Hash Rate - Ep 119 - The #1 Bittensor Subnet CHUTES

This episode features John Durban from Shoots, the leading Bittensor subnet, revealing how they simplify AI model deployment and are bridging the gap between the AI and crypto worlds through innovative incentive mechanisms and a focus on permissionless access.

Introducing Shoots: Simplifying AI Deployment

• John Durban, representing Shoots (subnet 64), explains that Shoots is a "serverless compute platform." This term, while a misnomer as servers are always involved, signifies that Shoots abstracts away the complexities of infrastructure management for AI model deployment.
• Users bring their code, and Shoots handles scaling, Docker container creation (Docker containers are standardized units of software that package up code and all its dependencies so the application runs quickly and reliably from one computing environment to another), driver installation, user management, and payments.
• The platform is specifically designed for AI models, though John mentions a potential future for CPU-based applications like website hosting.
• Mark Jeffrey, the host, likens the ease of use to Squarespace, highlighting how Shoots removes the painful and time-consuming infrastructure setup, which can often consume 90% of a developer's time.

Alternatives to Shoots

• John Durban outlines alternatives depending on the user's needs:
  • Hyperscalers like AWS or Google Cloud for renting raw servers, requiring manual setup.
  • Frontier LLM (Large Language Model – an AI model trained on vast amounts of text data to understand and generate human-like language) providers like ChatGPT or Claude for pure LLM inference.
  • More direct competitors include Replicate or Modal, which offer similar features for deploying arbitrary apps with GPU selection and scaling.
• The key differentiator for Shoots is its focus on simplifying the entire AI deployment stack within the Bittensor ecosystem.

Shoots' Impressive Growth Metrics

• John Durban reveals updated user and token processing numbers, emphasizing significant growth:
  • User count has risen from a previously reported 200,000 (May) to 371,000.
  • Total tokens served have surged from around 2 trillion to approximately 7.3 trillion.
  • Shoots recently hit a milestone of processing 100 billion tokens per day. John contextualizes this by stating, "for context, this is a third of what Google was processing just one year ago."
• These metrics, publicly available on shoots.ai/app/research, showcase rapid adoption and usage.

Contextualizing Shoots' Market Share in AI Serving

• Mark Jeffrey inquires about Shoots' percentage of the broader AI landscape.
• John Durban explains the difficulty in quantifying this due to a lack of published numbers from direct API calls to frontier labs like Claude.
• However, he points to Open Router (a platform that aggregates and provides access to various AI models) rankings as a proxy. On Open Router, DeepSeek models, largely served by Shoots and Targon, rank very high (e.g., #5 and #6), suggesting significant usage.
  • John notes, "if you included the free and the non-free, I think that would actually put it closer to number three," behind Anthropic's Sonnet 4 and Google's Gemini 2.0 Flash.
• This indicates a substantial demand for open-weight models, which Shoots facilitates access to.

Bridging the AI and Crypto Divide

• John Durban highlights a core mission for Shoots: to break down the "giant chasm between the AI world and the crypto world."
• Many AI developers are hesitant about crypto. Shoots aims to demonstrate the value of Bittensor's decentralized, permissionless mechanism, which crypto enables.
• Offering free or heavily discounted access to models, like getting new models like DeepSeek up quickly, is a strategy to attract AI users and show the benefits of a decentralized platform where models are locked to specific weights, ensuring consistent API response quality.
• Strategic Implication: Investors should note Shoots' strategy of using free tiers and rapid model deployment to onboard users from the traditional AI space, potentially creating a large user base before full monetization.

The Shoots Business Model: Subsidies and Revenue

• Shoots offers models for free or at a steep discount (e.g., 1/6th the cost of AWS self-hosting) because it is subsidized by emissions from the Bittensor TAO chain, similar to how Bitcoin miners earn Bitcoin.
• John Durban confirms, "yes, we are definitely 100% subsidized by Bittensor emissions."
• The platform recently added fiat payments alongside TAO payments.
  • TAO: The native cryptocurrency of the Bittensor network, used for staking, governance, and payments within the ecosystem.
• A key challenge with fiat is regulatory compliance, such as OFAC (Office of Foreign Assets Control – a U.S. Treasury department that enforces economic and trade sanctions), especially for a decentralized platform aiming to serve users globally, including those in embargoed countries.

Tokenomics: Staking Revenue and Operational Costs

• Revenue from TAO payments is automatically staked to the Shoots Alpha token (the specific token for the Shoots subnet). John explains, "by staking we're basically increasing the value of Shoots for all token holders."
• There's consideration to burn Alpha tokens received as revenue instead of staking, which would permanently remove them from supply, having a similar effect of locking up value.
• Fiat revenue is currently manually converted to TAO, then staked to the Alpha token, with all payment data being public via their API.
• Miners and validators on the Shoots subnet sell their earned Alpha tokens to cover server bills and operational expenses. The subnet owners (like Shoots itself) also sell a portion of their 18% emission share to cover employee salaries and other costs.
• John acknowledges the apparent contradiction of staking revenue while also selling emissions but frames it as setting a precedent for a future where revenue will exceed expenses, with the system already in place to direct all revenue to benefit the Alpha token.

