The promise sounds almost too straightforward: let artificial intelligence models fight for your business in real time, pay only for what you use, and watch the savings pile up. No long-term contracts, no manual negotiations. Just a live orderbook where sellers undercutting each other.
That's the wager Surplus Intelligence is making. The Wyoming-based startup deployed its settlement contract on Base blockchain between February and March 2026, building what it calls "a decentralized exchange for LLM inference." Think of it less like ChatGPT's subscription model and more like a stock exchange—where every API call triggers an automated auction, and the cheapest healthy model wins the request.
As of mid-June 2026, Surplus reported daily usage trending toward over 2 billion input tokens per day excluding cache. Analytics snapshots from around that period show over 1 million requests and 38.25 billion input tokens handled across a 28-day window. Those are marketing claims, not independently verified figures. But if even half-true, the platform has quietly reached a scale that matters.
An Orderbook for Inference
Here's how it works. Model sellers—Venice AI, Bankr LLM Gateway, OpenRouter, and others—list their prices on Surplus. The platform's router checks the orderbook, stored in memory and refreshed from DynamoDB roughly every 30 seconds, then tries the cheapest option first. If that seller fails or times out, the system immediately cascades to the next cheapest without the buyer noticing a hiccup.
Every API call settles on-chain in USDC. The platform calculates cost by multiplying tokens against the seller's listed price, then calls a settlement contract on Base to move funds—about $0.001 to $0.002 in gas per transaction. The current fee structure sits at zero percent, with a multiplier of 10,000. Whether that changes as the platform matures remains an open question.
Developers already comfortable with OpenAI's API can swap in Surplus with a single line change—just point the base URL to api.surplusintelligence.ai. The homepage touts a 38% cut to your inference bill, though your mileage will vary depending on which models you're calling and when. The platform also supports an Anthropic-compatible wrapper for Claude Code, plus media generation endpoints for images, video, and audio.
The Discount Math
Recent data paints a rosier picture than most startups might admit to. According to company-reported analytics as of July 4, 2026, the platform logged 38.25 billion input tokens, 1.21 billion output tokens, and 30.19 billion cache tokens over a 28-day stretch. Seven-day discount tracking showed a mean of 57.7%, with the median bucket landing between 40% and 45%.
Surplus translates that into roughly $15,000 in weekly savings for users—close to $800,000 annually if the pace holds. Take those numbers with the usual grain of salt. They're self-reported, and the math assumes consistent usage patterns and baseline pricing that may not reflect what every developer would pay going through official channels.
Still, the underlying logic is hard to argue with: competition drives prices down. And in a market where OpenAI, Anthropic, and Google each charge different rates for similar capabilities, arbitrage opportunities exist. Platforms like CheapTokens and LLM Mart already trade discounted inference credits in secondary markets. Surplus is betting that an orderbook—constantly updating, transparently priced—can do that job more efficiently.
Settlement Rails and Payment Quirks
The settlement contract went live on Base sometime in early 2026, according to the platform's changelog. Public social presence didn't begin until May. By mid-June, Surplus had migrated to AWS infrastructure, and recent snapshots show 76 active marketplace models pulled from a catalog of 145 across allow-listed providers.
Payment options span more than just standard USDC wallets. The platform integrates x402 micropayments—an account-less HTTP 402 standard built on Permit2—and MPP/Tempo session-based micropayments. The documentation mentions fiat on-ramps for funding a Base USDC wallet for normal settlement flows, though specific details about how they operate or which providers are supported remain sparse.
The x402 integration ties into Bankr's LLM Gateway, a Base-native infrastructure provider positioning itself for agentic payments. Surplus also surfaces endpoints on x402scan, priced per request in USDC. It's a constellation of crypto-native payment rails that, for better or worse, assumes a certain comfort level with on-chain mechanics.
Tooling and Developer Integrations
The platform has rolled out integrations for Cursor, Aider, OpenCode, Continue, and Claude Code—tools that developers increasingly rely on for AI-assisted coding workflows. A minimum-discount routing feature lets buyers set thresholds, ensuring requests only route if projected savings clear a certain floor. That's a hedge against situations where the cheapest option might not be worth the hassle if savings are marginal.
Venice AI and Bankr LLM Gateway provide much of the active liquidity. Both are Base-enabled, both position themselves as infrastructure plays rather than consumer-facing products. The broader ecosystem around discounted inference—secondary credit markets, arbitrage platforms—has been growing quietly for the past year. Surplus is threading itself into that landscape, offering an orderbook-driven alternative to manually sourced credits or opaque reseller deals.
The Trust Questions
The documentation is refreshingly candid about what Surplus doesn't solve yet. Seller provider keys are encrypted using AES-256-GCM via AWS KMS envelope encryption, but the platform offers no cryptographic proof of exact model identity. You're trusting that when you call GPT-4, you're actually hitting GPT-4 and not some cheaper substitute. Routing defaults to trusted domains, which is another way of saying: we're checking reputation, not cryptographic proofs.
The settlement contract—a UUPS proxy—currently lacks enforced timelocks on admin upgrades or fee changes. That means the team could, in theory, push through changes without warning. For a platform branding itself as decentralized, that's a meaningful caveat.
No team members are identified publicly. Surplus Intelligence LLC is listed at 1309 Coffeen Avenue STE 19992, Sheridan, Wyoming—a jurisdiction favored by crypto startups for its light regulatory touch. A SURPLUS token exists on Base, with listings on LBank and KCEX alongside DEX pools, though no formal venture round or valuation has surfaced. The lack of public faces and conventional funding might be a red flag for some, or simply a reflection of how early-stage crypto infrastructure often operates: pseudonymously, with tokens instead of term sheets.
The Bet on Competition
Perhaps the most interesting question isn't whether Surplus can drive costs down—it appears to be doing that already—but whether developers will trust an orderbook run by an anonymous team with encryption but no cryptographic model verification. For those prioritizing cost efficiency over maximum decentralization, the trade-off might be worth it. The settlement is verifiable on-chain. The savings, at least according to company data, are real enough.
And the underlying thesis—that competition among sellers will keep driving prices down—has a certain elegance to it. As long as model providers keep undercutting each other for volume, and as long as the router keeps picking the cheapest option, buyers win. The platform just needs to stay fast, stay honest, and avoid becoming the single point of failure in a system that bills itself as decentralized.
For now, Surplus remains a live experiment in what happens when you apply exchange mechanics to inference. Whether that experiment scales beyond the crypto-native developer crowd will depend on trust, transparency, and whether the savings hold up as more users pile in. The orderbook is live. The bids are real. What happens next is anybody's guess.
