The Markets That Let AI Agents Pay for Themselves
Three onchain inference marketplaces—Orbio, Surplus Intelligence, and Antseed—are turning model access into a tradeable commodity. The economics are still unproven.
By The crypto-money Desk
Somewhere on Base, a software agent is routing its own inference request to the cheapest available seller, paying fractions of a cent per call, and logging the transaction onchain. It did not ask a human for budget approval. It drew on a token balance it acquired through a marketplace it found autonomously. This is not a whitepaper scenario. Surplus Intelligence has cited more than 120 million requests over 28 days, many costing fractions of a cent [1].
What crypto infrastructure gives AI is not just cheaper access to models. It gives access a price, and prices can be owned, traded, earned, and burned. That structural shift is small enough to miss in a single product announcement, but add several products together and the outline of a new market layer starts to come clear.
Three marketplaces have emerged as the clearest examples of this layer forming: Orbio on Robinhood Chain, and Surplus Intelligence and Antseed on Base [1]. All three offer discounted inference. All three use tokens. Beyond that, their approaches diverge sharply—in architecture, in customer focus, and in how honestly they can connect marketplace activity to token value.
Orbio began as something simple: a token that distributed OpenRouter credits to holders. It has since grown into a platform offering access to hundreds of AI models, many at steep discounts, alongside a launchpad for developers to deploy agents and their own tokens that trade in a pair with ORBIO [1]. The product is now explicitly shaped around one kind of customer—the onchain AI agent.
The mechanism at the center of Orbio is a two-token design. ORBIO is the platform's primary token [1]. CREDIT represents one dollar of inference through Orbio's gateway [1]. Sell that CREDIT on the onchain order book, or burn it to activate API balance across 400-plus models and tools [2]. The arithmetic can work in a buyer's favor: CREDIT recently traded at about $0.47, meaning an agent could purchase 100 CREDIT for roughly $47 and activate $95 of inference balance after fees [1].
For an autonomous agent, this matters in a specific way. Agents can pay for their own prompts, using funds from their wallet to buy and then burn CREDIT in the pursuit of their own 'longevity' [1]. Agent-token launches that trade against ORBIO can separately direct creator fees toward funding their own inference [1]. The plumbing is elaborate, and deliberately so—Orbio is trying to build an economy where agents are self-sustaining rather than dependent on human wallet top-ups.
The question is whether this can sustain itself without continued speculation [1]. Orbio has reported around 2.9 million inference requests since September, but only about 11% of CREDIT minted has been activated, leaving most issued credits unused so far [1]. The credits exist. The agents are not yet burning them at scale. Whether that gap closes through genuine demand or stays open is the central test of the model.
A marketplace that works on paper can still stall when the incentives settle into an unexpected equilibrium. Tokens accumulate as yield. Yield sits idle. The activity that was supposed to flow through the system pools instead at the point of lowest friction. Orbio's data captures that dynamic exactly—and the team knows it.
Surplus Intelligence takes a structurally simpler position. It operates an order book where sellers compete to resell access to AI models—OpenAI, Anthropic, Venice, OpenRouter—at prices they set, with competition driving discounts that can approach 90% off standard provider prices [1]. When a request comes in, Surplus's gateway routes it to the cheapest available seller, switching automatically if one fails [1]. Buyers pay by card or USDC, including through x402, which makes the marketplace accessible to autonomous agents [1].
That simplicity has produced the most reported usage of the three platforms. The 120-million-request figure over 28 days is significant even accounting for the low per-call cost [1]. Surplus has also added a feature called Secret Masking: when enabled, it automatically replaces sensitive information like API keys and private keys with placeholders before forwarding prompts to providers, then restores that information in the response [1]. For agents handling credentials, the feature reduces one real category of exposure—though Surplus still processes the original prompt [1].
The gap in the Surplus story is the token. SURPLUS plays no role in inference payments or routing, nor does it have a documented mechanism connecting marketplace revenue to token holders [1]. The team proposed a token migration on October 16 to give it greater control over the token's development, but nothing is set in stone, leaving Surplus with a functioning and increasingly active marketplace, though no established way for that usage to translate directly into token value [1].
Usage without tokenomic integration is its own kind of stress test. A product can succeed on the metrics that matter to users—volume, price, reliability—while leaving investors holding something whose value depends entirely on a future mechanism that has not yet been built. Both can be true at once. The question is which truth matters more to the people in the room.
Antseed is describing itself as 'BitTorrent for AI inference' [1]—a peer-to-peer network where buyers connect directly with independent providers, rather than routing through a centralized gateway. Buyers discover available providers, compare their prices and reputations, and choose who to use [1]. Providers can offer anything from cheap access to widely used models to more specialized services like model-fingerprinting checks, designed to detect whether a seller is actually serving the model it advertises [1].
The ANTS token, which is not yet publicly tradable, is designed to support the marketplace through incentives and reputation [1]. Buyers and sellers can earn newly minted ANTS through recognized paid activity, while staking tokens can improve a provider's reputation and reward eligibility [1]. Unlike Orbio, these rewards are not financed by trading fees—Antseed distributes newly minted tokens, with the possibility of eventually using marketplace fees to buy back and burn ANTS [1]. No such mechanism has been implemented yet, so the direct relationship between growing inference demand and token value remains unproven [1].
The architecture Antseed is proposing does address a real problem. A centralized gateway is a single point of failure and a single pricing authority. A peer-to-peer marketplace with provider reputation scores and model-fingerprinting distributes both the risk and the decision-making to buyers. Whether that complexity attracts users or repels them depends on how much the average inference buyer actually cares about provider selection—a question the market has not yet answered.
Each of the three architectures solves for something different. One optimizes for agent autonomy. One optimizes for price and volume. One optimizes for openness and verification. They are not competing on the same axis, which means their eventual success or failure will tell different stories about what buyers in this market actually want.
What connects all three is a structural bet: that tokenizing inference—making compute access something you can earn, hold, trade, and burn—creates utility that raw API access does not. CREDIT is inference as an asset [2]. ANTS is reputation made liquid. SURPLUS is, so far, a marketplace searching for its token rationale. The token thesis is clearest at Orbio, most absent at Surplus, and most nascent at Antseed.
None of these platforms have yet demonstrated that inference demand alone—stripped of token speculation—can sustain the economics they are promising. Orbio's 11% CREDIT activation rate is an honest data point [1]. Surplus's 120 million requests are real volume, but the token sits disconnected from it [1]. Antseed's burn mechanism is a proposal, not a function [1]. The infrastructure is being built in advance of the demand that would validate it.
That is not necessarily a problem. Most infrastructure gets built before the use cases that justify it arrive. The onchain inference market is still early enough that the three platforms profiled here are less competitors than they are three different hypotheses about what the layer should look like. One prioritizes agent autonomy, one prioritizes price, one prioritizes open provider selection [1]. When the AI agent economy scales to the point where inference costs actually matter at volume, the hypothesis that got the architecture right will have a significant head start. The rest will be a cautionary case study in tokens that were never quite connected to the thing they were supposed to represent.