Architect Launches Liquid Inference Auction for AI Model Requests

Architect Financial Technologies launched Liquid Inference on Oct. 7, putting large language model serving requests through a live auction that compares provider bids for each job. The platform routes each request to the lowest-priced offer that meets the buyer’s requirements, applying competition to the serving stage after a model has been deployed.

Buyers can set a maximum spend, specify model capabilities, and set response-time and throughput thresholds. They can also restrict approved regions, data retention, and the providers or models allowed to handle requests. A lower-priced offer does not qualify if it fails one of those conditions.

Providers can post quotes for a set period, then change or withdraw them. Before generation begins, the platform sets a maximum price for the request and charges for metered usage. The receipt identifies the selected provider, the request’s spending ceiling, the final charge and how competitive the award was.

The service works with clients compatible with OpenAI and Anthropic. Inference is when a deployed model processes a request and generates a response; Liquid Inference handles that work, not model training.

Brett Harrison, Architect’s founder and CEO, described inference as a growing technology expense, saying, “Inference is quickly becoming one of the largest line items in every technology budget.” The launch included no like-for-like comparison showing whether buyers might save money against other providers.

Architect’s test phase covered hundreds of tasks across more than 700 models. Architect uses randomized test requests and response fingerprints to check whether providers serve the models they claim to offer. Architect has not published the full procedure.

Marketplace data may help Architect develop a reference index for inference-token prices. The company has not settled how to construct the index.

Liquid Inference’s terms define buyers as Architect customers and independent inference providers as subcontractors, with no direct contract between providers and buyers. Architect classifies the service as neither a financial, investment nor digital-asset product. Its separate plans for a U.S. exchange offering GPU-compute futures and options remain under regulatory review.