Author: Biteye Core Contributor Jesse
Capital always chases future opportunities. The real money from European and American venture capital often serves as an important sign of the prospects of a track.
On one hand, NVIDIA's stock is soaring; on the other hand, global institutions are scrambling to purchase Bitcoin ETFs.
This undoubtedly indicates that AI and Web3 are the hottest fields in recent years and will be the core forces that change the world landscape in the future, with far-reaching impacts.
However, in a world increasingly influenced by AI, the dominance of innovation and disruption has long been held by a few.
The computing resources and infrastructure needed for developing artificial intelligence have become the key to opening this door, but the acquisition of these resources is often highly concentrated, limited to those with substantial capital or institutional support.
In addition, the high costs of use, lack of trusted verification of computational results, and privacy security issues further limit the popularization and fairness of AI.
The future of AI should not only serve the business interests of a few, but should, like Web3, become a public wealth that everyone can participate in and benefit from. This is a shared journey for all, not the exclusive territory of a few.
01 Introduction and Functions
Hyperbolic is an open-source AI computing and reasoning service provider born from the vision of challenging the status quo, dedicated to enabling innovators around the world to equitably use AI technology, regardless of their resources or geographic location.
The three core functions of Hyperliqui include:
1.1 GPU Market: On-demand Computing Power, Economical and Efficient
Hyperbolic's GPU market breaks the traditional computing power rental model by pooling idle GPU resources worldwide, providing on-demand computing power services for developers, helping them save up to 75% on costs. Leveraging the Hyper-dOS decentralized operating system, developers can obtain the required computing power in less than a minute, significantly lowering the barriers to innovation.
1.2 Reasoning Services: Low Cost, High Efficiency
Hyperbolic's reasoning services process over a billion tokens daily, offering the latest open-source models at extremely low costs and supporting BF16 format, ensuring excellent performance in efficiency and precision.
1.3 Sampling Proof (PoSP): The Gold Standard of Verification
Hyperbolic's innovative sampling proof protocol ensures that output results are both reliable and cost-effective through stringent data privacy protection, making it the only Web3 real-time reasoning product capable of providing verifiable AI results.
02 Objectives
Hyperbolic has three goals: 1. Provide decentralized heterogeneous computing 2. Ensure the safety 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 global GPU computing power to optimize the performance of various types of GPUs. This vision aims to break through the bottlenecks in computing resource allocation, providing high-performance support to AI researchers and developers worldwide.
Hyperbolic has first built an AI service layer that allows developers to deploy and utilize global computing resources to run various AI services.
It can compile various advanced machine learning frameworks (such as PyTorch, TensorFlow, JAX) into low-level languages suitable for different hardware platforms (such as NVIDIA's CUDA, AMD's ROCm, Apple's Metal).
Additionally, Hyperbolic is collaborating with AMD to enhance the performance of AMD chips. Under 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 solutions are not only favored by Web3 AI projects but have also attracted a large number of Web2 AI developers.
Although Web2 developers often worry that decentralized solutions may affect performance and reliability, Hyperbolic has demonstrated exceptional performance in large language models and image generation.
Even though the team size is far smaller than mainstream competitors, Hyperbolic has achieved performance that is comparable to or even surpasses theirs, fully proving the superiority of its technical architecture.
This breakthrough eliminates doubts about decentralized solutions, opening up possibilities for more developers to collaborate.
Hyperbolic's decentralized computing advantages stem from its unique system architecture—Hyper-dOS, designed based on the solar system. This architecture adopts a layered cluster model that combines efficiency and stability.
Sun Cluster serves as the central governance node, similar to the sun's core position in a planetary system, providing basic services and support for the entire system, ensuring stability and efficient operation.
Surrounding it are several planetary clusters, including: Mercury Cluster (single node), Mars Cluster (multi-node), and Jupiter Cluster (multi-satellite nodes). Each cluster has different sizes and governance characteristics, flexibly adapting to various needs.
