Written by: Jesse, core contributor of Biteye
Editor: Crush, core contributor of Biteye
Capital always chases future opportunities. The real money of European and American venture capital is often an important indicator of the prospects of a track.
On the one hand, Nvidia's stock is rising steadily; on the other hand, global institutions are scrambling to buy Bitcoin ETFs.
This undoubtedly shows that AI and Web3 are the hottest fields in recent years, and will also be the core force in changing the world pattern in the future, with far-reaching impact.
However, in a world increasingly influenced by AI, the power to innovate and disrupt has long been in the hands of a few.
The computing resources and infrastructure needed to develop artificial intelligence have become the key to opening this door, but access to these resources is often highly concentrated and limited to those with strong capital or institutional support.
In addition, high usage costs, lack of reliable verification of computing results, and privacy and security issues further limit the popularity and fairness of AI.
The future of AI should not only serve the commercial interests of a few people, but should become a public asset that everyone can participate in and benefit from, just like Web3. This is a common journey for everyone, not the exclusive domain of a few.
01 Introduction and Functions
Hyperbolic is an open source AI computing and reasoning service provider born out of a vision to challenge the status quo and enable innovators around the world to have equal access to AI technology, regardless of their resources or geographic location.
Hyperliqui's three core functions include:
1.1 GPU Market: On-demand computing power, cost-effective
Hyperbolic's GPU market breaks the traditional computing power rental model. By gathering idle GPU resources around the world, it provides developers with on-demand computing power services, helping them save up to 75% of costs. Relying on the Hyper-dOS decentralized operating system, developers can get the required computing power in less than a minute, greatly lowering the threshold for innovation.
1.2 Reasoning Service: Low Cost, High Efficiency
Hyperbolic's inference service processes more than one billion tokens every day, provides the latest open source models at a very low cost, and supports the BF16 format, ensuring excellent performance in efficiency and accuracy.
1.3 Proof of Sampling (PoSP): The Gold Standard for Verification
Hyperbolic's original sampling proof protocol ensures that the output results are both reliable and cost-effective through strict data privacy protection, making it the only Web3 real-time inference product that can provide verifiable AI results.
02 Goal
Hyperbolic has three goals: 1. Provide decentralized heterogeneous computing 2. Ensure the security and verifiability of decentralized artificial intelligence 3. Protect privacy in decentralized AI.
2.1 Providing decentralized heterogeneous computing
Hyperbolic is committed to building a scalable system that integrates the global GPU computing power to optimize the performance of various types of GPUs. This vision aims to break the bottleneck of computing resource allocation and provide high-performance support to AI researchers and developers around the world.
Hyperbolic first built an AI service layer, allowing developers to deploy and utilize global computing resources to run different AI services.
It can compile various high-level machine learning frameworks (such as PyTorch, TensorFlow, JAX) into low-level languages adapted to different hardware platforms (such as NVIDIA's CUDA, AMD's ROCm, Apple's Metal).
In addition, Hyperbolic has also worked with AMD to improve the performance of AMD chips. With Hyperbolic's optimization, the input throughput of the Llama3-8B model on the AMD MI250 platform increased by 120.4% and the output throughput increased by 144.8%.
Hyperbolic's solution is not only favored by Web3 AI projects, but also attracts a large number of Web2 AI developers.
Although Web2 developers often worry that decentralized solutions may affect performance and reliability, Hyperbolic has demonstrated excellent performance in the fields of large language models and image generation.
Even though the team size is much smaller than that of mainstream competitors, Hyperbolic still achieves performance comparable to or even exceeds them, fully demonstrating the superiority of its technical architecture.
This breakthrough eliminates doubts about decentralized solutions and opens up the possibility of collaboration for more developers.
Hyperbolic's decentralized computing advantage stems from its unique system architecture, Hyper-dOS, which is inspired by the solar system. The architecture adopts a hierarchical cluster model that combines efficiency and stability.
Sun Cluster is the central governance node, similar to the core position of the sun in the planetary system, providing basic services and support for the entire system to ensure stability and efficient operation.
Surrounding it are multiple planetary clusters, including: Mercury Cluster (single node), Mars Cluster (multiple nodes) and Jupiter Cluster (multiple satellite nodes). Each cluster has different scale and governance characteristics, which can be flexibly adapted to different needs.
Three key features of the system
Automatic scaling: The cluster can automatically expand or shrink according to computing needs, flexibly responding to load changes.
Self-healing: The system automatically detects problems and recovers from failures, ensuring stable operation.
