# **Polkadot Treasury Proposal: Polkadot Super AI Center**

#### Proposer: **Hashforest LLC (US)**
#### Funding Request: **$1,500,000 USD (375,000 DOT)**
#### Date: **2024.10.02**

---

## **1. Abstract**

This proposal seeks $1.5 million (375,000 DOT) in funding to create the **Polkadot Super AI Center**, an infrastructure project aimed at integrating AI into the Polkadot ecosystem. Operated by Hashforest LLC during the first year, the center will offer GPU-powered computing resources designed specifically for training and deploying decentralized AI models. The AI Center will serve Polkadot's ecosystem and the broader Web3 and Web2 AI developer community by providing access to computational power, thus driving the adoption of decentralized AI. After the first year, the Polkadot community will have full ownership, creating a lasting, decentralized infrastructure that aligns with Web3 principles.

---

## **2. Introduction**

### **Background:**
The use of artificial intelligence (AI) is accelerating rapidly across every industry, and AI-driven products are proliferating globally. However, centralized AI development faces several challenges, including expensive computational resources, limited access to specialized hardware (GPUs), and the difficulty in transitioning AI services to decentralized systems. This proposal aims to provide a solution to these challenges through a decentralized infrastructure built within the Polkadot ecosystem, one of the most scalable and versatile blockchain networks in the world.

Polkadot already hosts several key projects in decentralized computing and Web3 infrastructure that make it ideal for AI integration. Notably:
- **Phala Network** provides confidential computing solutions with trusted execution environments (TEE).
- **Bittensor** offers decentralized machine learning services built on Substrate.
- **OriginTrail** delivers decentralized knowledge graphs crucial for data-rich AI applications.

The Polkadot Super AI Center will leverage these existing capabilities and provide the hardware backbone necessary for large-scale AI training and deployment, allowing AI developers to build decentralized AI applications on Polkadot.

---

## **3. Rationale and Problem Statement**

### **Why Decentralized AI?**
In the near future, decentralized AI will become a critical part of Web3 infrastructures. AI models need powerful computing resources to train and deploy. The current AI landscape, however, is dominated by centralized services from a few large tech companies (e.g., Google, Microsoft, OpenAI). This centralized control results in restricted access, high costs, and vulnerability to single points of failure.

**Decentralized AI**, by contrast, offers:
- **Trustless Infrastructure**: Decentralization ensures that no single entity can control the data or computational power behind AI models.
- **Permissionless Access**: Developers and organizations can access compute resources and deploy AI models without reliance on centralized authorities.
- **Privacy and Confidentiality**: Confidential computing ensures that sensitive data used in AI models (e.g., medical or financial data) remains secure and private.
- **Community Ownership**: A decentralized model allows for community ownership and governance over critical AI infrastructure, aligning incentives and ensuring fair access.

### **Challenges for AI Builders:**
1. **Access to GPUs**: AI companies and developers need affordable and scalable access to GPUs for training AI models. The cost of GPUs, cloud services, and the infrastructure needed for large-scale AI projects is prohibitively high.

For instance:
- AI software costs can range from $7,200 to over $300,000 per year, depending on usage.
- A single NVIDIA H100 Tensor Core GPU can cost between $389,000 and $495,000.

2. **Complexity and Centralization**: Centralized AI platforms charge high fees and control the data pipelines, limiting innovation. In contrast, decentralized AI offers transparency and democratizes access to compute resources.

3. **Web2 Limitations**: Current AI solutions in the Web2 ecosystem rely on expensive, centralized cloud providers, leading to inefficiencies. The Polkadot Super AI Center aims to bypass these bottlenecks by leveraging decentralized infrastructure.

