The bill for AI kept arriving in the wrong currency
Every developer using frontier AI models knows the ritual: five API keys, five dashboards, Stripe top-ups that run dry at the worst moment, and a different rate limit on each provider. Alan Orwick's team at Dominant Strategies watched that friction pile up and noticed something stranger underneath it. The people — and increasingly the agents — who most need inference are often the ones least equipped to pay for it in the traditional way. An AI agent cannot hold a credit card, pass a KYC check, or top up a balance before it knows whether a call is worth making.
Orwick came to the problem through blockchains, not APIs. As a computer science student at the University of Texas at Austin, he founded Texas Blockchain in 2017, then worked at Apple building supply-chain data pipelines before co-founding Quai Network and becoming CEO of Dominant Strategies, the firm that builds it. Quai began as university research in 2019 and grew into a proof-of-work Layer-1 blockchain with a deliberately unusual thesis: energy itself can be money. The network's dual-token design pairs QUAI, a store of value, with Qi, an energy-linked medium of exchange, and the project has reported raising around $15 million from investors including Polychain Capital.
That thesis sat mostly in the world of payments and monetary theory until the team asked a practical question: what if the same hardware already burning electricity to secure a chain could spend its work directly on something developers want? Inference was the obvious answer. The most expensive thing a developer buys today is tokens from someone else's model, and the cheapest thing many of them own is idle compute.
x402 Work: a payment rail made of puzzles
Hashference launched in late October 2026 alongside x402 Work, an open payment standard from the same team. The idea is simple enough to state in one sentence: any device pays per request with its own hardware. Send an API request with no API key and no account, and Hashference answers with a small puzzle over HTTP 402, the web's long-reserved payment-required status. The device solves it in a few seconds, sends the solution back, and the answer comes through. No signup, no card, no balance to manage.
For heavier use there are three more ways to pay. Leave an idle CPU, GPU or ASIC mining overnight through a Stratum connection and wake up with credits that work in Claude Code, Cursor, OpenCode and Cline. Buy credits outright with QUAI, Qi, USDT, USDC, ETH or BTC. Or stake QUAI through Hashference's plans in exchange for monthly platform credits. The team published x402 Work as an open standard, with hardware profiles for SHA-256d, KawPoW and Scrypt so that CPUs, GPUs and ASIC miners can all participate, and submitted its specification to the x402 Foundation.
The design also fits how agents actually operate. Hashference ships a Model Context Protocol server so compatible agents can discover models, check balances and make requests on their own — and, through x402, pay for individual calls with computing work rather than waiting for a human to fund an account. That is the point of the rail: a machine that earns its keep in compute should be able to spend compute, without a billing department in the middle.
One API, 45 models, no account
Behind the payment rail sits a deliberately boring product decision: one API that speaks the two dialects developers already use. Hashference is OpenAI- and Anthropic-compatible, so existing tools and code point at a new base URL — hashference.com/v1 — instead of learning a new SDK. The catalogue spans more than 45 live models from Anthropic, OpenAI, DeepSeek, Moonshot's Kimi, Qwen, Z.ai's GLM, xAI's Grok, Meta's Llama, Mistral and Google's Gemma, and the service is designed to drop into the coding assistants developers already live in, including Claude Code, Cursor, Codex, VS Code, OpenCode, Cline and Roo Code.
The onboarding is the pitch. The team advertises a ten-second, zero-key trial at hashference.com/zero-key: no email, no KYC, no card. That is a genuine break from how model APIs normally gate access, and it is only possible because the payment rail does not need to know who you are — only that the work was done. For a developer in a country where cards and verification are friction, or an agent running unattended, that difference is the whole product.
There are real limits to the model, and the team does not hide them. Proof-of-work per request costs seconds of local compute and real electricity, so it suits occasional calls and agents far better than high-volume production traffic; bulk users will still mine, buy or stake. But as a way to try a frontier model, or to let an agent pay its own way, it removes the two steps — signup and funding — that normally stand between a developer and a first token.
Why compute-as-payment matters for agents
Hashference is best understood as a bet on who the next big API customer is. If software agents become routine buyers of inference, the billing model built for humans — accounts, cards, monthly invoices — becomes the bottleneck. A standard that lets a device convert its own hardware into payment, request by request, gives agents a wallet they already own. Dominant Strategies is not the only team chasing agent micropayments, but it is one of the few grounding the attempt in a live blockchain economy rather than a slide deck: Quai's network, its miners and its dual-token system already exist, and Hashference is the first service running on top of the new rail.
Whether developers adopt it at scale will depend on the unglamorous details — model quality and latency, puzzle difficulty versus model cost, and how credits compare with plain token pricing. What is already clear is the shape of the idea: when the scarcest resource in AI is inference, and the most widely distributed resource is idle hardware, someone was going to build a bridge between them. Orwick's team built one where the toll is paid in work, and launched it with the models developers actually want on the other side.
The playbook
- Make payment invisible, not cheaper — Hashference's pitch is not a lower price per token — it is the removal of signup, cards and balances. The per-request puzzle turns payment into something that happens inside the request itself.
- Let the hardware be the wallet — A CPU, GPU or ASIC is something developers already own. Mining for credits and paying per request with work converts a sunk cost — idle machines — into spendable inference.
- Build the rail, then the first service — Dominant Strategies shipped x402 Work as an open standard and Hashference as its first application, so the payment idea is bigger than any single product — and other services can ride the same rail.