# Praecise > Praecise builds AI infrastructure from the machine up: the hardware AI runs on (devices, local computers, private clusters and dedicated GPU servers), Furcate, the open operating system it develops and stewards, and Praecise Cloud, the same stack operated as a service. Own the machines and run them, or let Praecise run them. ## Infrastructure (praecise.com) - [Praecise Metal](https://praecise.com/metal): Owned machines, placed, powered and cooled to measured targets. Website: https://metal.praecise.com. Console: https://console.metal.praecise.com - [Praecise Compute](https://praecise.com/compute): A local Linux machine for AI, agents, applications and services. Available; Furcate and its software support the silicon. - [Praecise Sense](https://praecise.com/sense): Devices that observe the physical world: motion, environment, sound and more. Available; Furcate and its software support the silicon. - [Praecise Edge](https://praecise.com/edge): Embedded compute for inference, control and real-time work. Available; Furcate and its software support the silicon. - [Praecise Interface](https://praecise.com/interface): Physical interfaces for people and agents: displays, controls and audio. Available; Furcate and its software support the silicon. - [Praecise Furcate](https://praecise.com/furcate): One open operating system for machines that work together, at every tier. Open source at https://furcate.org; any hardware maker can build on it, and it integrates most closely with Praecise hardware. Praecise Furcate (tuned images, model serving, fleet operations, support) at https://furcate.xyz. ## Praecise Cloud: the stack operated as a service (praecise.cloud) - [Praecise Cloud](https://praecise.cloud): Apps, agents, models and data, built from source and run on machines Praecise owns. Console: https://console.praecise.cloud - [Praecise Silo](https://praecise.com/silo): A microVM for every agent: its own kernel, its own disk, and no network. - [Praecise Inbox](https://praecise.com/inbox): An address and a searchable record for every agent and person. Website: https://inbox.praecise.com - [Ipnops](https://praecise.com/ipnops): Agents that run payments, accounting, commerce and mail inside limits you set. Website: https://ipnops.ai - [Skerry](https://praecise.com/skerry): Payment rails for agents, including stablecoin settlement. Website: https://skerry.xyz. Bridge and Stripe perform the regulated acts under their own licences. ## Research and open source (the engineering base, not products) - [Research notes](https://praecise.com/research): AI, data engineering, and the machines that run them. Measured on hardware, negative results included. - [Praecise Harness](https://github.com/Praecise/praecise-harness): A framework for AI agents. A folder is an app. TypeScript, Node 22+, npm package `@praecise/harness`. Serves chat UI, REST, MCP, A2A and AG-UI; any MCP server or OpenAPI document becomes tools; durable workflows with approval gates; SQL, document, time-series and vector stores. - [Praecise Engine](https://github.com/Praecise/praecise-engine): An inference acceleration layer for large language models. Rust with a C ABI. Speculation policy, n-gram self-speculation without a draft model, licence-checked drafter catalogue, request shaping, admission control. llama.cpp is the first backend; verified on NVIDIA Grace Blackwell. ## Blog: writing on AI infrastructure - [Blog index](https://praecise.com/blog) ยท [RSS](https://praecise.com/blog/rss.xml) - [A folder is an agent](https://praecise.com/blog/a-folder-is-an-agent): Most agents today live as prompts and settings inside someone's dashboard, where nobody can diff, review or reproduce them. Agents should be files: a folder you can read, version and put through review like any other code that acts on your behalf. - [An agent needs an address](https://praecise.com/blog/an-agent-needs-an-address): New standards for agent identity are still being drafted. Email already gives an agent a verified domain, authenticated sending and a durable record, and the largest mailbox providers now enforce the rules that make it trustworthy. - [An open operating system for machines you own](https://praecise.com/blog/an-open-operating-system-for-machines-you-own): Europe's Cyber Resilience Act now makes manufacturers answer for the security of every connected product. The only trust claims that survive that kind of scrutiny are ones anyone can read and check, rooted in keys the owner holds in hardware. - [CPUs are the agent bottleneck](https://praecise.com/blog/cpus-are-the-agent-bottleneck): An agent spends most of its time running tools, sandboxes and browsers, not generating tokens. The measurements and the chip makers now agree, and the CPU-to-GPU ratio in data centres is moving with them. - [Cloud concentration in the agent era](https://praecise.com/blog/cloud-concentration-in-the-agent-era): Three providers take close to two thirds of the world's cloud infrastructure spend, and regulators in the EU and UK are still working out what to do about it. Agents raise the stakes, because an agent's compute, mail, payments and record can all end up with one provider. - [Cross-border payouts at agent speed](https://praecise.com/blog/cross-border-payouts-at-agent-speed): The G20 will miss most of its 2027 targets for cheaper, faster cross-border payments, and remittance costs have barely moved. When software starts paying across borders, those costs stop being a policy disappointment and become an engineering problem that can be routed around. - [Edge cells: running AI where the machines are](https://praecise.com/blog/edge-cells-running-ai-where-the-machines-are): Three 2025 cloud outages and a national blackout showed what happens when