---

## **4. Ecosystem Fit and Strategic Importance**

### **Why Polkadot?**
Polkadot's existing infrastructure, combined with its innovative multi-chain architecture, makes it the ideal home for decentralized AI. Several features of the Polkadot ecosystem enable a strong foundation for the Polkadot Super AI Center:
- **Scalability**: Polkadot’s parachain model enables scalability by allowing multiple blockchains to operate independently but securely interoperate.
- **Security**: Polkadot’s shared security model ensures the center’s integrity without sacrificing decentralization.
- **Polkadot AI Builders**:
Several AI companies and projects are actively developing within the Polkadot ecosystem, leveraging its blockchain infrastructure to advance AI capabilities. Here are some of the key players:
- PolkaBot AI: PolkaBot AI is a recently introduced AI-powered chatbot for the Polkadot ecosystem. It's designed to assist users with inquiries related to Polkadot's Proof-of-Stake (PoS) system. Built on NeuroWebAI and OriginTrail's decentralized AI, PolkaBot AI offers insights and information about the Polkadot ecosystem.
- OriginTrail: A revolutionizing digital information management through its decentralized knowledge graph (DKG). It ensures data discoverability and verifiability, which is crucial for trusted AI solutions. OriginTrail proposes the Verifiable Internet for AI to address challenges like intellectual property rights mishandling and potential AI model collapses.
- Phala Network: Phala has introduced the AI-Agent Contract, which combines AI capabilities with blockchain technology. This platform allows AI agents to operate autonomously and generate value within the ecosystem. Phala's confidential computing framework ensures sensitive data processing without exposure, making it suitable for AI applications in finance, healthcare, and other sectors.
- Bittensor: Launched in 2021, Bittensor has become a significant player in the AI-blockchain intersection. It provides a decentralized, community-driven approach to AI development, countering the centralization of AI power among large corporations. Bittensor creates a decentralized AI marketplace across subnets, offering services like data storage, AI-powered chatbots, and pricing oracles.
- Fetch.ai: While not exclusively a Polkadot project, Fetch.ai has integrated with peaq, a network in the Polkadot ecosystem. This integration allows developers building on peaq and its canary network krest to leverage Fetch.ai's agents for automating and optimizing various business processes in the Economy of Things, such as helping drivers find available parking spots or charging stations.

The Polkadot Super AI Center complements the existing decentralized computing offerings of Phala and Bittensor while introducing critical infrastructure for large-scale AI model training and deployment.

### **Polkadot's Competitive Advantage in AI**:
- **GPU Leasing Market**: By creating a marketplace where AI developers can rent GPUs via decentralized credits or grants, the Polkadot ecosystem will stand as the go-to solution for decentralized AI needs.
- **Innovation and Research**: By supporting Web3 and AI companies with decentralized GPU power, Polkadot will position itself as a pioneer in Web3 AI research and development.

---

## **5. Goals and Objectives**

The primary goal of the **Polkadot Super AI Center** is to establish a decentralized hub where AI developers can train, deploy, and scale AI models using the Polkadot infrastructure.

### Key Objectives:
- Deploy a **GPU-powered AI center** consisting of **16 NVIDIA H100 Tensor Core GPUs** for decentralized AI training.
- Launch AI **model inference services** with limited-access gates, allowing developers to securely deploy and scale AI applications.
- Provide **GPU compute grants** to the Polkadot ecosystem, incentivizing Web3 AI development.
- Foster partnerships with top AI and Web3 companies to **co-develop AI research** and publish joint research papers.
- Onboard **20+ AI applications** within the first year of operation.

---

## **6. Customer Plans, Pricing, and KPIs**

The AI Center will offer its resources to various customer segments with the following pricing structures and KPIs:

| **Customer Segment** | **Plan** | **KPIs** |
|-----------------------------------|------------------------------------------|-------------------------------------|
| **Parity / Web3 Foundation** | Free ($50k credits/month) | No direct KPIs, internal use |
| **Polkadot Ecosystem Teams (30)** | Free ($30k credits/month) | Blog/social media mentions |
| **Web3 Partners (20)** | 50% market price ($5k credits/month) | Research papers, press releases |
| **AI Partners (20)** | 50% market price ($5k credits/month) | Co-research papers |
| **Public (20)** | 80% market price | 20+ paid companies onboarded |

---

## **7. Milestones and Deliverables**

### **Milestone 1 - Initial Setup (6 months)**
- **Objective**: Establish an AI training center with **16 NVIDIA H100 GPUs**.
- **Deliverables**:
- Purchase and installation of hardware (16 GPUs).
- Integration of popular open-source AI models such as Llama3.2 for training.
- Public dashboard for monitoring GPU usage and AI job allocation.

### **Milestone 2 - Model Inference Services (3 months)**
- **Objective**: Provide secure and gated access to AI model inference services.
- **Deliverables**:
- Set up secure, on-demand inference services for developers.
- Build a user-facing interface allowing Web3 and Web2 developers to upload and run AI models.

### **Milestone 3 - Onboarding and Partnerships (3 months)**
- **Objective**: Onboard at least **20 AI projects** and publish **5 collaborative research papers**.
- **Deliverables**:
- Collaborate with leading Web3/AI companies to onboard their projects.
- Publish 5 research papers highlighting the results of AI models trained on decentralized infrastructure.