Customizability: Each cluster can be individually configured according to specific needs, providing highly flexible services.
This layered architecture not only ensures high availability and scalability of the system, but also achieves a balance between autonomy and overall coordination. Users only need to have a machine or a cluster, install Hyper-dOS, and then they can easily access the Hyperbolic network, obtain global computing resources, and achieve seamless collaboration.
2.2 Ensuring the security and verifiability of decentralized AI
A key challenge in decentralized networks is how to ensure that the results generated by random nodes are correct. Security and verifiability have always been unresolved issues in deployed AI systems.
Currently, popular verification mechanisms for AI include consensus/voting, optimistic mechanisms, and zero-knowledge proofs.
The consensus/voting mechanism requires multiple nodes to run the same request at the same time and determine the answer by majority vote. However, this approach is very expensive. If 10 nodes process the same request, the overhead will increase 10 times.
The Optimistic Mechanism (OPML) verifies the result by allowing a single node to generate the result and setting a challenge window (usually 7 days) for other nodes to raise objections.
However, this approach is not practical in real-time scenarios. For example, if a user asks "What are the fun places in Singapore?", it is meaningless if they have to wait 7 days to confirm whether the answer is correct.
Zero-knowledge proof excels in privacy and verification, but its computational cost is too high to be put into practical use in the short term.
To solve these problems, Hyperbolic, together with experts from the University of California, Berkeley and Columbia University, proposed a new verification mechanism based on Nash equilibrium, called "Proof of Sampling" (PoSP). This mechanism is centered on sampling verification rather than a comprehensive check of all results.
Normally, only one node generates the result, but the network will randomly ask another node to regenerate it with a certain probability. If the results of the two nodes are inconsistent, an arbitration process will be initiated. Dishonest nodes will be subject to high economic penalties.
The threshold formula for staking and rewards derived from a mathematical model shows that as long as the probability of inspection is higher than the threshold, the system can reach a pure Nash equilibrium in game theory, ensuring that all nodes choose to be 100% honest for their own benefit.
This sampling proof mechanism is not only effective for AI reasoning, but can also be applied to areas such as AI training and fine-tuning, and even extended to services outside the AI field, such as L2 Rollup and data availability.
Hyperbolic is working with re-staking protocols such as EigenLayer and Karak to build a universal verifiable service layer (AVS), so that other AVS service providers can also use this verification mechanism to ensure the security and reliability of their services.
2.3 Protecting Privacy in Decentralized AI
In a decentralized AI network, how to ensure data privacy and model integrity at the same time is a big problem that needs to be solved. When your data is distributed on nodes around the world, security faces severe challenges.
Existing technologies such as fully homomorphic encryption (FHE), zero-knowledge proof (ZKP) and multi-party computation (MPC) can solve these problems in theory, but in practical applications they will greatly reduce the computing speed and cannot meet the needs of real-time reasoning.
Hyperbolic uses the Trusted Execution Environment (TEE) technology on NVIDIA's latest Hopper and Blackwell GPUs to provide an efficient privacy protection solution.
Through TEE technology, it is equivalent to creating a "privacy safe" on the GPU: although the outside world cannot spy on the data content, the GPU can still complete data processing normally.
Moreover, this privacy-preserving mechanism only loses about 1% of computing performance during inference.
Hyperbolic will introduce a confidential computing layer throughout the decentralized network. This will ensure that data and AI models are always secure during use, providing users with reliable privacy and security protection.
03 Application Scenarios of Hyperbolic
AI Agent is the hottest track at present. AI Agent can achieve many innovative functions through Hyperbolic:
3.1 Support encrypted payment
AI Agent can be paid through cryptocurrency, making it self-sustaining and independently operating.
3.2 Hosting Custom Models
Each AI Agent can have its own unique characteristics and skills to form personalized services.
3.3 Self-evolution capability
Through continuous fine-tuning and learning, AI Agents can continuously improve their capabilities based on user needs or environmental changes, making them more efficient and intelligent.
3.4 Verifiable Reasoning
The reasoning process of AI Agents is transparent and verifiable, which ensures their independence and protects them from external control or malicious interference, thus enhancing user trust.
3.5 With memory function
With the help of Retrieval Augmentation Generation (RAG) technology, AI Agents can record and store information about interactions with users to form long-term memories. This enables them to provide more thoughtful services, such as remembering user preferences.
3.6 Inter-Agent Communication
AI agents can communicate and collaborate with each other to form a network for solving complex tasks. For example, different agents can collaborate to complete a multi-step project.
