How will the strongest new star Hyperbolic reshape the AI ​​track?

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 to change the world pattern in the future, with far-reaching influence.

However, in a world increasingly influenced by AI, the dominance of innovation and subversion has long been in the hands of a few people.

The computing resources and infrastructure required to develop artificial intelligence have become the key to opening this door, but the acquisition of these resources is often highly concentrated and limited to those with strong capital or institutional support.

In addition, high usage costs, lack of credible 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 territory of a few.

01 Introduction and Features

Hyperbolic is an open source AI computing and reasoning service provider, born out of a vision to challenge the status quo, dedicated to enabling innovators around the world to use AI technology equally, regardless of their resources or geographical location.

Hyperliqui's three core functions include:

1.1 GPU Market: On-demand computing power, economical and efficient

Hyperbolic's GPU market breaks the traditional computing power leasing 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 reasoning 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 Proof of Sampling protocol ensures that the output results are both reliable and cost-effective through strict data privacy protection, making it the only Web3 real-time reasoning product that can provide verifiable AI results.

02 Goals

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 Provide 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 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 use 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 also cooperates with AMD to improve the performance of AMD chips. With the optimization of Hyperbolic, 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 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 outstanding performance in the fields of large language models and image generation.

Even with a much smaller team size than mainstream competitors, Hyperbolic has achieved performance comparable to or even exceeding them, fully demonstrating the superiority of its technical architecture.

This breakthrough eliminates doubts about decentralized solutions and opens up the possibility of cooperation 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 layered 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 scales 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 its scale according to computing needs, and flexibly respond to load changes.

• Self-healing: The system can automatically detect problems and recover from failures to ensure stable operation.

• Customizability: Each cluster can be personalized according to specific needs to provide 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, and after installing Hyper-dOS, they can easily access the Hyperbolic network, obtain global computing resources and achieve seamless collaboration.

2.2 Ensure the security and verifiability of decentralized artificial intelligence

There is a key challenge in decentralized networks: 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 voting. However, this approach is very costly. 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 a result and setting a challenge window (usually 7 days) for other nodes to raise objections.

But 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 you have to wait 7 days to confirm whether the answer is correct.

Zero-knowledge proofs excel in privacy and verification, but the computational cost is too high to be practical in the short term.

To address these issues, Hyperbolic, in collaboration 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.

Usually, only one node generates a result, but the network randomly asks 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 pledge and reward derived from a mathematical model shows that as long as the probability of checking is higher than the threshold, the system can reach a pure Nash equilibrium state in game theory, ensuring that all nodes choose 100% honesty for their own benefit.

This sampling proof mechanism is not only effective for AI reasoning, but can also be applied to AI training, fine-tuning and other fields, 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 jointly 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 computing (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 peek into the data content, the GPU can still complete data processing normally.

Moreover, this privacy protection mechanism only loses about 1% of computing performance during the reasoning process.

Hyperbolic will introduce a confidential computing layer throughout the decentralized network. This will ensure that data and AI models are always in a safe state during use, providing users with reliable privacy and security guarantees.

03 Application scenarios of Hyperbolic

AI Agent is the hottest track at present. AI Agent can realize many innovative functions through Hyperbolic:

3.1 Support encrypted payment

AI Agent can make payments through cryptocurrency to achieve self-sustaining and independent operation.

3.2 Hosting customized models

Each AI Agent can have exclusive 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 according to 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 is not subject to external control or malicious interference, enhancing user trust.

3.5 Possessing 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 considerate 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 Calling

AI Agents can integrate and use a variety of external APIs and tools to greatly expand their functional scope. For example, calling a weather API to plan a user's itinerary, or using 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 verification nodes. This not only enhances network security, but also earns rewards by verifying transactions, further achieving self-sufficiency.

Recently, Hyperbolic’s collaboration with Virtuals Protocol, the most popular Base chain AI launch platform, has provided strong technical support for AI agents, comprehensively improving their performance and self-development capabilities.

By directly connecting Virtuals Protocol’s agents to Hyperbolic’s infrastructure, each agent can obtain the highly scalable computing resources, stable reasoning capabilities, and seamless dynamic interactive experience provided by the Hyperbolic API, maintaining efficient and consistent performance regardless of the number of agents or the complexity of the task.

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", 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 this is thanks to Hyperbolic's scalable computing network, which enables 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 areas such as games, virtual assistants, education, and content creation.

04 Comparison with Competitive Products

4.1 Partnerships

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 rental, 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 efficient operation on GPUs, with performance comparable to or even exceeding that of centralized systems.

4.3 Superior 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 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 Live Products

Unlike many Web3 AI projects that are still under development or have limited access, Hyperbolic has already launched two live products. Currently, more than 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, and has successfully implemented 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 only 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 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 has 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).

Variant partner Jesse Walden highly recognized Hyperbolic: "Hyperbolic is the first company we have seen that truly solves the problem of 'trust cost' in decentralized GPU networks while maintaining high levels of performance, quality and user experience."

Hyperbolic is in a leading position in financing in the field of Web3 AI, which fully proves that its technical strength and product feasibility have been favored and trusted by "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 two years at an astonishing speed.

Before founding Hyperbolic, he worked as a quantitative researcher at Citadel Securities and 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 was 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 figures.

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 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 AI at the University of Hong Kong, and 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 competitive optimization solutions for enterprises' spending on expensive API calls and high-cost machine rentals.

On the premise of 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 rental agreements, Hyperbolic has 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

Hyperbolic provides a rich GPU option for developers who are unable to advance their project testing due to limited GPU resources, and the price is only a fraction 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 Centers

Hyperbolic provides a platform for data centers that have not achieved the expected return on investment for existing resources or hope to break through the traditional book value limitations 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 and transform them into high-quality assets that continuously 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 and image generation, voice reading, etc.

In the future, Hyperbolic will also build AI agents on Base for users to use. You can stay tuned.

Hyperbolic webpage:

app.hyperbolic.xyz?utm_source=x&utm_campaign=seriesA&utm_content=biteye

08 Summary

Hyperbolic provides a GPU market, inference services, and the gold standard verification protocol of sampling proof, setting a new benchmark for reliable and high-performance AI for Web3 by maximizing GPU performance, higher-precision models, and secure and economical 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.