### **Milestone 4 - Market Fit Evaluation (3 months)**
- **Objective**: Submit a report assessing the market fit of decentralized AI infrastructure in Polkadot.
- **Deliverables**:
- Detailed analysis on the adoption, performance, and business opportunities of decentralized AI on Polkadot.
- Recommendations for future expansion.

---

## **8. Governance, Monitoring, and Transparency**

### **Governance Model**:
The operation of the Polkadot Super AI Center will be overseen by a **Polkadot AI Committee**. This committee, composed of key stakeholders from the Polkadot ecosystem, including representatives from the Web3 Foundation, Hashforest LLC, and community members, will ensure that the center operates transparently and efficiently.

- **Monthly Reports**: Hashforest LLC will submit detailed monthly reports on GPU usage, operational efficiency, and grant allocations.
- **Community Governance**: After the first year, the Polkadot community will take ownership of the center through a governance token or another form of decentralized control.

---

## **9. Detailed Cost Breakdown**

### **Hardware Costs**:
- **2 NVIDIA DGX H100 systems**: $778,000 to $990,000 (8 GPUs each).
- **Additional servers and storage**: ~$50,000.
- **Networking equipment**: ~$50,000.

### **Infrastructure and Setup**:
- **Power and cooling**: $80,000 (power distribution units, UPS, cooling systems).
- **Facility rental (California, 1 year)**: $60,000.
- **Rack and cabling infrastructure**: $10,000.

### **Operational Costs (Annual)**:
- **Power consumption** (35 kW @ $0.18/kWh): ~$54,000.
- **Staffing**:
- **Full-time system administrator**: $120,000.
- **Part-time technician**: $60,000.
- **Maintenance and support contracts**: $50,000.
- **Internet and connectivity**: $24,000.
- **Miscellaneous costs (insurance, permits)**: $30,000.

### **Total Setup Cost**: $1,252,000 to $1,464,000.
### **Annual Operating Cost**: ~$308,000.
--

## **10. Why Hashforest LLC?**
Hashforest LLC is a US-based company with a strong track record in building decentralized infrastructures, particularly within the Polkadot ecosystem. Since 2019, Hashforest has been leading the development of Phala Network, a confidential computing platform on Polkadot, and has played a pivotal role in pioneering decentralized compute services within Web3.

The team's technical expertise and deep familiarity with the Polkadot tech stack, including Substrate and Rust, make Hashforest uniquely qualified to manage the Polkadot Super AI Center during its first year of operation. In particular, Hashforest has extensive experience in:
- Building and managing large-scale decentralized computing networks.
- Developing confidential compute technology using trusted execution environments (TEE), critical for AI use cases involving sensitive data.
- Providing AI services, including running AI models for various Web3 companies and integrating more than 150 AI models into decentralized APIs.

The company's track record, including creating the first Nvidia H100 TEE benchmark report, positions it as an ideal partner to manage the early phases of the Polkadot Super AI Center, ensuring both technical excellence and long-term viability.

---

## **10. Conclusion**

The **Polkadot Super AI Center** will serve as a key infrastructure component in the Web3 ecosystem, offering decentralized, community-owned AI compute power. This project not only aligns with Polkadot’s mission to bring decentralization to the forefront of technology but also positions the network as a leader in the AI space. By funding this proposal, Polkadot will secure a foothold in the decentralized AI landscape, fostering innovation, attracting top-tier AI projects, and enhancing DOT’s utility.

The proposed governance, transparency measures, and clear milestone roadmap ensure that the project remains accountable to the Polkadot community, delivering tangible value in both Web3 and AI spaces.

# Q&A List

### **Question 1: Has anyone gauged demand/interest within the ecosystem in buying these GPU credits? Otherwise, this would just benefit Hashforest.**

**Response:**

Yes, the demand within the ecosystem for GPU credits is both real and significant. A typical AI company spends between **60% to 80%** of its income on AI model inference costs. For example, if a company like **Bittensor** or any other Polkadot-based AI app generates **$10,000/month** in income, it is likely spending **$6,000 to $8,000** on GPU resources or services from providers like OpenAI or other centralized GPU centers to run open-source models.

This means that:
- **A)** If all the AI-related projects within the Polkadot ecosystem are spending over **$100,000 per month** on inference costs, it would take just **one year** to recover the initial investment of the Polkadot Super AI Center.
- **B)** If Polkadot projects collectively are not spending at least $100,000 per month on GPU costs, it indicates that "real AI use cases" are not fully developed within the ecosystem. In that case, providing access to more affordable and decentralized GPU resources can **massively increase demand** by lowering the cost barriers to entry for AI projects.

