Trust is the foundation of artificial intelligence. Without it, even the most powerful AI models are just black boxes generating unreliable outputs. At Hyperbolic we are champions of decentralized AI infrastructure. This truth becomes even more critical when we adopt decentralization for AI—a move that promises greater accessibility and lower costs, but introduces new challenges in ensuring reliable results.
We are the only company delivering scalable, verifiable decentralized AI that's already processing over a billion tokens daily for more than 40,000 developers. Through our groundbreaking Proof of Sampling (PoSP) protocol, developed in collaboration with researchers from UC Berkeley and Columbia University, we've solved the trust problem in decentralized AI while maintaining the cost benefits that decentralization promises.
The Hyperbolic Decentralized Operating System (Hyper-dOS) was developed to provide a robust and scalable backend architecture to efficiently manage a vast network of globally distributed GPUs. These compute resources are organized into a solar system-like network as independent planets coordinated by a sun cluster that governs and sustains each planetary cluster. This central cluster provides essential services and support to ensure the stability and efficiency of the overall system.
Hyper-dOS is so easy to use that, once installed on an underutilized machine, its compute power is seamlessly integrated with our distributed network in minutes for AI builders to quickly and easily rent scalable and cost-effective compute.
While this model provides an intuitive way for owners of GPUs to monetize their valuable resource while making it accessible to builders changing the game in AI, trusting the behavior of third party nodes poses a different challenge. How can one be certain that their inference output from a third-party node is valid?
The Decentralized AI Verification Problem
Verification has been a challenge for decentralized technologies for a long time, with just a few methodologies emerging to address the issue. These traditional verification methods often rely on redundant computation or complex cryptographic proofs, which can introduce significant computational overhead to the system.
While such time and computation-heavy methods may have worked for the cryptographic use cases required in decentralized finance, they do not present as a practical solution for verification in decentralized AI inference. It isn’t realistic to expect users of AI to wait for 10 days as the generated result of their request is verified. Further, current verification best practices can increase the cost of inference by up to 300%, negating any previously held cost advantage of decentralized GPU networks.
Hyperbolic’s Proof of Sampling: a Game-Theoretical Approach to Verification of Decentralized Systems
Hyperbolic's novel Proof of Sampling (PoSP) takes a fundamentally different approach to verification. Developed by our co-founder and CEO Jasper Zhang in collaboration with researchers from UC Berkeley and Columbia University, it secures the decentralized network by relying on a Nash Equilibrium-like approach to verifying outputs.
A Nash Equilibrium can be easily understood in the context of a train ticket verification system. Say a ticket costs $10, and to increase efficiency by reducing the number of train conductors that are required on each train, tickets are checked via a random sample of travelers with a $100 fine if a traveler is found to be without a ticket. One would think that this would open the door for grifters to feel more confident boarding the train without a ticket and thus getting a free ride. Logically, however, if the train conductors check just 1 in 10 tickets, there is a 10% chance that a grifter would need to pay a $100 fine, which works out to $10 for every ride in 10 rides, the same cost as the ticket. The optimized random sampling rate removes any incentive for a rider to act dishonestly while allowing for greatest possible efficiency for the train conductors.
Just as the train conductors don’t have to check 100% of the tickets and are still able to ensure that 100% of the riders pay for their tickets, PoSP saves the computational time and energy of checking outputs from all nodes, and instead verifies a strategic proportion of outputs.
Proof of Sampling in a Decentralized GPU Network
Our verification process begins when a client submits an inference request to the Hyperbolic network. The request is assigned to available third-party nodes based on computational requirements. Meanwhile, the system computes the appropriate verification sampling rate based on node reputation and stake, increasing for newer nodes in the network and decreasing for more nodes that have been proven reliable.
If an output is selected for verification, the system duplicates the work on trusted validator nodes. Should a validator find that the output is invalid, it will send the computation back through the orchestrator to trigger an arbitration process.
Because the cost of arbitration and the probability of validation are established to remain as a Nash Equilibrium, we have created a system where honest computation is the dominant strategy for participating GPUs. We have thus avoided having to spend the computational energy on checking every single result from our decentralized network, maintaining cost efficiency and ensuring practical usability.
Join Hyperbolic’s Verified AI Ecosystem
PoSP represents a fundamental breakthrough in making the utilization of coordinated decentralized GPU resources both trustworthy and cost-effective.
At Hyperbolic, our mission is to build an open and accessible AI ecosystem. With PoSP, we are empowering AI builders looking to shape the future of AI by making verified GPU resources available to all.
Join our community of AI builders confidently running AI inference on our decentralized GPU network and take your ideas hyperbolic at app.hyperbolic.xyz.
About Hyperbolic
Hyperbolic is democratizing AI by delivering a complete open ecosystem of AI infrastructure, services, and models. Through coordinating a decentralized network of global GPUs and leveraging proprietary verification technology, developers and researchers have access to reliable, scalable, and affordable compute as well as the latest open-source models.
Founded by award-winning Math and AI researchers from UC Berkeley and the University of Washington, Hyperbolic is committed to creating a future where AI technology is universally accessible, verified, and collectively governed.
Website | X | Discord | LinkedIn | YouTube | GitHub | Documentation
