1. A Marketplace Where Providers Bid for Each Request
Architect Financial Technologies, led by chief executive Brett Harrison, launched Liquid Inference around October 7, 2026, an AI inference auction in which model providers compete in real time to serve each customer request. Harrison argues that pricing inference the way a trader prices a position, from the marginal cost of serving one more request, can push the cost of running AI models down for buyers.
The product reframes how access to artificial intelligence is usually sold. Most developers today pick a provider, accept its posted rate for each model, and pay that rate regardless of how busy or idle the provider's hardware happens to be. Liquid Inference replaces the fixed menu with a live bidding process, so the price a buyer pays reflects what suppliers are actually willing to accept at that moment.
Harrison's argument rests on a simple claim about markets: when sellers have to compete against one another for each order, prices move toward the real cost of production. The company says its design brings that logic, long familiar from stock and futures exchanges, to the business of running trained AI models.
2. What Is AI Inference and Why Does It Cost So Much?
Inference is the stage at which a trained AI model produces an answer to a prompt. Training a model is a large, one-off effort; inference happens every time a person or an application sends it a question, and it consumes computing power on each occasion. As AI tools spread into search, coding, customer service and office software, inference has become the recurring bill that companies pay to keep those tools running.
That bill depends on specialized processors, usually graphics processing units, which are expensive to buy, power and house. A provider that owns a large fleet of them has to recover the cost whether the machines are busy or not. The result is a market in which the same model can be offered at different rates by different operators, and in which the posted rate says little about how much spare capacity any one operator has at a given hour.
Because demand fluctuates through the day and across regions, a good deal of hardware sits partly idle at various moments. An auction is one way to put that idle capacity to work, since a provider with unused machines has an incentive to quote a lower price than one whose machines are fully booked.
3. How Does the Liquid Inference Auction Work?
Under the model Architect describes, providers bid in real time for model-serving jobs, and the buyer chooses the offer that best fits the job. Each request effectively becomes a small trade: the buyer states what it needs, providers respond with quotes, and the work goes to the supplier whose terms match.
Price is only one of the factors a buyer can weigh. The platform lets customers consider cost, latency and the way their data is handled, not just the cheapest quote. A team building a voice assistant, where every fraction of a second matters, may therefore pick a different provider from a team running a large batch of document summaries that can wait. Both can use the same marketplace and still get what they value most.
This is where the comparison with a financial exchange is most direct. On an exchange, a visible order book lets buyers and sellers see each other's prices, and the trade clears where interests meet. Architect, which builds trading infrastructure for markets tied to computing resources, is applying the same discipline to a service that has so far been sold mainly through price lists.
4. Why Marginal Cost Is the Central Idea
The concept at the center of Harrison's pitch is marginal cost, meaning the extra expense of serving one additional request once the infrastructure is already in place. For a provider whose machines are running anyway, that extra expense can be small, because the hardware is already bought and the building already powered.
Harrison's illustration involves a provider with idle graphics processors at three in the morning. Because the added cost of one more job is low at that hour, the provider can afford to quote a low price and still come out ahead compared with leaving the machines unused. Under a fixed price list, that flexibility is rarely passed on to customers; in an auction, competition pushes providers to reveal it.
Economists describe this process as price discovery, the way open competition brings out what a good or service is truly worth at a given moment. Harrison contends that competitive bidding improves price discovery for inference in the way that traditional exchanges have done for other assets, and that this can lead to a meaningful reduction in what buyers pay.
It is worth noting that the claim concerns the potential of the mechanism rather than a measured saving. The argument is that the structure of the market gives lower prices a route to buyers, not that any particular discount has been demonstrated across the industry.
5. Who Could Benefit From Auction-Based Pricing?
Developers and companies that buy inference in volume are the obvious beneficiaries, because even small differences in the price per request add up quickly at scale. A business that runs millions of model calls a day has a strong reason to shop around, and a marketplace that does the shopping automatically lowers the effort involved.
Smaller providers may also gain. An independent operator with a modest fleet of processors can find it difficult to win customers against the biggest cloud platforms, whose brands and sales teams dominate procurement decisions. In an open auction, a lesser-known supplier can compete on the terms of the job itself, with price and performance, rather than on name recognition.
For buyers, there is a practical consideration as well: different jobs have different needs. Some prompts handle sensitive information and call for tight controls over where data goes, while others are routine and can be sent wherever it is cheapest. A system that lets the buyer express those preferences in the request, rather than committing to a single vendor in advance, matches the way many organizations already think about risk and cost.
6. Where Does Liquid Inference Fit in Architect's Business?
Architect did not start with inference. The company already offers financial products built around computing power, including compute forwards and options, which allow buyers and sellers to hedge against swings in the price of compute. A forward fixes a price for delivery at a later date, and an option gives the holder the right, but not the obligation, to transact at a set price.
Liquid Inference adds something different to that lineup: a spot-style marketplace, where the work is bought and delivered immediately instead of being contracted for the future. Taken together, the products cover both ends of the timeline, with the longer-dated instruments helping firms plan around the cost of computing and the spot auction setting the price of the work being done right now.
In practice, that means a company could lock in a rough expectation of its future computing costs through a forward while still sending individual requests to the auction each day. The two layers address different risks: one concerns the cost of capacity over months, the other the price of a single answer at a single moment.
The combination reflects a wider idea that computing power is becoming a commodity in its own right. If processing capacity is bought and sold in volume, traded for delivery on different dates and priced continuously, then the tools familiar from energy and metals markets begin to apply to it. Harrison, who built his reputation in trading technology, is betting that the same tools will prove useful for AI.
7. What to Watch as the Platform Develops
The first test is whether a real auction attracts enough supply and demand to produce prices that beat the standard rate cards. A marketplace is only as useful as the number of participants on each side, and a thin market can produce erratic quotes. Sustained participation from providers and from buyers will determine whether the mechanism works as Harrison describes.
A second question is whether buyers will accept the trade-offs that come with variable pricing. Fixed rates are easy to budget for, while auction prices can move. Companies that value predictability may prefer to use the forwards and options Architect already sells to manage that uncertainty, which is one reason the firm presents its products as a connected set.
For now, the launch is an early experiment in treating AI inference as something to be traded, and its results will show whether competition among providers brings the lower costs that Harrison expects.
8. How Does This Differ From Conventional Cloud Pricing?
Conventional cloud services publish a price list and hold it steady until the provider decides to change it. Customers can negotiate discounts for large commitments, but day-to-day the rate does not respond to how busy the provider is. That arrangement is simple and predictable, yet it means the supplier, not the market, sets the price.
An auction reverses that relationship. The buyer defines the job, the providers compete, and the price emerges from their bids. Compared with the posted-price model, the buyer gains the chance to pay less when supply is plentiful, and accepts that the price may be higher when capacity is scarce.
The model also changes what a provider needs to be good at. Under fixed pricing, success depends largely on sales channels and brand. Under an auction, it depends on running hardware efficiently, since the operator that can serve a request at the lowest real cost can bid most aggressively and still earn a margin. That pressure, if it materializes, tends to reward operational skill over marketing.