Hyperbolic

  Executive Briefing

Hyperbolic is a San Francisco-based AI infrastructure company founded in late 2022 with a deceptively simple thesis: open-source model weights are only meaningful if compute is affordable enough to actually run them. The company operates two interlocking products — a serverless inference API and a decentralized GPU marketplace — both positioned at prices it claims are 3–10x below traditional cloud providers. By aggregating underutilized GPUs from data centers, mining farms, and even personal machines worldwide, rather than building or leasing its own data centers, Hyperbolic bets that supply-side diversity can structurally undercut hyperscaler pricing without sacrificing reliability.¹

The company is led by its two co-founders. Jasper Zhang (CEO) holds a mathematics Ph.D. from UC Berkeley completed in under two years, is a multi-time national olympiad gold medalist (Chinese Mathematical Olympiad, Russian Mathematical Olympiad, Alibaba Global Math Competition), and previously worked as a quantitative researcher at Citadel Securities and as a senior blockchain researcher at Ava Labs. Yuchen Jin (CTO) holds a computer science Ph.D. from the University of Washington under Professor Arvind Krishnamurthy, with prior roles at ByteDance AI Lab, Microsoft Research Azure, and OctoAI — where he led the Relax project within the Apache TVM compiler stack. The combination of theoretical mathematics, quantitative finance, distributed systems, and compiler-level ML optimization shapes how Hyperbolic approaches the problem of cheap, verifiable inference at scale.²³

By mid-2025 the platform had crossed 200,000 registered developers and reported processing more than one billion tokens per day, with month-over-month growth rates at times exceeding 300 percent.⁴ Official integrations with Hugging Face, Vercel, Poe by Quora, and OpenRouter, alongside academic pricing programs at UC Berkeley, Stanford, Cornell, and NYU, signal a deliberate positioning as the infrastructure layer for the open-source AI ecosystem rather than a competitor to enterprise hyperscalers. Hyperbolic's most technically distinctive contribution is its Proof of Sampling (PoSP) cryptographic verification protocol, co-developed with UC Berkeley and Columbia University researchers, which provides game-theoretically enforced honesty guarantees across decentralized inference nodes — an integrity layer that centralized cloud competitors do not offer.⁵

Having raised a reported total of approximately $20 million across a pre-seed, a $7 million seed round led by Polychain Capital and Lightspeed Faction, and a $12 million Series A led by Variant Fund and Polychain Capital, Hyperbolic remains small relative to peers like Together AI or CoreWeave but punches above its weight in developer mindshare and open-source ecosystem alignment.⁶⁷ Its roadmap points toward premium enterprise tiers, a proprietary coordination blockchain, and a GPU futures exchange — a trajectory from consumer inference API toward full-stack decentralized AI compute fabric.

  At a Glance

  Origins & Founding

Hyperbolic takes its name directly from CEO Jasper Zhang's doctoral dissertation, "Guts, Dehn Fillings and Volumes of Hyperbolic Manifolds," a signal that the company's identity is inseparable from its founders' academic roots.² Zhang's founding observation was expressed as a clean paradox: "open source does not mean open access." Publishing model weights achieves little if the compute required to run them costs $10,000 per month at prevailing hyperscaler rates — pricing that eliminates independent researchers, early-stage startups, and most academic teams entirely. He identified two underexploited data points: personal computing devices sit idle for more than 19 hours per day on average, and GPU utilization within commercial data centers frequently falls below 50 percent. The gap between supply (stranded compute globally) and demand (developers who cannot afford hyperscaler rates) was, in Zhang's framing, a coordination failure solvable by software.²

Zhang co-founded Hyperbolic with Yuchen Jin (CTO), whose compiler and distributed-systems background provided the technical counterpart to Zhang's mathematical and financial perspective. Jin had spent years optimizing ML inference at ByteDance and contributed to Apache TVM through OctoAI's Relax project — experience directly applicable to squeezing performance out of heterogeneous, decentralized hardware. The company's founding thesis extended beyond mere price arbitrage: by building cryptographic verification into the inference layer through PoSP, Hyperbolic aimed to make decentralized compute trustworthy enough for production use cases, not just hobbyist experimentation.³

The initial product was an inference service hosting popular open-source models at a fraction of the cost charged by AWS, Azure, and Google Cloud. The longer-term plan was always a full GPU marketplace enabling training workloads as well — a two-sided platform that aggregates supply from GPU owners and connects it to developers who need compute on demand, governed by Hyperbolic's proprietary Hyper-dOS distributed operating system to provide reliability atop inherently unreliable decentralized hardware.⁹