By reducing costs for AI projects, the Polkadot Super AI Center could significantly stimulate the development of AI applications in the ecosystem, making it more attractive for both existing and new AI developers to build on Polkadot. This creates a flywheel effect, where lower costs lead to more development, which in turn increases the usage of Polkadot's decentralized infrastructure. Thus, this proposal benefits not only Hashforest but the entire Polkadot community by addressing a crucial need for decentralized, affordable AI compute power.

Additionally, many AI companies and Web3 projects have already shown strong interest in accessing decentralized GPU resources, especially as a more affordable alternative to centralized services like OpenAI.

---

### Question 2: What happens if funding is abandoned after the first year? Will the DGX H100’s be sold with the returns going back to the DOT treasury?

Response:

In the unfortunate event that funding is abandoned after the first year, the community will still have full ownership of the infrastructure, as the GPUs and all associated assets (including the DGX H100 systems) will belong to the Polkadot community from the outset. This means that, in case of a funding shortfall or project termination, the community would have full control over the disposal of assets.

If the community deems it necessary to sell the hardware, any proceeds from such sales would be returned to the Polkadot Treasury. The primary goal of this proposal is to ensure long-term, decentralized community control over the AI Center, so every precaution will be taken to ensure that it remains a valuable, ongoing resource for Polkadot and the wider Web3 ecosystem.

Hashforest’s role in the first year is purely operational, acting as a steward to manage the infrastructure, with monthly reports submitted to the Polkadot AI Committee and the community. If funding or operations need to shift after that period, the transition will be handled transparently and with community involvement.

### **Question 3: How does the proposal benefit Hashforest? Are you just asking for money?**

**Response:**

No, this is not simply a request for money nor a profit-driven initiative. The **Polkadot Super AI Center** proposal is structured as a **non-profit project** that benefits the entire Polkadot ecosystem, not just Hashforest. Regardless of who operates this initiative, the hard costs associated with building and maintaining a GPU-powered AI infrastructure are substantial, and the majority of the funding goes directly toward these fixed expenses.

To clarify:
1. **Hardware Costs**: The largest portion of the requested funding is allocated to the purchase of **NVIDIA DGX H100** systems, which are currently priced between **$389,000 to $495,000** per unit. This pricing is consistent across the industry, regardless of the operator, and is transparent and verifiable by referencing the **NVIDIA price list** (which can be reviewed here: [[source](https://marketplace.uvation.com/nvidia-h100-tensor-core-gpu-80gb-pcie/?gad_source=1)]).
2. **Non-profit Structure**: This proposal is not designed to generate profit for Hashforest. In fact, approximately **98% to 100%** of the funding requested is dedicated to hard costs—primarily the **GPU hardware**, infrastructure setup, and operational expenses. Hashforest will not profit from this proposal; rather, our role is to manage the setup and operation of the center during the first year, ensuring the infrastructure is built and maintained properly.
3. **Risk of Loss**: Given the non-profit nature of this initiative, Hashforest bears financial risks, especially with **DOT price volatility**. If the value of DOT declines, Hashforest would actually face potential **losses**, as the budget is pegged to hard dollar costs for equipment, power, and other operational expenses. Our commitment to seeing this project through demonstrates that our interest lies in advancing Polkadot’s AI capabilities, not in generating revenue for ourselves.
4. **Long-term Community Ownership**: After the first year, the Polkadot community will have full ownership of the **Super AI Center**, which means any future revenues or operational decisions will be entirely governed by the community. Hashforest’s involvement is limited to the initial setup and operation, ensuring that the center functions as intended and benefits the broader Web3 ecosystem.

In summary, this proposal is about building infrastructure for the **Polkadot community**, and not about benefiting Hashforest financially. Whether it’s Hashforest or another entity, the costs remain the same due to the nature of the hardware and infrastructure involved.

### **Question 4: Seems like a classic "let's get as much money as we can now and figure it out later (or just plain wing it)" concept ?**

**Response:**

**Solid ROI on GPU Investment**
You’re right in noting that the proposal focuses on GPU investments as a key infrastructure need. Here’s why this is a reliable strategy:

- **Return on Investment**: The ROI for GPU-based AI infrastructure is well-established. In AI sectors, the returns on GPU utilization can range from **30% to 80% per year**, depending on model complexity and demand. This isn’t speculative; it's based on real-world use cases across industries.
- **Economic Impact**: What makes this proposal unique is that **all income generated by the AI Center will be used to buy back and burn DOT**, making this initiative inherently deflationary for the Polkadot tokenomics. This creates a long-term value proposition for DOT holders, as few other projects offer mechanisms for direct DOT buybacks linked to AI infrastructure revenue.