3.7 Flexible API and tool calls
AI Agent can integrate and use a variety of external APIs and tools to greatly expand its functionality. For example, it can call weather APIs to plan trips for users, or use financial tools to provide investment advice.
3.8 Autonomous computing capabilities
They can have their own computing devices and run tasks independently. This means that AI Agents can get rid of their dependence on centralized servers and become more decentralized and independent.
3.9 Become a blockchain verification node
AI Agents can even participate in blockchain networks and serve as validation nodes, which not only enhances network security but also allows them to earn rewards by validating transactions, further achieving self-sufficiency.
Recently, Hyperbolic’s collaboration with Virtuals Protocol, the hottest Base chain AI launch platform, has provided strong technical support for AI agents, comprehensively improving their performance and self-development capabilities.
By connecting Virtuals Protocol agents directly to Hyperbolic’s infrastructure, each agent gains access to the highly scalable computing resources, stable reasoning capabilities, and seamless dynamic interaction experience provided by the Hyperbolic API, maintaining efficient and consistent performance regardless of the number of agents or task complexity.
This collaboration not only enhances the computing power of AI agents, but also improves their adaptability and intelligence in diverse application scenarios.
For example, Hyperbolic’s infrastructure provides persistent memory and personality development capabilities for the game’s intelligent NPCs (non-player characters).
In the game (Legendary Quest), Virtuals Protocol's advanced AI agents are integrated. These NPCs are able to maintain consistent personalities based on player interactions, adjust behavior patterns based on past experiences, and even continue to develop their own plots when players are offline.
All of this is made possible by Hyperbolic’s scalable computing network, which enables these NPCs to make complex decisions and evolve their personalities without affecting game performance.
The collaboration enables developers to turn AI concepts into real-world solutions, driving innovation in gaming, virtual assistants, education, content creation, and more.
04 Comparison with competitors
4.1 Partnership
Hyperbolic has won the trust of leading AI companies such as Hugging Face, Quora, Black Forest Labs, and Nous Research, and has also been supported by top universities such as Stanford University, New York University, and University of California, Berkeley.
Developers can seamlessly create and share AI applications on Hugging Face Spaces through Hyperbolic's inference API, greatly simplifying the deployment and distribution process.
In addition, doctoral students and postdoctoral researchers at Stanford University, Cornell University, and New York University can enjoy up to 75% discounts on GPU rentals, significantly reducing computing costs.
Hyperbolic’s AI models, including the base model, are now available on Quora’s Poe platform, enabling developers to easily create and deploy chatbots and monetize them directly through the platform.
4.2 Optimizing Performance
Hyperbolic's proprietary compiler ensures that the GPU runs efficiently, with performance comparable to or even exceeding that of centralized systems.
4.3 Excellent model quality
All models use BF16 precision, providing superior accuracy and performance, ahead of competitors who still use FP8.
4.4 Data Privacy and Security
Hyperbolic solves the security issues in AI verification through the Proof of Sampling protocol (PoSP) while achieving minimal computational overhead, which is more advantageous than zkML, opML and consensus-based alternatives. In addition, Hyperbolic does not store user data at all, further protecting privacy.
4.5 Mature real-time products
Unlike many Web3 AI projects that are still under development or have limited access, Hyperbolic has already launched two live products and has over 40,000 Web2 developers using its services.
4.6 Unified Computation and Reasoning
Hyperbolic is the only company that can provide both GPU computing and inference services on the same platform, successfully implementing a unified computing solution.
In summary, compared with Web2 AI companies with team sizes 10 to 30 times larger, Hyperbolic has achieved comparable or even superior performance with just a streamlined team, while providing more cost-effective services through Web3 mechanism design.
In the field of Web3 AI, Hyperbolic is far ahead with its leading technology and has won the trust of Web2 developers. Hyperbolic has built a high-speed and convenient bridge between the AI fields of Web2 and Web3, becoming an important cornerstone for promoting the development of the industry.
05 Financing
On December 10, Hyperbolic announced that it had completed a $12 million strategic financing round led by Variant and Polychain Capital, bringing the company’s total financing to $20 million.
The round also attracted notable investors such as Chapter One, Lightspeed Faction, Bankless Ventures, IOSG, Vertex, GSR, Wintermute Ventures, Blockchain Builders Fund, Alumni Ventures and Ambush.
Previously, Hyperbolic had completed a $7 million seed round of financing led by Polychain Capital and Lightspeed Faction; earlier, it also received $725,000 in pre-seed round financing, with investors including Chapter One and Samsung Next.
