Liquid Inference, announced by Architect Financial Technologies on October 7, 2026, is an online marketplace where providers quote prices to serve large-language-model requests, and each request goes to the lowest-priced offer that meets the buyer’s rules. Architect describes this as exchange-style price discovery for model output. The company’s claims about how the auction works, what it costs, and how well it performs are its own. This article explains the mechanics as Architect describes them and marks where independent evidence does not yet exist.
How a request moves through the auction
Architect’s launch announcement and product introduction describe the same sequence for every job. The steps below follow that description.
As an Amazon Associate I earn from qualifying purchases.
- You name a model and set your rules. Each request can carry its own constraints, covered in the next section. Auto-routing can instead choose a model for a particular unit of work.
- A maximum price is locked before the first token. Architect says the price cap is fixed before generation begins, so the ceiling is known at the start of the job.
- Eligible providers compete. The request is auctioned among providers quoting the named model. Offers that cannot satisfy your rules are excluded from the award.
- The lowest qualifying offer wins. Architect says the award goes to the cheapest offer that passes every configured rule.
- You pay only for metered usage. The final charge reflects metered use rather than the cap itself.
- A fixed receipt is issued. It records the winning provider, the price cap, the final charge, and the competition depth at award time, meaning how many providers were in the auction when it was decided.
The receipt’s competition figure is the number to watch. The launch named seven providers as its first cohort, so the number of providers quoting any given model may be small at first, and a thin field gives the auction less to work with.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →The controls buyers can set
Architect lists the following per-request controls. Each one narrows which offers can win.
#1 Best Overall
- NVIDIA Volta GV100 Architecture — 4,608 CUDA Cores, 640 1st-Gen Tensor Cores delivering 14 TFLOPS FP32 and 112 TFLOPS deep learning performance for AI training, inference, HPC, and scientific computing workloads
- 32GB HBM2 ECC Memory — 900 GB/s Bandwidth — High-bandwidth memory on a 4096-bit bus with ECC error correction provides the memory capacity and throughput required for the largest AI models, simulations, and datasets
- PCIe 3.0 x16 Interface — 250W TDP — Standard PCIe Gen3 connectivity with passive cooling designed for enterprise rack server deployment in HPE ProLiant, Dell PowerEdge, and Supermicro platforms with adequate chassis airflow
- NVLink — Scale to 96GB Unified Memory — Connect two V100 GPUs via NVLink at 300 GB/s bi-directional bandwidth to scale GPU memory from 32GB to 96GB for larger AI training and HPC workloads
- Multi-Precision Computing — Supports FP64 (7 TFLOPS), FP32 (14 TFLOPS), FP16 (112 TFLOPS) and INT8 precision modes for flexible deployment across training, inference, and scientific simulation workloads
- Per-job cost cap: the maximum price you will accept for that job.
- Maximum time to first token: a latency ceiling for when the first output token arrives.
- Minimum throughput: a floor on output speed for the job.
- Approved regions: the geographic locations where a provider may serve the request.
- Zero data retention: a requirement that the request be handled under zero data retention terms.
- Provider and model allow lists: limits on which providers and models may bid.
These controls are described in Architect’s announcement. Whether a given provider can meet a given latency, throughput, or retention setting in practice has not been independently verified.
Compatibility and Auto-routing
Architect positions Liquid Inference as both a chat-style interface and an API service. It says the API works with OpenAI-compatible and Anthropic-compatible clients, including official SDKs and coding tools such as Claude Code, Codex, Cursor, and Cline, and that existing users can switch without changing their code. Treat the no-code-change claim as Architect’s statement. Independent integration testing of these tools is not established.
Auto-routing is the feature that selects a model for a unit of work. It is useful when you want the auction to decide between models as well as providers, but it also means the model that runs a job may differ from the one you would have named yourself. Keep the allow lists in place if that matters for your workload.
Rank #2
- Delivers 25 TFLOPS of AI inference performance using four MemryX MX3 accelerators in a compact M.2 2280 M-key module.
- Connects through PCI Express Gen 3 and above for straightforward integration into embedded and edge systems with an M.2 M-key slot.
- Includes 42 MB of distributed on-chip weight storage to support low-latency, batch-1 pipelined inference.
- Designed for low-power deployment with a 1 W–12 W operating range and host-managed airflow requirements.
- Supports ONNX, TensorFlow, Keras, and TFLite models through the MemryX software stack on Linux and Windows hosts.
Who supplies the inference
Liquid Inference is a service provided by Architect Financial Technologies Inc. Architect says it sources inference from independent providers acting as subcontractors, and that those providers have no contract with buyers. Your agreement for the service is with Architect, not with the provider that serves a given request.
The first launch partners
Architect named akashML, Boundless, engy, Grizzly, Hyperbolic, Pearl, and Zro as the first cohort of inference-provider launch partners. According to the product introduction, providers can register models and quotes through a REST/WebSocket API, and provider payouts are processed through Stripe. Architect says providers pay no fees. That is a launch-era claim, and the terms may change.
