. That can matter when the verifier is an outside auditor, a lender, a trading counterparty, or an AI agent that needs a checked result before taking an action.

A conventional database asks you to trust the operator. If a company tells you its loan book has a certain average collateral ratio, you normally trust its systems, internal controls, accountants, and later audits.

A proving system changes the division of labor. The operator still holds the data, but the recipient can check whether the declared calculation followed the declared rules, without obtaining every customer balance or transaction record.

There is an important limit. Nate Holiday, chief executive of MakeInfinite Labs, which developed Space and Time, told Decrypt that proof can show correct processing of a dataset, but cannot establish that the original data entry was true.

If a loan officer entered the wrong collateral value, a perfect proof will verify a perfectly calculated wrong result. You still need data controls, reconciliations, legal records, and people accountable for the source information.

Space and Time says supported queries over millions of rows can be run and proven in under a second using GPU acceleration, although performance depends on the query, data set, and hardware. The system works with Google Cloud BigQuery and can run on a customer’s own servers.

That last point affects the token question. A business can use the enterprise product without operating a blockchain, buying specialized proving hardware, or necessarily creating immediate demand for SXT, Space and Time’s network token.

Holiday said network use enters when an application needs to connect with onchain markets, where workloads are paid through Space and Time. For you, that means enterprise adoption and token demand are related, but they are not the same event.

The token must pay for work

AI-and-crypto projects often combine three separate things: a software product, a token, and a claim that decentralization will lower costs. You should separate them before assigning value to any of them.

The strongest model is payment-linked usage. Render coordinates GPU providers and customers through its RNDR token, while Akash has used a burn-mint mechanism intended to link AKT supply to spending on compute services, Coin Bureau reported.

That structure gives you a measurable chain. A customer buys compute, the network processes the job, suppliers receive payment, and token activity follows demand for a service that someone actually used.

Bittensor takes a different approach. Its subnets compete to supply specialized AI services, and Coin Bureau reported estimated annualized customer revenue rising from about $35 million in September to more than $50 million in early October.

Those numbers are still small against centralized cloud infrastructure. CoreWeave, a centralized AI cloud provider, reported about $2.58 billion of revenue in one quarter, according to Coin Bureau, showing how far decentralized compute remains from hyperscaler scale.

The gap is not automatically a failure. Open networks may serve users who want geographically distributed supply, flexible capacity, or an alternative to a small number of cloud providers, but they face the same hardware, uptime, and power constraints.

We would look first for verifiable jobs and disclosed customer payments. We would skip projects whose token could disappear without changing how the product operates, because that makes the token a fundraising layer rather than infrastructure.

AI chips make the economics brutally physical

The AI trade has been framed as a software story, but its costs begin with chips, power lines, cooling equipment, land, and permits. Every AI model answer ultimately runs through expensive hardware consuming electricity around the clock.

Nvidia’s H100 SXM5 reportedly costs about $3,320 to manufacture and has sold around $27,000, while the B200 reportedly costs about $6,750 to manufacture and has sold near $40,000. Those reported spreads help explain why accelerator supply has become a strategic bottleneck.

Newer systems are becoming larger and more memory-intensive. Nvidia’s Vera Rubin NVL72 is specified with 288 GB of HBM4 memory and rated for 50 FP4 petaFLOPS of inference, while AMD’s MI455X is specified with 432 GB of HBM4.

Those specifications matter to decentralized GPU networks because they reveal the mismatch between a marketplace’s advertised supply and the equipment required for frontier workloads. A spare consumer GPU may be valuable, but it is not interchangeable with a dense data-center cluster.

Power is an even harder constraint. Decrypt reported that new gas turbines in the United States now take an average of six years to deliver, with procurement discussions already focused on 2030 and 2031 delivery windows.

AI developers have responded by seeking on-site generation, including aeroderivative turbines based on aircraft-engine technology. Crusoe ordered 13 50-megawatt units from ProEnergy in April, representing 650 megawatts intended for future U.S. data centers, Decrypt reported.

Used turbines can cost roughly $1.5 million to $1.7 million per megawatt, compared with about $2 million to $2.5 million for new units, according to equipment brokers and energy analysts cited by Decrypt. The lower sticker price still leaves fuel, emissions controls, permits, and maintenance.

The International Energy Agency expects global data-center electricity consumption to rise from 485 terawatt-hours in 2025 to 950 terawatt-hours in 2030. AI-focused facilities are projected to triple consumption over that period, according to Decrypt.

That is why we would be cautious with claims that crypto can cheaply democratize AI compute. The marketplace software may be open, but the best chips, reliable power, data-center operations, and grid connections remain expensive and heavily concentrated.

Algorithmic money still needs redemption rules

Dr. Pippa Malmgren, founder of the Geopolitical Institute and a former adviser to President George W. Bush, has argued on Bitcoin Magazine that AI’s ability to analyze vast numbers of scenarios will push finance toward faster, more algorithmic forms of money.

That is a view about direction, not an operating system that has already arrived. An AI agent can calculate which invoice to pay or which collateral threshold was breached, but it needs explicit permission, auditable instructions, and a settlement asset.

Stablecoins are the likely settlement layer in that model because they move continuously and can be programmed into onchain workflows. Yet the value of a stablecoin depends less on the AI agent than on the issuer’s reserves and redemption process.