  History & Timeline

    2022–2023: Founding and Pre-Seed

Hyperbolic was established in late 2022 by Zhang and Jin. The company raised an initial $725,000 pre-seed round from Chapter One and Samsung Next to fund early product development and infrastructure. The founding period focused on architecting Hyper-dOS, establishing the decentralized node network, and launching an initial serverless inference offering with a small selection of open-source models.¹

    2024: Seed Round, Marketplace Launch, and PoSP

The year 2024 marked Hyperbolic's first public breakout. On 30 July 2024, the company announced a $7 million seed round led by Polychain Capital and Lightspeed Faction, with participation from Chapter One, LongHash, Bankless Ventures, Republic Digital, Nomad Capital, CoinSummer Labs, and Third Earth Capital, alongside notable angel investors including Balaji Srinivasan and Illia Polosukhin (NEAR Protocol founder) and Sandeep Nailwal (Polygon co-founder). The round received coverage in CoinDesk, reflecting the crypto-native dimension of Hyperbolic's investor base.⁶

During 2024's second half, Hyperbolic launched the GPU Marketplace in alpha, reporting 10,000 GPUs rented to 1,000 users and 33,000 GPU hours processed. The inference service reached 98,000 unique users and processed more than 59 million inference requests across a catalog of 38 open-source models. The Proof of Sampling (PoSP) cryptographic verification protocol was introduced and subsequently adopted by EigenLayer and Karak Network as an active verification service (AVS).⁵¹⁰ Partnerships with Poe by Quora (model access), NEAR Protocol, Virtuals Protocol, Black Forest Labs (FLUX.1), Nous Research (Hermes 3), and Hugging Face Gradio were established across this period.

On 10 December 2024, Hyperbolic announced a $12 million Series A led by Variant Fund and Polychain Capital, with co-investment from Chapter One, Lightspeed Faction, Bankless Ventures, IOSG, Vertex, GSR, Wintermute Ventures, Blockchain Builders Fund, Alumni Ventures, and Ambush. Additional angels included Sreeram Kannan (EigenLayer), Devin Walsh (Uniswap Foundation), Ethan Sun (MyShell), Ying Sheng and Lianmin Zheng (LMSYS/Chatbot Arena), Alex Atallah (OpenRouter), Nicola Greco (Protocol Labs), and Thomas Scott (ex-Worldcoin). The round brought total reported funding to approximately $20 million.⁷

    2025: Scale, Ecosystem Integrations, and Product Expansion

February 2025 saw Hyperbolic added as an official Hugging Face serverless inference provider, joining Nebius and Novita — a distribution milestone that placed Hyperbolic's endpoints directly within the world's largest open-model hosting platform.¹¹

In March 2025, Hyperbolic reported being the first inference provider to support DeepSeek-V3-0324 on Hugging Face, while GPU hours processed grew 2.6x month-over-month. The developer base reached 165,000 and Lex Fridman mentioned Hyperbolic alongside Together AI and Fireworks AI as representative cost-efficient inference options. By April 2025, the platform had 195,000+ registered developers processing 112 billion tokens per month — a 320 percent month-over-month increase. The company launched Auto Top-Ups for seamless balance management, Custom Port Configuration for GPU rentals, and released the @hyperbolic/ai-sdk-provider package for Vercel's AI SDK. Hyperbolic was announced as a launch partner for Coinbase's x402 crypto payment protocol.⁴¹²

May 2025 brought an MCP (Model Context Protocol) Server release enabling LLMs to orchestrate GPU compute via natural language, addition of DeepSeek-R1-0528, a student GPU credit program offering up to 75 percent discounts for PhD students and postdocs, a co-sponsorship of the Agents and MCP Hackathon with Gradio and Hugging Face (over $16,500 in prizes), and a domain rebrand from hyperbolic.xyz to hyperbolic.ai.¹³

By mid-2025, the platform crossed 200,000 developers and reported processing more than one billion tokens per day.⁴

  What They Offer — Products & Platform

Hyperbolic organizes its offering around four main product lines, all accessible through a unified dashboard and OpenAI-compatible API layer.