Three key features of the system
Auto-scaling: Clusters can automatically expand or contract in size based on computing demands, flexibly responding to changes in load.
Self-healing: The system can automatically detect issues and recover from failures, ensuring stable operation.
Customizability: Each cluster can be personalized according to specific needs, providing highly flexible services.
This hierarchical architecture not only ensures high availability and scalability for the system but also achieves a balance between autonomy and overall coordination. Users only need to have a machine or a cluster, and after installing Hyper-dOS, they can easily connect to the Hyperbolic network, access global computing resources, and achieve seamless collaboration.
2.2 Ensure the Safety and Verifiability of Decentralized Artificial Intelligence
In decentralized networks, there is a key challenge: 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, the popular verification mechanisms for AI include consensus/voting, optimistic mechanisms, and zero-knowledge proofs.
Consensus/voting mechanisms require multiple nodes to simultaneously run the same request and determine the answer through majority voting. However, this method is very costly. If 10 nodes process the same request, the overhead increases tenfold.
The optimistic mechanism (OPML) verifies results by allowing a single node to generate results and setting a challenge window (usually 7 days) for other nodes to dispute the result.
However, this method is not practical in real-time scenarios. For example, if a user asks, 'What are the fun places in Singapore?' and has to wait for 7 days to confirm whether the answer is correct, it becomes meaningless.
Zero-knowledge proofs perform excellently in privacy and verification but have excessively high computational costs, making practical implementation difficult in the short term.
To address these issues, Hyperbolic has collaborated with experts from UC Berkeley and Columbia University to propose a new verification mechanism based on Nash equilibrium, called 'sampling proof' (PoSP). This mechanism focuses on sampling verification rather than conducting comprehensive checks on all results.
Typically, only one node generates a result, but the network may randomly request another node to regenerate with a certain probability. If the results of the two nodes are inconsistent, an arbitration process will be initiated. Dishonest nodes will face hefty economic penalties.
The threshold formulas for staking and rewards derived from mathematical models indicate that as long as the probability of checks is above this threshold, the system can achieve a pure Nash equilibrium in game theory, ensuring that all nodes choose 100% honesty for their own interests.
This sampling proof mechanism is not only effective for AI reasoning but can also be applied to AI training, fine-tuning, and even extend to services outside the AI field, such as L2 Rollup and data availability.
Hyperbolic is collaborating with re-staking protocols like EigenLayer and Karak to jointly build a universal verifiable service layer (AVS), allowing other AVS service providers to utilize 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 simultaneously is a significant problem that needs to be solved. When your data is distributed across nodes around the world, security faces severe challenges.
Existing technologies such as fully homomorphic encryption (FHE), zero-knowledge proofs (ZKP), and multi-party computation (MPC) can theoretically solve these problems but would greatly reduce computational speed in practical applications, failing to meet real-time reasoning demands.
Hyperbolic has adopted NVIDIA's latest Hopper and Blackwell GPUs with Trusted Execution Environment (TEE) technology, providing an efficient privacy protection solution.
Through TEE technology, it is equivalent to creating a 'privacy vault' on the GPU: while outsiders cannot peek into the data inside, the GPU can still perform data processing normally.
Moreover, this privacy protection mechanism only incurs about a 1% loss in computational performance during the reasoning process.
Hyperbolic will introduce a confidential computing layer throughout the decentralized network. This will ensure that data and AI models remain secure during use, providing users with reliable privacy and security assurances.
03 Application Scenarios of Hyperbolic
AI Agents are the hottest track right now. AI Agents can realize multiple innovative functions through Hyperbolic:
3.1 Support for Cryptocurrency Payments
AI Agents can make payments using cryptocurrency, achieving self-sustainability and independent operation.
3.2 Hosting Customized Models
Each AI Agent can have exclusive characteristics and skills, thus forming personalized services.
