Novelty Search April 24, 2025

Okay, here are the detailed, narrative-driven show notes for the podcast transcript, tailored for Crypto AI investors and researchers, following all specified guidelines.

Show Notes: Celium Compute - Forging a Permissionless GPU Market on BitTensor

Episode: Novelty Search (April 24, 2025)
Guest: Fish (Founder, Celium Compute & Dura)
Host: Mars

1️⃣ Episode Introduction

This Novelty Search episode from April 24, 2025, unpacks Celium Compute's rapid six-month evolution on BitTensor, detailing their battle against miner exploits to forge a stable, permissionless GPU-as-a-Service platform aiming to disrupt the cloud compute market.

2️⃣ Structured and Narrative Show Notes

GPU Market Context & Celium's Rationale

• Fish frames the discussion by highlighting the massive scale and growth of the GPU market, projected to exceed 10x growth within eight years.
• He notes the immense investments by tech giants (XAI, Meta, Grock) acquiring hundreds of thousands of GPUs, underpinning Nvidia's multi-trillion dollar valuation and the pervasive integration of AI (like LLMs - Large Language Models, used for generating human-like text) into daily workflows.
• However, Fish points out the prohibitive cost for smaller players to own GPU infrastructure (data centers, power, maintenance), citing analysis showing variable costs alone matching 5-year rental costs for H100s.
• Celium aims to bridge this gap, offering accessible, rentable GPU power via BitTensor's Subnet 51 for users without billions to invest.

Celium's Origin: From Miner Pain Points to Platform Vision

• Fish shares his background as an early, large-scale BitTensor miner, primarily focused on LLM inferencing (running trained AI models to generate outputs like predicting the next word/token).
• He describes the inefficiency of managing multiple GPU providers, each with different terms, costs, and requiring lengthy sales negotiations.
• "I was having to bounce across multiple different providers with each provider having their different benefits, their pros and cons... and I was having to spend a bunch of calls uh time on sales calls trying to negotiate different rates." - Fish
• This experience, combined with his BitTensor knowledge, sparked the idea for Celium: a unified platform incentivizing high-quality, in-demand GPUs from BitTensor miners to provide a valuable compute service.

Overcoming Miner Exploits: The Battle for Security & Stability

• Fish candidly discusses the significant challenges faced in Celium's first six months, describing miners as "pretty brutal" in finding ways to illegitimately claim rewards on Subnet 51.
• Initial GPU verification relied on checking Nvidia libraries, which miners bypassed using hooks to misreport hardware (e.g., claiming H100s when using 4090s).
• Celium evolved its verification to performance-based checks, similar to Subnet 64, using matrix multiplications to measure actual speed (Flops - Floating-Point Operations Per Second, a measure of compute speed) and capacity (VRAM), allowing accurate GPU identification regardless of miner claims.
• Other exploits included running multiple mining containers on one GPU (countered by checking Nvidia UUIDs and monitoring all containers on the machine) and proxying tasks to other machines (countered using SSH interactive shells that prevent proxying).
• Fish acknowledges past instability but emphasizes continuous, rapid improvements have led to a much more robust and reliable platform today. Strategic Insight: Decentralized physical infrastructure networks (DePIN) face unique security challenges; Celium's iterative approach to exploit mitigation is crucial for building trust and usability.

Dynamic Incentives & Feature Enhancements

• Celium implemented dynamic GPU incentives based on rental demand. When flooded with 4090s but short on rented-out H200s/B200s, they programmatically increased incentives for the latter and decreased for the former.
• This leverages BitTensor's core strength: miners adapt rapidly (within hours) to incentive changes, helping Celium balance supply and demand for different GPU types.
• A major recent breakthrough is Docker-in-Docker support (using sysbox for security without privileged access), allowing users to run containerized applications (like Subtensor) within their rented Celium container—a capability Fish notes competitors like Vast and RunPod lack. Docker is a platform for packaging and running applications in isolated environments called containers.
• Other quality-of-life improvements include remote reboot, UI-based SSH key management, and verified custom template creation (ensuring templates include necessary components like SSH servers).

API, Tooling, and Performance Metrics

• The entire Celium workflow is accessible programmatically via an API (Application Programming Interface - a way for software components to communicate).
• Fish shouts out Distributed Tensor for creating a Celium CLI (Command-Line Interface - a text-based way to interact with software). Celium is also developing its own advanced tooling, including a Kubernetes-like configuration system for automated GPU provisioning based on user specs (e.g., VRAM requirements).
• Current performance shows significant traction: ~$7,000/day in rentals, 500 unique users, and 124% month-over-month growth, achieved within the first six months. Actionable Insight: Rapid growth metrics suggest market validation, but investors should watch if stability supports continued scaling.