On-Demand GPU Clusters allow developers to rent individual GPUs or multi-GPU nodes by the hour with provisioning completing in under 60 seconds. The catalog spans NVIDIA H100 SXM, H100 PCIe, H200, A100 (PCIe and SXM4), and RTX 4090 hardware. There are no long-term commitments and no quota limits, distinguishing the offering from hyperscalers that impose capacity reservation requirements. AMD and Intel GPU support has been described as planned but was not confirmed as available as of mid-2026.⁹

Serverless Inference API provides an OpenAI-compatible HTTP endpoint for calling hosted open-source large language models and image-generation models on a per-token basis, with no GPU provisioning required. The service handles scaling automatically and supports rate limits of 60 to 600 requests per minute depending on account tier.

Reserved and Bulk Clusters offer dedicated GPU reservations with guaranteed 99.5 percent uptime SLA and discounted prepaid pricing for teams running production workloads or long-horizon training jobs.

Dedicated Endpoints provide single-tenant GPU hosting for bring-your-own-weights deployments or private instances of open models, delivering private endpoints for 24/7 high-throughput inference with isolation guarantees.

Platform-level features include: pre-configured Docker images for PyTorch, TensorFlow, and CUDA environments; the MCP Server for LLM-driven GPU orchestration; support for USDC crypto payments via Coinbase x402 as well as standard credit cards; Auto Top-Ups for continuous availability; and a Model Context Protocol integration that allows AI agents to spin up and manage compute programmatically. Business Cloud and Enterprise Cloud premium tiers with stronger SLA guarantees at pricing claimed to be up to 70 percent below legacy cloud rates had been announced but their formal launch status as of mid-2026 is unconfirmed.⁹¹²

  Technology & Infrastructure

Hyperbolic's infrastructure design rejects the owned-data-center model in favor of a software coordination layer atop a heterogeneous, globally distributed pool of GPUs. Supply comes from commercial data centers, mining farms transitioning from cryptocurrency, and private machines — a model analogous to Airbnb for compute rather than Marriott.

The platform's central software component is Hyper-dOS, Hyperbolic's proprietary distributed operating system. Hyper-dOS provides automatic load balancing, failure recovery, and multi-GPU clustering, enabling the platform to present reliable endpoints atop hardware that it does not own or control directly. Inference is served through vLLM and SGLang frameworks, which handle continuous batching, KV cache management, and efficient tensor parallelism across the GPU fleet.⁹

The most technically novel infrastructure element is the Proof of Sampling (PoSP) protocol. PoSP is a cryptographic verification mechanism co-developed with researchers at UC Berkeley and Columbia University. It creates a game-theoretic Nash Equilibrium in which rational node operators find it unprofitable to falsify inference outputs, providing integrity guarantees without requiring trusted hardware enclaves or centralized auditing. Critically, PoSP adds less than one percent computational overhead to each inference request — making verifiable compute practical rather than prohibitively expensive. The protocol has been adopted by EigenLayer and Karak Network as an active verification service, extending its reach beyond Hyperbolic's own infrastructure.⁵

GPU provisioning completes in under one minute from request to usable instance. At peak the network has been reported to process over one billion tokens per day and 112 billion tokens per month, with the underlying GPU node count not publicly disclosed.⁴

  Model Catalog & Performance

Hyperbolic's serverless inference catalog spans large language models across several open-weight families as well as text-to-image models. The LLM offering is centered on Meta's Llama family, with the platform notably serving Llama 3.1 405B in both BF16 and FP8 precision — it has claimed to be the only platform offering the base (non-instruct) variant in BF16.¹ Other Llama models served include Llama 3.3 70B Instruct and Llama 3.1 8B Instruct.

DeepSeek models feature prominently, including DeepSeek-R1, DeepSeek-V3-0324, and DeepSeek-R1-0528 — with Hyperbolic being the first Hugging Face inference provider to support DeepSeek-V3-0324 at the time of its release.¹² The Qwen series from Alibaba is well-represented, including QwQ-32B for reasoning tasks and multiple Qwen 2.5 variants. NousResearch/Hermes-3 and Mistral variants round out the instruct-tuned LLM catalog.

For image generation, Hyperbolic serves Black Forest Labs' FLUX.1 and FLUX.2 models, Stability AI's SDXL, and multiple Stable Diffusion variants. As of end-2024 the catalog stood at 38 open-source models and has continued to expand.