In addition, the angel investor lineup for this round of financing is also very strong, including Sreeram Kannan (EigenLayer), Devin Walsh (Uniswap Foundation), Ethan Sun (MyShell), Daniel Shorr (Modulus), Bidhan Roy (Bagel), Ying Sheng and Lianmin Zheng (LMSYS), Dillon Rolnick (Nous Research), Alex Atallah (OpenRouter), Chainyoda, Comfy Capital, Nicola Greco (Protocol Labs), Alex Atallah (OpenRouter) and Thomas Scott (formerly Worldcoin).
Jesse Walden, partner at Variant, highly recognized Hyperbolic: "Hyperbolic is the first company we have seen that truly solves the 'trust cost' problem in decentralized GPU networks while maintaining a high level of performance, quality, and user experience."
Hyperbolic is in a leading position in financing in the Web3 AI field, which fully proves that its technical strength and product feasibility have won the favor and trust of "smart money" in the industry.
06 Team Background
Co-founder Jasper Zhang graduated from the Department of Mathematics at Peking University and obtained a Ph.D. in Mathematics from the University of California, Berkeley in just two years at an astonishing speed.
Prior to founding Hyperbolic, he worked as a quantitative researcher at Citadel Securities and as a senior blockchain researcher at Avalanche.
Co-founder and part-time CTO Yuzhen Jin holds a PhD in computer science from the University of Washington and served as a senior engineering manager at OctoAI before founding Hyperbolic.
Hyperbolic's team members all have backgrounds from top universities, the founders have a solid technical foundation, and many team members have previously worked together at Avalanche.
The company's advisory team is also composed of top industry professionals.
Dr. Reynold Xin is the co-founder and chief architect of Databricks, a major contributor to Apache Spark, and the author of the most cited paper at SIGMOD.
Prof. Raluca Ada Popa is an associate professor at the University of California, Berkeley, co-director of RISELab and SkyLab, and co-founder of Opaque Systems.
Prof. Ciamac C. Moallemi is a professor at Columbia Business School, a research advisor at Paradigm, and the director of the Briger Family Digital Finance Lab.
Prof. Yi Ma is the head of the Department of Computer Science and a chair professor in the field of AI at the University of Hong Kong. He is also a professor of computer science at the University of California, Berkeley, and a fellow of IEEE, ACM, and SIAM.
07 How to participate
7.1 Company
Hyperbolic provides a competitive optimization solution for enterprises' spending on expensive API calls and high-cost machine rentals.
While ensuring stable service quality, Hyperbolic's technical support can help enterprises reduce costs by up to 75%.
At the same time, in response to the inefficient use of resources caused by long-term GPU leasing agreements, Hyperbolic launched a resource reallocation mechanism that allows customers to sublease idle equipment to the platform. This model not only improves asset utilization, but also finds the optimal balance between flexibility and cost control.
7.2 Researchers
In order to solve the problem that developers cannot advance their project testing due to limited GPU resources, Hyperbolic provides a wide range of GPU options at a fraction of the price of traditional cloud service providers such as AWS. By providing cost-effective resources, Hyperbolic provides developers with the most competitive solutions in the market, helping them quickly turn innovative ideas into reality.
7.3 Data Center
Hyperbolic provides a platform to achieve higher returns for data centers that are not achieving the expected return on investment for existing resources or are looking to break through the limitations of traditional book value.
7.4 Individual
The potential of high-performance GPUs should not be limited to the gaming field. Through Hyperbolic, individuals can rent out GPUs and turn them into high-quality assets that continue to generate income. Currently in the whitelist stage, you can register first.
In addition, Hyperbolic provides a variety of large models for personal use. Users can perform activities such as text, image generation, and voice reading.
In the future, Hyperbolic will also build AI agents on Base for users to use. Stay tuned.
08 Conclusion
Hyperbolic provides a GPU marketplace, inference services, and the gold standard proof-of-sampling protocol, setting a new bar for reliable, high-performance AI for Web3 by maximizing GPU performance, higher precision models, and secure and cost-effective solutions.
The emergence of Hyperbolic has made decentralized AI move from concept to practice. With its multi-source computing strategy, competitive pricing, and deep understanding of Web2 and Web3 customer needs, Hyperbolic occupies a unique position in the ecosystem.
Hyperbolic's efforts in promoting the democratization and efficient use of computing resources will drive the development of the AI track and bring continuous innovation and growth to the industry.