Competitive Advantage: Permissionless Onboarding & Cost

• Fish identifies Celium's core advantage as its fully automated, permissionless onboarding for GPU providers: no KYC (Know Your Customer) checks, no contracts, no vendor lock-in.
• This contrasts sharply with competitors requiring lengthy agreements, hardware lock-ins (up to a year or more), or stringent operational requirements (detailed below).
• By accessing a global, unrestricted pool of GPU providers (including those in jurisdictions potentially excluded by competitors), Celium can offer significantly lower prices—claiming to be roughly half the cost of alternatives while striving for comparable quality. Strategic Insight: Permissionless access is Celium's key lever for potentially disrupting incumbents on price and supply diversity.

Competitor Terms vs. Celium's Approach

• Fish highlights specific terms from competitors' agreements:
  • Vast: Can charge providers for taxes, withhold payments arbitrarily, and performs infrequent (weekly) machine verification.
  • RunPod: Requires providers to have 24/7 on-site security, track physical access, submit logs, and meet comprehensive setup criteria.
• Celium avoids these barriers; verification is automated and fast (minutes), payments are on-chain and transparent, and the only requirement is suitable hardware. Providers also earn baseline rewards (currently ~20% of potential emissions) just for availability, plus a share of rental revenue paid in the Subnet 51 token.

Platform Features & User Experience

• Celium offers a wide variety of GPUs and is expanding into one-click custom deployments like image/video generators (demoed during the call).
• Users can create and share their own private or public custom templates via DockerHub.
• The platform provides detailed metrics dashboards (GPU usage, utilization), SSH key management via UI, and flexible payment options (crypto, fiat via Stripe with multi-currency support).
• Upcoming features include public uptime/reliability trackers for individual provider GPUs, building on an existing public database of rental history per miner coldkey.

Future Roadmap & Confidential Compute

• A high priority is integrating TEE (Trusted Execution Environment) and Confidential Compute capabilities. Fish defines this as enabling workloads to run encrypted on a GPU, preventing the hardware owner from accessing the data.
• He mentions discussions with FollowCloud (a TEE service provider) to implement this.
• Strategic Insight: Confidential Compute is critical for attracting users with sensitive data or workloads (e.g., validators needing to protect hotkeys), significantly expanding Celium's addressable market and building trust.

Q&A Highlights

• Stability Source: Continuous iteration and creative solutions to counter miner exploits were key to recent stability gains.
• Storage Verification: Miners are checked by deploying random official templates (10-20GB) and potentially larger user-requested images (e.g., 100GB LLMs); failure due to insufficient storage resets their score.
• Proxy Solution: Primarily addressed via interactive SSH shells preventing proxying, alongside performance/matrix checks verifying the GPU directly.
• Cost Fairness: Rental prices are market-driven; emissions primarily incentivize availability, with only 20% currently active to ensure sustainable growth rather than unused oversupply. Rental revenue is shared back with miners.
• Scaling: Celium could attract thousands more high-end GPUs by increasing incentives but is deliberately focusing on platform features and stability first.
• Token Value (SN51): Value accrues to token holders and validators through direct, stake-weighted access to the underlying GPU hardware, not direct revenue share. Validators get "free" GPU access proportional to their stake. Fish suggests this utility could eventually make validator emissions unnecessary.
• Open Source / Direct Access: Validators can already bypass the Celium front-end and access allocated miner GPUs directly via SSH using their validator keys. The subnet is permissionless.
• Current Users: Primarily within the BitTensor ecosystem (e.g., miners on other subnets like SN19). Celium is exploring proof-of-concepts to directly serve these users (e.g., one-click SN19 miner templates).
• Dura Background: Fish created Dura initially to provide credibility for hiring developers while scaling his early BitTensor mining operations. Dura now focuses broadly on adding value to the BitTensor ecosystem (tools like TaoshiScan/Town Market Cap, subnet advising, validation, mining), though Fish is now concentrating his efforts on Subnet 51.
• Emission Burning: Fish advocates for adjusting/burning miner emissions (and potentially validator/owner emissions) based on actual value creation and demand, arguing against full emissions if the network isn't fully utilized or stable, promoting long-term sustainability.
• Weight Copying: Fish argues it's not a significant issue on SN51 because the primary incentive for honest validation is the valuable, stake-weighted access to GPUs. Weight-copying validators don't offer this utility and thus don't attract miner delegations, effectively marginalizing them.
• Docker Security: Celium uses sysbox to enable Docker-in-Docker functionality without granting privileged root access, mitigating major security concerns for GPU providers.

3️⃣ Reflective and Strategic Conclusion

Celium leverages BitTensor's dynamic incentives for a permissionless, cost-competitive GPU market, rapidly iterating despite miner exploits. Investors and researchers should monitor stability, feature maturity (especially TEE/Confidential Compute), and hardware diversity as key indicators of its disruptive potential in the decentralized compute landscape.