Performance benchmarks from Artificial Analysis position Llama 3.3 70B on Hyperbolic at approximately 62 tokens per second output throughput with a time-to-first-token of roughly 2.15 seconds — meaningfully slower than Groq's LPU-powered inference (which reaches 400–800 tokens per second on comparable model sizes) but competitive with other GPU-cluster-based inference providers such as Together AI and Fireworks AI.¹⁴

  Pricing & Performance Position

Hyperbolic's central value proposition is price. Serverless inference for Llama 3.1 8B Instruct is listed at $0.10 per million input tokens, and Llama 3.3 70B Instruct at approximately $0.40 per million tokens (blended rate), with the catalog ranging from roughly $0.10 to $4.00 per million input tokens. The company claims 3–10x lower inference pricing than leading competitors, though exact comparisons shift as rivals adjust their rates.¹

On the GPU rental side, pricing is refreshed weekly based on marketplace supply-and-demand dynamics. Reported rates (which may have changed) include: RTX 4090 at approximately $0.30–0.35 per GPU-hour; A100 at approximately $1.80 per GPU-hour; H100 SXM at approximately $1.49–1.50 per GPU-hour. Spot GPU rentals are advertised starting at $0.16 per GPU-hour. Hyperbolic claims H100 SXM pricing approximately 9x cheaper than AWS and H200 pricing approximately 4.4x cheaper — figures that reflect the decentralized supply model reducing the premium typically embedded in hyperscaler margin and data-center overhead.⁹

For academic users, Hyperbolic offers GPU compute credits with discounts of up to 75 percent for PhD students and postdoctoral researchers, supporting its open-access positioning. The May 2025 student credit program formalized this.¹³

Pricing is marketplace-dynamic and subject to weekly change. Figures are reported estimates only.⁹

  People & Leadership

Jasper Zhang (Co-Founder and CEO) brings an unusual profile to infrastructure entrepreneurship. His mathematics doctorate from UC Berkeley, completed in an accelerated timeline, reflects the kind of structural problem-solving that informs how Hyperbolic frames the GPU utilization puzzle. His competition pedigree — gold medals at the Chinese Mathematical Olympiad, Russian Mathematical Olympiad, and Alibaba Global Math Competition — and his quantitative finance experience at Citadel Securities add rigor to business model design. His prior blockchain research role at Ava Labs connects to the crypto-native funding and payment strategy.²

Yuchen Jin (Co-Founder and CTO) brings the systems engineering depth. His CS doctorate at University of Washington under Professor Arvind Krishnamurthy placed him at the intersection of distributed systems and ML. At ByteDance AI Lab he worked on large-scale model serving; at Microsoft Research he contributed to Azure-scale infrastructure; and at OctoAI he led the Apache TVM Relax project, a compiler-level framework for optimizing model deployment across hardware backends. This background directly informs Hyper-dOS's design and the choice of vLLM and SGLang as inference serving frameworks.³

The company's advisor board includes several prominent figures. Ciamac C. Moallemi, a Columbia University professor, provides guidance on GPU exchange design and tokenization. Reynold Xin, co-founder of Databricks, advises on large-scale distributed systems. Raluca Ada Popa, a UC Berkeley professor specializing in privacy and secure computation, informs the cryptographic integrity work. Yi Ma, affiliated with HKU and UC Berkeley, contributes perspective on AI model transparency. Employee headcount has not been publicly reported.

  Funding, Ownership & Business

Hyperbolic has raised approximately $20 million across three rounds, maintaining independence with no corporate parent.⁶⁷

The pre-seed round of $725,000 came from Chapter One and Samsung Next. The $7 million seed (announced 30 July 2024) was led by Polychain Capital and Lightspeed Faction, with a notably crypto-native investor syndicate that included LongHash, Bankless Ventures, Republic Digital, and Nomad Capital alongside angels Balaji Srinivasan, Illia Polosukhin, and Sandeep Nailwal.⁶ The $12 million Series A (announced 10 December 2024) was led by Variant Fund and Polychain Capital, co-invested by Chapter One, Lightspeed Faction, Bankless Ventures, IOSG, Vertex, GSR, Wintermute Ventures, Blockchain Builders Fund, Alumni Ventures, and Ambush.⁷ Post-Series A valuation has not been publicly disclosed.

The business model combines two revenue streams: GPU rental revenue (usage-based hourly billing on the marketplace) and inference revenue (per-token billing on the serverless API). Both streams benefit from the decentralized supply model, which lowers Hyperbolic's own capital expenditure relative to companies that must own or lease data center capacity. Planned future revenue channels include tokenized AI model revenue-sharing and a GPU futures exchange, both contingent on a forthcoming proprietary Hyperbolic blockchain whose launch timeline as of mid-2026 remains unconfirmed.⁷

The crypto-oriented investor base is not incidental: it reflects the company's stated long-term architecture as a decentralized coordination layer for compute, with economic incentives enforced on-chain rather than through traditional contractual arrangements.

