The world is advancing toward a future in which over 10 billion humans coexist with trillions of autonomous agents in a superintelligent universe. However, the cloud-hosted systems that currently dominate AI’s operational models lack the architectural elasticity to support this growing demand for compute resources. Championed by managed GPU clusters and centralized data centers, cloud AI relies on infrastructure that expands the operational scope of AI, requiring users to literally “rent” intelligence or the tools to develop it.

Tether argues that this system has a weak advantage, similar to scaling a database by simply buying a bigger server. AI scaling, in contrast, amplifies intelligence and availability. Cloud-hosted AI falters in both cases. Infinite scalability and universality in AI are therefore driven by the ability to deploy intelligent systems in any environment using readily available toolsets.

Tether’s AI research and development team believes that this ability is inherent in edge-optimized, local AI: self-hosted infrastructure for AI inference, training, and development on user-grade devices. Essentially, local AI converts intelligence into a portable capital asset readily available to the user, rather than a rented utility, as with cloud-hosted AI.

Operational Constraints for AI Developers and Businesses

Over 4,200 data centers — 43% of the world’s total — are located in the United States, eight times more than second-placed UK and third-placed Germany, which hosts more than twice as many data centers as all of Africa. This availability bias cuts across several other core infrastructure areas for AI development and routine use, impacting cost-effectiveness, performance, and data sovereignty.

Furthermore, data centers and other AI infrastructure, swamped with resource demands from unicorns and AI startups, continually adjust rental fees to offset operating costs and generate revenue. By extension, the high capital expenditure (CapEx) required for AI infrastructure is forcing companies to restructure their balance sheets. Major AI companies are expected to spend $500 billion on capital costs this year, a 30% increase from 2025 records. This is projected to reach $1.3 trillion by 2030, growing at 25% per annum.

Beyond availability and cost-efficacy, reliance on third-party infrastructure introduces additional operational factors and points of failure. Unplanned outages and abrupt changes in usage terms could significantly affect developers and end users.

In contrast, local AI operates autonomously, runs on infrastructure available to everyone, and works on any system. They are agnostic, run on-premise, and eliminate barriers to availability.

Reputation Capital of Centralized AI

In an ideal scenario, anyone should be able to confidently use AI tools without worrying about what happens to their data post-execution. However, third-party access to user data opens channels for data mismanagement — made more consequential by the quality of data unintentionally supplied by users during model training, fine-tuning, and inference.

34% of cybersecurity leaders identified “data leaks through generative AI” as their top concern for 2026, surpassing hacker capabilities for the first time. Additionally, the majority of the 670 data breach incidents reported in the first quarter of 2026 are either directly AI-driven or involve centralized data management infrastructure.

“Third-party” in this context includes AI companies and as-a-service infrastructure providers contracted to serve as a liaison between AI tools and users. Tens of the former are already headed to court in major class-action lawsuits over the handling of user data.

Local AI positions users as the only control point. User data is stored on-device and managed by software that has zero contact with external systems.

Tether AI’s research and development focuses on advancing machine intelligence as a readily available utility worldwide through technologies that let anyone build or use AI tools anywhere.

Universality for Superintelligence and Agnostic, On-Premise AI

Tether believes that efficient, self-hosted AI can transform machine intelligence into a new foundational element that powers new possibilities in dynamic systems. This will set in a new paradigm in which superintelligence is owned by the user. However, effective agnostic, on-premise AI can only be achieved through creative engineering — including modifications from the model level to the complete architecture that accommodate design differences. Tether is leading innovations in pursuit of this and contributing to open-source efforts to localize AI and abstract its complexities, expanding opportunities for further advancements in local and edge-first AI.

The goal is to decouple AI from the current siloed, controlled, and fragile model. Tether is re-engineering artificial intelligence and modifying existing technologies to achieve infinite, scalable intelligence.

The first task is to build a base infrastructure that operates as a self-governed unit. To this end, Tether developed the Pear Runtime and co-founded Holepunch. Pear Runtime and Holepunch employ decentralized resource networks, databases, and communication protocols to achieve a serverless P2P backend for edge applications.

Next, Tether addresses the heavy computational overhead of AI models by developing resource-efficient models and infrastructure that run locally on user-grade devices and across heterogeneous environments. It launched QVAC (QuantumVerse Automatic Computer), an AI research team, and a development framework for local-first and edge-first AI research and development.

This unit has led the development of:

  • A fine-tuning framework for Bitnet’s 13-billion-parameter LLM on regular devices, bypassing the GPU limitations of the ternary quantized model
  • QVAC Fabric LLM: A high-throughput local-first AI framework that transforms regular devices into sovereign compute machines for AI development, model training, and inference
  • An edge-first parameter-efficient fine-tuning framework for training the QVAC Fabric LLM locally on everyday devices and heterogeneous GPUs

Building on the QVAC framework, Tether’s AI coverage has expanded to include tools for deploying intelligence across diverse systems, from personal computers to interfaces for controlling smart homes and appliances. This includes runtime environments, training data, fine-tuning frameworks, and edge-optimized AI applications.

Tether has built a suite of tools that put local AI into practice across every layer of the stack, including:

  • QVAC Genesis I & II: Synthetic datasets for training AI models on STEM disciplines
  • WDK: An AI-ready wallet development framework that enables autonomous agents to build local and edge-first cryptocurrency wallets
  • QVAC MedPsy: A range of 1.7B and 4B parameter local-first medical AI models that run on heterogeneous everyday devices and produce more accurate and precise results than some larger medical AI models
  • QVAC Workbench: A general-purpose AI tool for researching, coding, and execution on edge devices

Developers equipped with these tools can implement local AI integrations across diverse systems through a serverless backend, models that run on resource-limited systems, and UI modules that simplify usage. This is further reinforced by the QVAC SDK, which consolidates all of Tether’s AI-related achievements to date. The QVAC SDK is a toolkit of prebuilt modules for components of the QVAC AI infrastructure, providing usage guides, integration contexts, and functional samples. It enables developers to build intelligent on-premises applications for any system without requiring permissions.

Committing to a Human-Centric Future for AI

Artificial intelligence is arguably the most human-targeted internet-based technology in history. In an AI-dominated future, the current user base will be only a fraction of the demand scale. However, institutional capital expenditure capacity is not unlimited, and the bloat in rental costs has no ceiling.

Tether’s local-first approach to AI acknowledges the relevance of machine intelligence to humans and its deep connection to everyday life. In response, it is dedicated to developing universally accessible AI that is modular enough to be embedded in the fabric of any device, system, or environment — from industrial servers to the smallest chip in a light bulb. In all of these cases, users own their AI, can build on their own terms without permission or external constraints, choose their own biases, and control how their data is used. Practically, this is the only way to ensure that superintelligence is successfully delivered to the billions of humans it is meant for.