everything depends on a distant link. Running compute in homes, businesses and districts as cells that govern themselves is a practical answer on resilience, energy, data locality and latency. - [Ipnops: agents that run operations](https://praecise.com/blog/ipnops-agents-that-run-operations): The business tools already do the work; what is missing is something that makes them work together. Ipnops connects payments, accounting, commerce and mail, and acts across them with permission, never ownership, inside limits the owner sets. - [MCP servers are the new APIs](https://praecise.com/blog/mcp-servers-are-the-new-apis): Every business that published an API will publish an MCP surface for agents to act through. The protocol is now settled enough to build on, and its record so far shows that running these servers safely is an infrastructure job, not a side project. - [Power is the new capacity limit](https://praecise.com/blog/power-is-the-new-capacity-limit): The constraint on AI compute has moved from the chip to the connection. What is scarce now is electricity that has been secured, placed and measured, and the operators who can account for every watt will set the pace. - [Praecise Cloud: where agents run](https://praecise.com/blog/praecise-cloud-where-agents-run): Agents are moving from demos into production, and the question is no longer which model to call but where the agent lives. Praecise Cloud builds, places and supervises agents, MCP servers and their backends from source, on hardware Praecise owns. - [Praecise Furcate: machines that keep working](https://praecise.com/blog/praecise-furcate-machines-that-keep-working): As agents begin to read and act on physical sites, the question is who holds authority over the machines when the network is gone. Furcate is the open operating system that keeps that authority in hardware keys on site, and Praecise Furcate is the hardware, operations and support Praecise builds on it. - [Praecise Inbox: an address for every agent](https://praecise.com/blog/praecise-inbox-an-address-for-every-agent): When agents act for a business, the first question after anything goes wrong is who said what. Praecise Inbox gives every agent, person and machine its own address on a verified domain, and keeps everything said in one record a person can open and search. - [Praecise Metal: the machines underneath](https://praecise.com/blog/praecise-metal-the-machines-underneath): Buying racks, accelerators, power and cooling is the easy part of running AI on owned hardware; the operating capability is what takes years. Praecise Metal is that capability, measured from instruments on the machine, from core clusters to edge cells. - [Skerry: payments agents can make](https://praecise.com/blog/skerry-payments-agents-can-make): Agents are about to become customers, suppliers and treasurers, and the payment system was built for a person clicking a button. Skerry is one programmable surface over virtual accounts, stablecoin conversion and cross-border payouts, scoped for agents; Bridge and Stripe perform the regulated acts under their own licences. - [Software now outnumbers people online](https://praecise.com/blog/software-now-outnumbers-people-online): In 2026 automated requests for web pages passed human ones, and traffic from agents grew fastest of all. At the same time, a large share of corporate agent projects are forecast to be cancelled. Both facts point to the same management task. - [Speculation pays on structure, not prose](https://praecise.com/blog/speculation-pays-on-structure-not-prose): Our measurements show speculative decoding speeding up SQL by 3.5x and code by 1.5x on one GPU, and doing nothing for prose. The useful decision is whether to speculate at all, and memory bandwidth sets the ceiling on everything else. - [The books close themselves](https://praecise.com/blog/the-books-close-themselves): E-invoicing mandates across Europe, the Gulf and Asia are turning invoices into structured data that software can read and act on. The hard problem in finance operations is no longer typing numbers in. It is deciding what an agent is allowed to do with them. - [The record is the control](https://praecise.com/blog/the-record-is-the-control): Regulators and standards bodies already require AI systems to keep logs. For agents that act on someone's behalf, the record is more than a compliance artefact: it is the mechanism that makes delegation safe in the first place. - [What an agent needs from a cloud](https://praecise.com/blog/what-an-agent-needs-from-a-cloud): GPUs answer the question of how an agent thinks. A working agent also needs compute that stays up, an address and a record, a way to pay, and limits its owner sets. Here is how the market covers each today. - [Who allowed the spend?](https://praecise.com/blog/who-allowed-the-spend): Card networks, Google, Coinbase and stablecoin regulators are each building a piece of how agents pay. The control that matters most is simpler than any of them: a limit the owner set, and a record of every payment made under it. ## Contact Early workloads are placed by conversation: contact@praecise.com. Say what you run and where it runs today. ## Pages - [Home](https://praecise.com): the gaps, the stack, the suite, who it serves. - Products: [Metal](https://praecise.com/metal), [Furcate](https://praecise.com/furcate), [Cloud](https://praecise.com/cloud), [Inbox](https://praecise.com/inbox), [Ipnops](https://praecise.com/ipnops), [Skerry](https://praecise.com/skerry). - [Open source](https://praecise.com/lab): Praecise Harness and Praecise Engine. - [Engineering notes](https://praecise.com/research): research notes measured on hardware. - [About](https://praecise.com/about) - [Contact](https://praecise.com/contact)