Canada’s Nov. 5 budget proposal draws that line clearly. Federally regulated banks would be prohibited from issuing stablecoins directly, and would instead use separate non-financial subsidiaries operating within the country’s Stablecoin Act framework.

The framework would require redemption at face value in the tracked currency, reserves at least equal to outstanding token face value, and segregated custody intended to shield reserves from issuer and custodian creditors, CryptoSlate reported.

It also explicitly excludes stablecoins from public deposit insurance and government guarantees. If you hold a bank-affiliated stablecoin, you would not hold an insured bank deposit merely because a bank is connected to the issuer.

The Stablecoin Act received royal assent on March 26, but the broader framework is targeted for fall 2027 commencement. The market effect remains uncertain because regulations, issuer structures, and actual products have not yet been finalized.

Market rotation is a story, not yet a flow measurement

Coin Bureau has argued that AI employees receiving liquidity through private-company tender offers may direct some proceeds into Bitcoin, decentralized compute, and AI-linked crypto projects. It cited OpenAI employees selling roughly $6.6 billion of stock in one transaction.

The story is plausible, but the evidence does not establish a broad transfer of AI wealth into crypto. A tender offer shows that private shares became cash, not where that cash was invested after settlement.

Public insider sales have the same limitation. Coinbase chief executive Brian Armstrong and two insiders filed plans to sell up to $2.18 billion in Coinbase shares, with sales scheduled to begin in November, but a filing plan does not identify a buyer’s destination for proceeds. [2]

Markets also tend to compress separate narratives into one trade. Bitcoin rose from about $63,500 to around $78,000 over the period Coin Bureau compared with a 1.4% decline in the Nasdaq 100, but correlation and timing do not establish a capital rotation.

A more concrete crossover is corporate infrastructure. MARA reported $634.4 million in third-quarter net income on bitcoin gains, then sold 13,597 BTC to repay $750 million in loans as it shifted toward AI infrastructure, The Block reported.

MARA’s financing illustrates the cost of that pivot. Its August bitcoin-backed loans totaled $600 million, split between a $300 million floating-rate Coinbase Credit loan and a $300 million Two Prime Lending loan fixed at 7.65%, both maturing in August 2028.

The company had 34,012 BTC at Sept. 30, with 19,260 BTC loaned or pledged as collateral. AI infrastructure may offer a new revenue path, but bitcoin-backed debt means price moves can affect collateral, financing flexibility, and the timing of asset sales.

AI is changing crypto’s infrastructure case most clearly where it creates a need for verifiable data, continuous settlement, and scarce compute. It is not yet proving that every AI-branded token, tokenized asset, or corporate pivot will create durable cash flow.

Frequently Asked Questions

What is verifiable execution in AI and crypto?

Verifiable execution means proving that an AI agent performed a calculation correctly on authorized data without exposing the underlying information. It provides cryptographic evidence that a query was run against specified records and the result is accurate, allowing auditors or counterparties to verify the work without accessing all raw data.

How does proof of SQL work on blockchain?

Proof of SQL is a system that generates a tamper-evident receipt for database queries. When a user asks a question, the system computes the answer and produces cryptographic proof that the result came from the stated records and query. This proof enables others to verify the computation without rerunning the entire workload.

Why do AI crypto tokens need to pay for compute work?

Tokens have economic value only when they are used to settle actual compute, data, or verification work. Linking token use to payment for services creates a measurable chain where customers pay for compute, providers receive compensation, and token activity reflects real demand for infrastructure rather than speculative claims.

What are the hardware and power challenges for AI in crypto?

AI infrastructure faces significant costs from expensive chips, power consumption, cooling, and physical hardware. Supply constraints on AI accelerators and power shortages are causing delays and operational limits for data centers, making hardware and energy availability key bottlenecks before crypto adoption issues arise.

Can AI-generated claims be trusted without verification?

No. While proofs can verify that a calculation was done correctly on given data, they cannot confirm the truthfulness of the original data entry. Data controls, reconciliations, legal records, and human accountability remain necessary to ensure the accuracy of source information.

How we researched this

This article was assembled from 3 video sources across 3 channels, 6 published articles, 2 cited references.

Nothing here is based on hands-on testing. Where a figure or finding appears, it belongs to the source cited beside it, and the writing says so rather than implying otherwise. Every source is listed below so you can check it.

Sources

Watch AI and Crypto: New Developments and Market Effects on Youtube

Also from the sources

Altcoin and Layer 1 Project Developments in November 2026

Altcoin and Layer 1 Project Developments in November 2026

Explore key altcoin and Layer 1 project developments, including NEAR issuance cuts, Quant's bank deal, and Aave's DeFi lending growth.

XRP and Altcoin Market Developments and Challenges in 2026

XRP and Altcoin Market Developments and Challenges in 2026

Explore XRP and altcoin market developments, comparing tokenization, payments, and challenges in 2026’s crypto landscape.

AI Impact on Cryptocurrency Development and Market Dynamics

AI Impact on Cryptocurrency Development and Market Dynamics

Explore how AI affects cryptocurrency development, mining, and market risks amid calls for slower AI progress in crypto infrastructure.

Altcoin Market Dynamics and Key Project Comparisons

Altcoin Market Dynamics and Key Project Comparisons

Explore altcoin market dynamics with a detailed comparison of Ethereum, Hyperliquid, and other key projects in 2026.

Topics

Infrastructure