### **Question 5: There isn't even any evidence that teams are actually asking for this.**

**Response:**

1. **Existing Proven Market**
We are not creating a market out of thin air here. This proposal is about serving an **existing demand** for compute power:

- **AI Companies Already Pay for GPUs**: AI teams across Web3 and traditional sectors are already paying significant amounts for compute power—whether through **OpenAI, Anthropic**, or centralized GPU cloud services. This market is proven and thriving.
- **Polkadot’s Brand Advantage**: If Polkadot becomes known as the ecosystem offering cost-effective and decentralized AI inference services, the **Polkadot-AI brand** can grow exponentially. The potential to build out a real "AI hub" in Web3 under the Polkadot brand is enormous.

2. **Aiming at the Potential Market, Not Just Current Adoption**
This proposal isn’t just about serving the **current demand** in the Polkadot ecosystem but also positioning Polkadot as the **go-to hub for AI developers**. We are looking ahead:

- **Encouraging Innovation**: If the Polkadot community is **content** with where adoption stands today, they can indeed vote "nay". But for those of us who see untapped potential, this is a simple way to bring **new Web3 and AI projects** into Polkadot's orbit.
- **Simplified Onboarding for New Users**: What makes this approach particularly valuable is that **AI developers don’t need to know Substrate, Phala, or the intricate workings of the Polkadot ecosystem**. All they need to do is **compare GPU pricing**, sign a contract, and get access to compute power. It’s a streamlined offer that makes onboarding easier than ever.

---

### **Question 6: How to access to the hardwares?**

We aim to provide easy and flexible access to the GPUs through a **Web2-friendly platform** that’s designed for a broad audience, including AI developers who may not be familiar with Web3 infrastructures.

- **Two Access Methods**:
- **Rent GPUs via API**: Users can rent GPUs directly, similar to how they would with other cloud services like io.net or AWS. They’ll have the flexibility to rent GPUs by hour, day, month, or even year.
- **Run AI Models via API**: We’ll provide a service similar to **TogetherAI** or **OpenRouter**, where users can run AI models (like open-source models) using credits. This allows developers to focus on deploying models without worrying about the complexities of infrastructure setup.

---

### **Question 7: Who will develop the frontend, APIs, and connectivity?**

Our team at **Hashforest LLC** will develop these platforms and APIs. Importantly, we are **not asking for additional funding** from the Polkadot treasury to build this frontend, API infrastructure, or the connectivity components.

- **Why no additional funding?**
We view this as a long-term product and business opportunity. By building this infrastructure ourselves, we can create a platform that not only benefits the Polkadot ecosystem but also has the potential to provide value to Hashforest as a long-term sustainable project. Given that potential benefit, we believe it’s fair to **deliver this infrastructure at no extra cost** to the Polkadot treasury.

**Development Timeline and Milestones**

As mentioned in the proposal milestones:
- **If we can secure the GPUs**, it will take approximately **three months** to build the platform and launch the **Inference Services**.
- We have a dedicated team ready to start development as soon as the hardware is in place.

Regarding **backlog concerns**, it’s true that obtaining GPUs like the **NVIDIA H100** can sometimes face delays, as there are often long lead times for delivery. We have accounted for this possibility in our timeline and procurement strategy. The goal is to ensure that the development of the platform runs in parallel with the arrival of the hardware.

---

### **Question 8: Accepted Payment Methods**

We plan to accept payments primarily in **fiat** (USD) or **USDC**, allowing us to reach a wider audience and more seamlessly integrate with existing AI and cloud customers.
- These payments will be used to **buy back DOT** and **burn it**, aligning with our commitment to increasing DOT value and contributing to Polkadot’s long-term tokenomics.
- **DOT Payments**:
While our initial focus is on fiat and USDC, we also plan to support **DOT payments** on the platform. This option will be available to users who prefer transacting directly within the Polkadot ecosystem.
- **Treasury Funding in DOT**:
The treasury request is made in **DOT**, not USDC, which does carry financial risk for us as the proposal team. If the price of DOT declines, our team will bear that risk, as the costs (especially for hardware) are tied to USD-based pricing. However, we understand the importance of contributing to the ecosystem and are prepared to take on that risk.