  Customers & Partnerships

Hyperbolic's customer base spans individual developers, academic researchers, and a growing set of named commercial partners. The 200,000+ registered developers figure (mid-2025) includes integration users from platform partnerships rather than purely direct sign-ups.⁴

Notable ecosystem partnerships include:

Hugging Face integrated Hyperbolic as an official serverless inference provider in February 2025, placing Hyperbolic endpoints alongside Nebius and Novita within the Hugging Face Inference Providers program — giving millions of Hugging Face users direct access to Hyperbolic's API with a single authentication token.¹¹

Vercel partnered with Hyperbolic through the AI Accelerator program, resulting in the @hyperbolic/ai-sdk-provider npm package that lets Next.js and Vercel AI SDK users route inference through Hyperbolic with minimal configuration.¹²

Poe by Quora integrated Hyperbolic to serve Llama 3.1 and NousResearch Hermes model access to its users.

Black Forest Labs and Nous Research established inference partnerships, with Hyperbolic serving FLUX.1 and Hermes 3 respectively.

EigenLayer and Karak Network adopted Hyperbolic's Proof of Sampling protocol as an active verification service, embedding the cryptographic integrity layer into their broader decentralized infrastructure stacks.⁵

Coinbase selected Hyperbolic as a launch partner for the x402 crypto payment protocol, enabling USDC-denominated GPU rental and inference payments.¹²

Chatbot Arena / LMSYS uses Hyperbolic infrastructure, reflected in the angel participation of LMSYS co-founders Ying Sheng and Lianmin Zheng in the Series A.

Academic institutions receiving GPU compute credits or partnerships include UC Berkeley, Stanford, Cornell, and NYU, reinforcing the open-access positioning among researchers who are disproportionate early adopters of new open-source models.

  Competitive Position

Hyperbolic competes across two overlapping markets simultaneously, which is both a strength and a complexity.

In serverless inference APIs, the primary competitors are Together AI, Fireworks AI, Groq, DeepInfra, Nebius AI Studio, Novita, and Replicate, as well as inference endpoints offered directly by hyperscalers. Hyperbolic's position in this market is lowest-cost pricing with OpenAI API compatibility and a broad open-model catalog. Its primary weakness relative to Groq is raw throughput: at roughly 62 tokens per second for a 70B model, Hyperbolic is 6–13x slower than Groq's LPU-accelerated inference, making it unsuitable for latency-sensitive applications. Relative to Together AI, the model catalog remains smaller and enterprise tooling less mature.

In GPU rental marketplaces, the competitors are Vast.ai, RunPod, Lambda Labs, CoreWeave, and hyperscalers. Here Hyperbolic competes on price and GPU variety, with the added differentiation of Hyper-dOS reliability — which addresses the principal weakness of decentralized competitors like Vast.ai (no reliability guarantees). The claimed 9x price advantage over AWS for H100 SXM, if accurate, is structurally significant.⁹

Cross-cutting differentiators that apply to both markets include: the PoSP verifiable inference layer (unique among inference API providers), crypto-native payment support (relevant to the Web3-adjacent developer base), academic pricing programs (serving a community often ignored by commercial providers), and an explicit open-source ecosystem focus that aligns with Hugging Face's positioning. The MCP Server integration is an early move on the agentic-AI infrastructure trend.

  Notable Events

• 30 July 2024 — $7M seed round announced; CoinDesk covers the round under the headline "Polychain and Lightspeed Fund Blockchain-Based AI Platform from Math Olympian," cementing the crypto-AI positioning.⁶
• 10 December 2024 — $12M Series A announced; Fortune and The Block cover the round; total funding reaches approximately $20M.⁷
• 2024 — PoSP protocol introduced and adopted by EigenLayer and Karak Network as an active verification service.⁵
• February 2025 — Added as an official Hugging Face serverless inference provider.¹¹
• March 2025 — First inference provider to support DeepSeek-V3-0324 on Hugging Face; platform's developer count reaches 165,000; GPU hours processed grow 2.6x month-over-month.¹²
• April 2025 — 195,000+ developers; 112B+ tokens/month (up 320% month-over-month); Vercel SDK provider launched; Coinbase x402 partnership announced.⁴¹²
• May 2025 — MCP Server released; domain rebranded to hyperbolic.ai; student GPU credit program launched; co-sponsored Agents and MCP Hackathon with Hugging Face and Gradio.¹³
• Mid-2025 — 200,000+ developers; 1B+ tokens/day throughput reported.⁴

  Outlook & Roadmap

Hyperbolic's roadmap reflects an intent to move up-market while preserving the decentralized, open-access ethos that defines its brand. The most concrete near-term priorities involve Business Cloud and Enterprise Cloud premium tiers, which promise enterprise-grade SLAs at pricing positioned as up to 70 percent below hyperscaler equivalents — addressing the reliability concerns that currently limit Hyperbolic's penetration into regulated industries and production-scale enterprise workloads.⁹ Whether these tiers have formally launched as of mid-2026 is unconfirmed.