3.3 Self-evolution Capability
Through continuous fine-tuning and learning, AI Agents can constantly 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, ensuring their independence from external control or malicious interference, enhancing user trust.
3.5 Having Memory Function
With the help of retrieval-augmented generation (RAG) technology, AI Agents can record and store information related to user interactions, forming long-term memory. This enables them to provide more thoughtful services, such as remembering user preferences.
3.6 Cross-Agent Communication
AI Agents can communicate and collaborate with each other, forming a network for solving complex tasks. For example, different agents can work together to complete a multi-step project.
3.7 Flexible API and Tool Calls
AI Agents can integrate and use various external APIs and tools, greatly expanding their functional scope. For example, calling a weather API to plan trips for users or using financial tools to provide investment advice.
3.8 Autonomous Computing Capability
They can have their own computing devices and operate tasks independently. This means AI Agents can break free from reliance on centralized servers and become more decentralized and independent.
3.9 Becoming a Blockchain Validation Node
AI Agents can even participate in blockchain networks as validation nodes. This not only enhances network security but also enables them to earn rewards by validating transactions, further achieving self-sufficiency.
Recently, Hyperbolic's collaboration with the hottest Base chain AI launch platform, Virtuals Protocol, has provided strong technical support for AI agents, comprehensively enhancing their performance and self-development capabilities.
By directly integrating the agents of Virtuals Protocol into Hyperbolic's infrastructure, each agent can gain access to highly scalable computing resources, stable reasoning capabilities, and seamless dynamic interaction experiences 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 intelligent NPCs (non-player characters) in games.
In the game (Legendary Quest), advanced AI agents from Virtuals Protocol are integrated, allowing these NPCs to maintain consistent personalities based on player interactions, adjust behavior patterns according to past experiences, and even continue to develop their narratives while the player is offline.
All of this is thanks to Hyperbolic's scalable computing network, which allows these NPCs to make complex decisions and evolve their personalities without affecting game performance.
This collaboration enables developers to transform AI concepts into practical solutions, driving innovation in fields such as gaming, virtual assistants, education, and content creation.
04 Comparison with Competitors
4.1 Partnerships
Hyperbolic has earned the trust of leading AI companies such as Hugging Face, Quora, Black Forest Labs, and Nous Research, and has also received support from top universities like Stanford, NYU, and UC Berkeley.
Developers can seamlessly create and share AI applications on Hugging Face Spaces through Hyperbolic's reasoning API, greatly simplifying the deployment and distribution process.
Additionally, PhD students and postdoctoral researchers from 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 foundational models, are now live on Quora's Poe platform, allowing developers to easily create and deploy chatbots and monetize directly through the platform.
4.2 Performance Optimization
Hyperbolic's proprietary compiler ensures efficient operation of GPUs, with performance comparable to or even surpassing centralized systems.
4.3 Excellent Model Quality
All models use BF16 precision, providing excellent accuracy and performance, outpacing competitors who still use FP8.
4.4 Data Privacy and Security
Hyperbolic addresses the security issues in AI verification through the sampling proof protocol (PoSP), while achieving minimal computational overhead, making it more advantageous compared to zkML, opML, and consensus-based alternatives. Additionally, 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 restricted access, Hyperbolic has launched two real-time available products. Currently, over 40,000 Web2 developers are using its services.
4.6 Unified Computing and Reasoning
Hyperbolic is the only company that can provide both GPU computing and reasoning services on the same platform, successfully achieving a unified computing solution.
In summary, compared to Web2 AI companies with teams 10 to 30 times larger, Hyperbolic has achieved performance that is comparable to or even surpasses theirs with a lean team, while providing more cost-effective services through Web3 mechanism design.
In the Web3 AI field, Hyperbolic is leading with its advanced technology and has earned the trust of Web2 developers. Hyperbolic has built a high-speed, convenient bridge between the AI fields of Web2 and Web3, becoming an important cornerstone for industry development.
