Verifiable spending on web3 AI model calls reached about $8.7 million in the 30 days through Oct. 12, with privacy-focused AI provider Venice accounting for roughly $7.2 million, or 83% of the total, according to a review of on-chain receipts, settlement records and disclosed supplier data covering 130 projects.
The findings place paid demand far below the activity implied by the trillion-token usage figures frequently promoted by decentralized AI networks. More than 100 projects outside Venice produced a combined estimated $1.5 million in monthly receipts, even as several reported processing hundreds of billions or trillions of tokens.
The review, conducted between Oct. 13 and Oct. 15, excluded self-reported token counts unless they could be connected to payments. Its $8.7 million estimate combined directly visible blockchain flows, OpenRouter supplier figures and project-published settlement documents, while removing duplicate counts where one project resold another provider’s capacity.
Venice’s dominance came from a mechanism introduced on July 17: the company buys back and burns $5 worth of VVV tokens for every $100 in API credits sold. The independent venicestats dashboard recorded 71,958 credit-burn transactions worth about $359,800 during the period. Applying Venice’s stated formula implies approximately $7.2 million in credit sales, equivalent to an annualized pace near $87 million if demand remained steady.
a steep drop after Venice
The next-largest projects were much smaller. Phala generated an estimated $520,000 through OpenRouter sales of trusted execution environment, or TEE, inference. Bittensor subnet SayGM recorded about $426,000, while Surplus Intelligence took in roughly $128,000, engy about $112,000 and Prime Intellect about $86,000.
TEE inference runs models inside protected hardware enclaves, allowing users to verify that processing took place in a secured environment. Phala’s position near the top of the non-Venice group indicates that buyers will pay for a differentiated infrastructure feature when it is embedded in a large routing marketplace.
The gap between Venice and the rest also reflects where payment capture occurs. Venice sells API credits directly, giving it a clear connection between customer spending and its token burn. Many decentralized networks sit further down the stack, selling capacity through intermediaries, subsidizing prices with token emissions or distributing free quotas that weaken the link between reported use and customer payments.
Seven decentralized suppliers — Phala, AkashML, Darkbloom, io.net, Prime Intellect, NEAR AI and Chutes — sold around 4.34 trillion tokens through OpenRouter in 30 days, producing about $760,000 in end-user payments. That equates to a small fraction of the token activity cited by the networks themselves.
OpenRouter’s supplier data also showed customer concentration. AkashML attributed 98% of its August use to OpenRouter, while Darkbloom’s share was about 95%. Such dependence gives the routing platform substantial influence over demand, pricing and the ability of smaller compute networks to retain customers directly.
routing platforms operate at a different scale
Stripe’s Aug. 19 acquisition of model-routing platform OpenRouter, reportedly valued at more than $7 billion, underscored the scale of the centralized model-access business. OpenRouter processed 656 trillion tokens over 30 days, according to the supplied data.
Using a 5% to 5.5% fee assumption and revenue estimates from research firm Sacra, the review estimated OpenRouter’s monthly payment volume at roughly $240 million to $270 million. The mapped web3 AI projects outside that platform therefore represented only about 3% to 4% of OpenRouter’s inferred volume.
Pricing data showed why raw token totals are a poor guide to commercial traction. SayGM received about $2.30 for every million reported tokens, compared with an estimated $0.39 platform-wide level for OpenRouter. Gonka and Twigpine were below $0.003 per million tokens.
Heavy caching, free allocations and discounts explain part of that spread. Surplus Intelligence processed 4.17 trillion tokens in 30 days but paid sellers about $128,000 on-chain. The project said 79% of its inputs came from cache, nearly half of token usage involved DeepSeek V4.1 Flash, and pricing was near one-tenth of listed rates. Buyer deposits were also concentrated: 96% came from a Bridge wallet associated with Stripe’s fiat-to-USDC flow.
0G Compute’s visible blockchain charges equaled about 1.6% of the amount its list pricing would have suggested. Prime Intellect reported nearly 1 trillion tokens per day, while Dolphin reported 332.2 billion over 30 days, but neither figure alone establishes comparable external spending.
x402 use remains concentrated outside AI calls
The x402 payment standard has drawn support from organizations including Stripe, Visa and Google, and is designed to let software agents make internet payments automatically. Its use in paid AI inference remained limited in the measured period.
Token-metered inference gateways on x402scan collected about $3,800 over 30 days. BlockRun led with roughly $3,186 across 349,000 transactions, suggesting an average charge below one cent per call. Venice’s implied $7.2 million of credit sales included only about $3,100 in x402 payments, less than 0.05% of its total.
Coinbase Bazaar listed 32,498 x402 resources, with large-model resources representing 3%. Of roughly 940,000 calls over 30 days, large-model calls made up 9%, and BlockRun accounted for 88% of that category.
Crypto payments have also become less prominent at OpenRouter. On-chain address estimates put crypto’s share of OpenRouter payments at about 5% to 6% in mid-2025, falling to 1.3% to 2.2% in the first quarter of 2026. OpenRouter disabled programmatic crypto top-ups on April 2 after Coinbase Commerce shut down its Charge API. Based AI likewise ended x402 and wallet top-ups in favor of card payments; its x402 receiving address had received about $1,122 since March.
quota tokens face an activation problem
Credit-denominated quota tokens expanded during the autumn through staking rewards, fee-sharing systems and dollar-priced usage credits. Their issuance often exceeded actual consumption.
Venice’s DIEM provided the largest visible example. Each staked token created $1 of daily quota, producing a 30-day quota ceiling near $834,000. Actual usage reached about $371,000, or 44% of available quota, alongside Venice’s much larger cash-credit business.
Orbio, launched on Robinhood Chain through Pons on Aug. 31, charged a 1.5% trading fee and converted half into a model-use token called CREDIT. DefiLlama data showed about $342,000 in fees from Sept. 16 through Oct. 12, creating roughly $170,000 in quota. Only about 34,300 CREDIT was activated over the latest four weeks, including 15,000 activated by an operations wallet for card users and about 19,000 from around 480 wallets.
Secondary-market trading was dominated by arbitrage contracts. They represented 80% of buyer-side CREDIT volume, and none of the tokens acquired by the five largest buyers was activated for model usage. Weekly average prices dropped from 0.82 USDG to 0.42 as daily trading fees declined from about $37,000 on Oct. 1 to $7,652 on Oct. 12.
The data draws a sharp distinction between token issuance, model traffic and money paid for access. Projects that directly control credit sales, routing or specialized supply recorded the clearest receipts, while many token-incentivized networks showed large computational volumes with limited independently verifiable customer spending.
To see how Web3 and AI reshape markets beyond models, explore Web3 AI and crypto’s broader impact on trading and infrastructure.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
