Computing platform shifts tend to evolve in couplets or triplets, often featuring new hardware, applications, and distribution innovations that propel one another forward. For example: the PC, the internet, and the web. Mobile, social, and the cloud.
We think a new and powerful couplet is emerging at the intersection of crypto(graphy) and AI. AI requires the massive coordination of GPUs and crypto uses incentives to herd resources. AI is probabilistic and crypto is deterministic. We believe crypto can solve two of the most pressing problems for AI: cost and trust (and specifically, cost of trust).
Let's unpack that:
Cost
It is currently extremely expensive to run models. The root cause is usually framed as a supply problem, a shortage in GPUs resulting from hoarding by the largest tech companies. But this isn’t really the case; GPUs are plentiful in data centers, mining farms, personal machines, and on-prem machines around the world. Rather, GPUs appear scarce because their supply is dispersed and uncoordinated. So what we really have is a coordination problem amongst a decentralized network of GPU suppliers, which makes them expensive.
Trust
A decentralized network of GPUs has a lower raw cost but introduces a new problem: the cost of trust. How can you trust that the model being run by a disparate network of participants is being run correctly? Crypto’s answer is traditionally to introduce significant overhead by having each node perform the same computation or to reduce the computational burden entirely.
But for AI models, this doesn’t work because having each node perform the same computation is too slow and reducing the size of models also reduces quality. Make no mistake, the verification problem exists in the centralized context as well (e.g., how do you know that ChatGPT is providing you GPT-4o versus GPT 3.5?), but OpenAI’s reputation can underwrite the trust more cheaply, albeit with no cryptographic rigor. When asking for a cookie recipe this verifiability may not matter, but when asking whether a malignant tumor is present in medical imaging, it certainly does. As AI progressively takes on more important work in society, the cost of trust will only increase. Crypto networks are ahead of the curve on this because they have to solve the verification problem in order to reduce costs.
Introducing Hyperbolic.
Hyperbolic is the first player we’ve seen with a solution to the cost-of-trust problem in decentralized GPU networks. One of the team’s key innovations that makes this possible is sampling-based verifiable machine learning (spML). It uses a random sampling protocol called Proof of Sampling to guarantee verifiability (assuming parties are acting economically rationally) amongst a decentralized network of GPU suppliers while maintaining the efficiency required to run the largest, highest-quality AI models out there. Hyperbolic makes it cheaper to verifiably run models without sacrificing performance or quality.
Early traction supports this. Hyperbolic is one of the only platforms to host Llama 3.1 405B’s base model in BF16 format, a massive open-source model with quality on par with OpenAI’s proprietary GPT-4o model—but running Llama 3.01 405B on Hyperbolic is 10x cheaper than paying to use OpenAI’s GPT-4o model. Integrations with leading AI platforms like Hugging Face’s Gradio, OpenRouter, and Quora’s Poe highlight Hyperbolic’s commitment to bringing the highest-quality models to the AI community. Well-known AI developers like Andrej Karpathy have used Hyperbolic to run open-source models because it is able to run higher-quality models, is cheaper, and has a better UX than competitors’ offerings.
But Hyperbolic is more than a formidable Web2 competitor and will be unmatched in its ability to service demand from Web3 apps. Web3 apps are currently forced to enter into a Faustian bargain when integrating AI: to get the performance they need, they must rely on centralized sources of AI inference, which is directly at odds with the decentralized ethos of the project and reintroduces the oracle problem. Because Hyperbolic will provide decentralization and performance and quality, Web3 apps will be able to utilize it without sacrificing one for the other.
We believe the team’s focus on first building a product that is competitive for all users (not just those in Web3) is the right approach. GPU supply is mercenary and will follow demand with little friction, so it is crucial to first attract demand to build the requisite stickiness. We expect that the demand from inference will sustainably attract GPU supply and achieve the economies of scale necessary to compete in the market long-term. An imperfect analogy to demonstrate this is Amazon’s approach to AWS. Amazon first focused on building demand for compute through products users loved (e.g., its marketplace), provided the compute supply necessary to support that demand, and eventually hit economies of scale such that it could launch AWS and offer that compute to third parties cheaper and better than competitors. After the core demand and supply are established on the Hyperbolic network, we believe the team will be in a strong position to expand across all layers of the AI stack, including training, data sourcing, and preprocessing.
The Hyperbolic founders are the strongest we’ve encountered in the space and have deep expertise in both crypto and AI that makes them uniquely positioned to tackle decentralized compute marketplaces for AI models. On the crypto side, CEO/co-founder Jasper Zhang is a math expert with expertise in proof verification in distributed systems. Jasper has won multiple Math Olympiads, earned a PhD in math from Berkeley in under two years (becoming the fastest person in the history of the institution to complete this five-year PhD), and was previously a quant at Citadel and a researcher at Ava Labs. On the AI side, CTO/co-founder Yuchen Jin is a machine learning and distributed systems expert. Yuchen received the prestigious China National Scholarship, has a PhD in CS from the University of Washington, and managed a team of engineers at OctoAI that built optimization solutions for AI models.
We’re excited to announce today that we have led Hyperbolic’s Series A round. We couldn’t be more excited to support Jasper, Yuchen, and the rest of the Hyperbolic team in their journey to make AI more accessible, verifiable, and open.
Disclaimer: This post is for general information purposes only. It does not constitute investment advice or a recommendation or solicitation to buy or sell any investment and should not be used in the evaluation of the merits of making any investment decision. It should not be relied upon for accounting, legal or tax advice or investment recommendations. You should consult your own advisers as to legal, business, tax, and other related matters concerning any investment. Certain information contained in here has been obtained from third-party sources, including from portfolio companies of funds managed by Variant. While taken from sources believed to be reliable, Variant has not independently verified such information. Variant makes no representations about the enduring accuracy of the information or its appropriateness for a given situation. This post reflects the current opinions of the authors and is not made on behalf of Variant or its Clients and does not necessarily reflect the opinions of Variant, its General Partners, its affiliates, advisors or individuals associated with Variant. The opinions reflected herein are subject to change without being updated.