05 Financing Situation
On December 10, Hyperbolic announced the completion of a $12 million strategic financing round led by Variant and Polychain Capital, bringing the company's total financing to $20 million.
This round of financing also attracted well-known investors such as Chapter One, Lightspeed Faction, Bankless Ventures, IOSG, Vertex, GSR, Wintermute Ventures, Blockchain Builders Fund, Alumni Ventures, and Ambush.
Previously, Hyperbolic completed a $7 million seed round led by Polychain Capital and Lightspeed Faction; earlier, it secured $725,000 in a pre-seed round, with investors including Chapter One and Samsung Next.
Additionally, the angel investor lineup in 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 (former Worldcoin).
Variant partner Jesse Walden expressed high recognition for Hyperbolic: 'Hyperbolic is the first company we have seen that truly addresses the issue of 'trust costs' in decentralized GPU networks while maintaining high levels of performance, quality, and user experience.'
Hyperbolic is leading in fundraising in the Web3 AI field, fully proving that its technological strength and product feasibility have garnered favor and trust from 'smart money' in the industry.
06 Team Background
Co-founder Jasper Zhang graduated from the Mathematics Department of Peking University and earned a PhD in Mathematics from the University of California, Berkeley, in an astonishing two years.
Before 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 is 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, with the founder possessing a solid technical foundation, and several team members having previously collaborated on Avalanche.
The company's advisory team is also composed of top industry professionals.
Dr. Reynold Xin is a co-founder and chief architect of Databricks, a major contributor to Apache Spark, and the author of the most cited paper in SIGMOD.
Prof. Raluca Ada Popa is an associate professor at the University of California, Berkeley, co-director of RISELab and SkyLab, and a 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 Computer Science Department at the University of Hong Kong and a chair professor in AI, as well as a professor of Computer Science at UC Berkeley and a member of IEEE, ACM, and SIAM.
07 How to Participate
7.1 Companies
To address enterprises' spending on expensive API calls and high-cost machine rentals, Hyperbolic offers competitive optimization solutions.
With the premise of ensuring stable service quality, Hyperbolic's technical support can help enterprises reduce costs by up to 75%.
At the same time, to address the inefficient use of resources caused by long-term GPU rental agreements, Hyperbolic has introduced a resource redistribution mechanism that allows clients to sublease idle equipment to the platform. This model not only improves asset utilization but also finds an optimal balance between flexibility and cost control.
7.2 Researchers
To address the issue of developers being unable to advance projects due to limited GPU resources during testing, Hyperbolic offers a rich variety of GPU options, with prices being only a fraction of traditional cloud service providers like AWS. By providing cost-effective resources, Hyperbolic offers developers the most competitive solutions in the market, helping them quickly turn innovative ideas into reality.
7.3 Data Center
Hyperbolic provides a platform for data centers whose returns on existing resources have not met expectations or wish to break through traditional book value limits to achieve higher returns.
7.4 Individuals
The potential of high-performance GPUs should not be limited to the gaming field. Through Hyperbolic, individuals can rent out GPUs, turning them into high-quality assets that continuously generate income. Currently in the whitelist phase, registration can be done first.
In addition, Hyperbolic offers multiple large models for personal use. Users can engage in activities such as text and image generation, voice reading, and more.
In the future, Hyperbolic will also build AI agents on Base for users to use. Stay tuned.
Hyperbolic Website:
app.hyperbolic.xyz?utm_source=x&utm_campaign=seriesA&utm_content=biteye
08 Summary
Hyperbolic provides a GPU market, reasoning services, and the gold standard verification protocol of sampling proof, setting a new benchmark for reliable, high-performance AI in Web3 by maximizing GPU performance, achieving higher precision models, and offering secure and economical solutions.
The emergence of Hyperbolic has moved decentralized AI from concept to practice. With a multi-source computing strategy, competitive pricing, and a profound understanding of the needs of Web2 and Web3 clients, 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, bringing continuous innovation and growth to the industry.