The longer-term vision centers on a proprietary Hyperbolic blockchain designed to coordinate decentralized GPU supply at scale, with on-chain incentives for node operators, tokenized AI model revenue-sharing, and a GPU futures exchange enabling compute capacity to be traded as a financial instrument. These ambitions were signaled at the time of the Series A but their development timeline and current status remain unconfirmed.⁷

On the inference side, expanded AMD and Intel GPU support has been described as planned since at least late 2024 but has not been confirmed as available. Planned platform integrations include Jupyter notebook environments and SkyPilot for ML workflow orchestration. The PoSP framework is expected to be generalized into a broader AVS applicable to other EigenLayer-based verification use cases beyond inference.

Strategically, the company frames its mission as building an "AI Rainforest" — a diverse, collaborative compute ecosystem rather than a winner-take-all monoculture. This framing is both philosophical positioning and a practical bet that open-source AI model development will continue to grow relative to closed model consumption, maintaining and expanding the addressable market for affordable open-model inference. The trajectory from early-stage API experiment to 200,000 developers and a billion tokens per day in under three years suggests the addressable market hypothesis is being validated, even as the enterprise revenue model remains unproven at scale.

---

  References

1. Hyperbolic — Platform overview and 2024 year in review
2. Meet Hyperbolic CEO Jasper Zhang — Medium / Hyperbolic Labs
3. Meet Hyperbolic CTO Yuchen Jin — Hyperbolic Blog
4. Monthly Recap April 2025 — Hyperbolic Blog
5. Proof of Sampling (PoSP) — Hyperbolic Blog
6. Polychain and Lightspeed Fund Blockchain-Based AI Platform from Math Olympian — CoinDesk, July 2024
7. AI Company Hyperbolic Raises $12M Series A — Fortune, December 2024
8. [Hyperbolic HQ discrepancy — platform lists San Francisco; Tracxn lists Irvine, CA; unresolved]
9. Deep Dive into Hyperbolic GPU Marketplace — Hyperbolic Blog
10. Variant and Polychain Capital Lead $12M Series A for Web3 AI Firm Hyperbolic — The Block, December 2024
11. Hyperbolic as Hugging Face Inference Provider — Hugging Face Docs
12. March 2025 Monthly Recap — Hyperbolic Blog
13. Monthly Recap May 2025 — Hyperbolic Blog
14. Hyperbolic Provider Benchmarks — Artificial Analysis

  References

1. Hyperbolic platform overview and open-access thesis — 2024 Year in Review. ↩ ↩² ↩³ ↩⁴

2. Jasper Zhang biography — Meet Hyperbolic CEO Jasper Zhang, Medium/Hyperbolic Labs. ↩ ↩² ↩³ ↩⁴

3. Yuchen Jin biography — Meet Hyperbolic CTO Yuchen Jin, Hyperbolic Blog. ↩ ↩² ↩³

4. Developer and throughput milestones — Monthly Recap April 2025; mid-2025 figures from Hyperbolic $20M total funding announcement. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹

5. Proof of Sampling (PoSP) protocol — Hyperbolic Blog. ↩ ↩² ↩³ ↩⁴ ↩⁵

6. $7M seed round — CoinDesk, 30 July 2024. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

7. $12M Series A — Fortune, 10 December 2024; The Block. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

8. HQ location discrepancy between hyperbolic.ai (San Francisco) and Tracxn (Irvine, CA) — unresolved. ↩

9. GPU Marketplace technical details and pricing — Deep Dive into Hyperbolic GPU Marketplace. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

10. EigenLayer and Karak Network PoSP adoption — Hyperbolic $20M announcement. ↩

11. Hugging Face inference provider integration — Hugging Face Docs. ↩ ↩² ↩³

12. March and April 2025 milestones — March 2025 Recap; April 2025 Recap. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

13. May 2025 product launches and rebrand — May 2025 Recap. ↩ ↩² ↩³

14. Throughput and latency benchmarks — Artificial Analysis: Hyperbolic. ↩