Scale and what the figures mean
The product introduction says the platform offered hundreds of open- and closed-weight models. Architect also reports 700+ models and hundreds of tasks completed successfully during its test phase. Both figures come from Architect itself. They describe the company’s own launch scale, not an independent benchmark, and they say nothing about production performance, uptime, or typical prices.
Rank #3
- ✅Powered by 26 Tera-Operations Per Second (TOPS) Hailo-8 AI Processor. 2.5W typical power consumption
- ✅Scalable, enabling simultaneous processing of multi-streams & multi-models
- ✅Enabling real-time, low latency and high-efficiency AI inferencing on the edge devices
- ✅Supports TensorFlow, TensorFlow Lite, ONNX, Keras, Pytorch frameworks
- ✅Supports Linux and Windows. Supports the temperature range of -40°C to 85°C
Introductory credits and referral terms
Architect has published several time-sensitive offers. Confirm that they are still available before relying on them.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
| Offer | Terms as Architect published them |
|---|---|
| Launch credit | $20 in credits for the first 500 users (Architect, 2026) |
| Direct referral credit | 20% of Architect fees for referred users |
| Second-level referral credit | 10% for referrals of referrals |
| New-account balance | New accounts receive a starting balance in inference tokens, according to the launch announcement; the amount is not stated in the announcement |
Architect says credits are service prepayments that can be used only for its service. They are non-transferable and have no cash value. Because they are prepayments, a credit is not a refund right or a savings guarantee.
The market argument behind the auction
Architect argues that inference supply is often sold through bilateral enterprise contracts, and that suppliers on existing platforms may not always have an incentive to offer their best price in real time. This is the company’s market rationale. Independent market-wide evidence for it is not established in the sources reviewed.
Rank #4
- 48GB AI graphics accelerator
Brett Harrison, Founder and Chief Executive Officer of Architect, said in the announcement: “Inference is quickly becoming one of the largest line items in every technology budget, but it is still only available via static list prices and bilateral contracts. Liquid Inference brings the core tools of modern exchanges to every AI request: competitive quoting, transparent market data, and firm limit prices.” That is a statement of product intent. It is not a finding about realized savings.
What is not yet established
- Auction behavior, service performance, billing, and receipts: the descriptions above come from Architect. No independent testing of these has been published as of October 9, 2026.
- Cost, speed, and quality: the available sources do not show that Liquid Inference is faster, cheaper, or higher quality than other inference options, and they report no measured savings.
- Availability: Architect says the service is not offered in every jurisdiction. Check the terms for your location before signing up.
- Output accuracy: Architect cautions that AI outputs may be inaccurate, incomplete, or inappropriate, and that caution applies to this service as to any other.
Liquid Inference is not a financial product
Architect also publishes about regulated financial-market offerings. Its press release mentions the American Innovation Exchange and other compute-related products, and notes that intended products are subject to applicable law and regulatory approval. Those products are separate from Liquid Inference. The company’s article states that Liquid Inference itself is not a financial, investment, or digital-asset product.
How to evaluate it against other options
If you are comparing Liquid Inference with a fixed-price API or a bilateral contract, use the same checklist on each:
Best Value
- Professional AI & Creator Workstation: AMD Radeon AI PRO R9700 GPU with 32GB GDDR6 is engineered for AI development, professional content creation, and compute-intensive workloads.
- Massive 32GB Memory Capacity: 32GB of GDDR6 memory on a 256-bit bus provides ample bandwidth for large AI models, 8K video editing, and complex 3D rendering.
- Advanced RDNA 4 with AI Accelerators: 64 Compute Units with 3rd Gen Ray Tracing and dedicated 2nd Gen AI Accelerators for groundbreaking AI performance and visual computing.
- Professional Blower Cooling: Efficient single blower design exhausts heat directly out of the chassis, ideal for multi-GPU workstation and server configurations.
- Enterprise-Grade Thermal Solution: Vapor chamber heatsink with industrial Honeywell PTM7950 thermal interface material ensures reliable cooling under sustained professional loads.
- Total metered cost per job, and whether the price cap is set low enough to matter.
- Latency limits, meaning maximum time to first token, and minimum throughput.
- Region, zero data retention, and any other data-handling requirements.
- Which models and providers are eligible under your allow lists.
- How Auto-routing chooses models, and whether that choice is acceptable for your work.
- The competition depth shown on receipts over a sample of your own jobs.
Run these checks on your own workload. The published materials document the controls but do not support a conclusion about which option is better for your use case.
Sources
The Bottom Line
Liquid Inference is a real product offering with a clear mechanism: per-request rules, a locked price cap, and a receipt for every job. Whether the auction lowers your costs depends on how many providers quote the models you use, which only your own receipts can show.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