Miner Incentives on Shoots

• Miners on Shoots supply GPU power to serve models. They are not supplying abstracted compute to a general grid; instead, they dedicate machines to specific models.
• The incentive mechanism for miners is based on four categories:
  1. Compute Units: Seconds of GPU usage.
  2. Number of Requests Served: Sheer volume of requests.
  3. Bounties for Cold Starts: Incentivizing quick loading of new or infrequently used models. A cold start refers to the delay when a model that is not actively running needs to be loaded into memory before it can process a request.
  4. Unique Shoot Score: Breadth of models a miner runs, encouraging models to stay "hot" (readily available).
• John Durban states, "we don't control what miners deploy at all... you as a miner are trying to optimize for those four categories." This decentralized approach to scaling and model availability is a core strength.
• Researcher Insight: The multi-faceted incentive mechanism is a key area for research, analyzing how it drives miner behavior and optimizes resource allocation for diverse AI models.

The Future Vision: Permissionless, Open-Weight AI

• John Durban's motivation stems from issues with frontier labs censoring responses, even for legitimate use cases like e-discovery.
• Shoots aims to be a permissionless, open-weight, open-source platform where builders know exactly what model they are using and what responses to expect, free from arbitrary censorship or API changes.
• Success for Shoots means becoming a "universal compute substrate" for AI and potentially other applications, fostering wider adoption and bringing anti-crypto AI users on board.
• Shoots is transitioning away from entirely free models to a tiered system (free tier + pay-as-you-go/subscriptions), similar to Google Gemini, to start generating more direct revenue.

Profitability and Sustainability

• John Durban discusses profitability:
  • One definition: Revenue exceeding the cost of emissions sold by the subnet team.
  • A broader definition: Revenue offsetting the sum of all emissions (owner, validator, miner).
• His initial goal for "reasonable" profitability is "offsetting all minor emissions with revenue."
• A major barrier to mass revenue and enterprise adoption is privacy. Businesses need assurance that their prompts and data remain confidential.
• Mark Jeffrey highlights the IP risk: "once you do that [upload to a general AI], the AI now knows your plot. It's now training data and your really unique premise may be spit out in an answer to somebody else."

Ensuring Privacy with Trusted Execution Environments (TEE)

• Shoots is working on implementing TEE (Trusted Execution Environments) for miners. TEEs are secure areas within a processor that guarantee code and data loaded inside are protected with respect to confidentiality and integrity.
• While TEEs like Intel's TDX (Trust Domain Extensions) offer encryption and cryptographic verification of the secure environment, they are not a panacea and can be vulnerable to side-channel attacks or misconfigurations.
• Shoots aims to use TEEs to prove that user data is always encrypted and that the running code is verifiable against open-source code, with logging disabled.
• John emphasizes the "don't trust, verify" philosophy: "You don't have to trust us that your data is private. You can actually just go check the report, compare the source code, and know for certain that it's never going to be leaked."
• Actionable Insight: The successful implementation and adoption of robust TEEs could be a major catalyst for enterprise adoption of Shoots and other decentralized AI platforms, addressing key privacy concerns.

What Does a Breakout Bittensor Subnet Look Like?

• Mark Jeffrey poses the question of what characterizes a successful, sustainable, "breakout" subnet, potentially reaching billion-dollar valuations.
• John Durban draws a parallel to Bitcoin mining, which generates no direct product revenue but incurs massive costs ($12 billion/year in power/equipment) based on speculation and the value of maintaining the ledger. He questions if providing "free and sort of incredibly powerful unlimited permissionless access to intelligence" via Bittensor could be worth a similar expense, even without direct revenue.
• He believes a breakout subnet with massive profit margins could skyrocket its token price if it bought back tokens, but such cycles can be short-lived, citing the boom and bust of AI agent/virtual tokens.
• Shoots aims to be the "stable foundation" for any AI application, regardless of future breakthroughs like AGI. They are also building second and third-order apps (e.g., an agents platform, a chat interface) on top of their foundational compute layer to capture higher profit margins.
• Investor Consideration: The path to a "breakout" subnet is still unclear. Investors should look for subnets with strong foundational technology, clear revenue models (even if initially subsidized), and strategies to build higher-margin applications or services on top.

The Bittensor Ecosystem: Challenges and Evolution (DTOs & Root Emissions)

• John Durban views Bittensor as still "pretty early in its evolution" in a business sense, with work needed on privacy, security, and enterprise adoption.
• He notes that Shoots, at 6 months old, achieved similar daily token processing numbers (around 160 billion at peak) as established players like Together AI, showcasing the power of Bittensor's incentive design.
• DTO (Dynamic TAO Offering): A mechanism in Bittensor that allows subnets to have their own tokens (Alpha tokens in Shoots' case) which are dynamically priced and traded against TAO. This has accelerated development and competition.
• The "Root is Too Damn High" Debate:
  • Mark Jeffrey raises the concern that high APY (Annual Percentage Yield – the real rate of return earned on an investment, taking into account the effect of compounding interest) on staking TAO directly to the root network (currently around 12%, previously as high as 25%) disincentivizes investment in individual subnets.
  • A limit of 128 subnets has been temporarily imposed, which John sees as a good move to prevent "unlimited dust paid out to all these subnets that aren't really doing anything" and to increase competitive pressure, though it might make it harder for new, talented teams to emerge.
  • John Durban believes high root APY can attract institutional capital that may not initially delve into subnet-specific investments. He suggests a balanced approach, possibly allowing root stakers to opt-in to receive their rewards as a proportional mix of Alpha tokens from various subnets (like an airdrop), reducing root sell pressure without drastic changes to the DTO system so early in its lifecycle.
• Strategic Implication: The evolution of DTO mechanics and root emission policies will significantly impact subnet viability and investor strategies. Monitoring these discussions and proposals within the Bittensor community is crucial.

Clarifying Roles: Namay and Core Development

• Mark Jeffrey asks about Namay, a core dev.
• John Durban clarifies that Namay is the founder of Rayon Labs (the entity behind Shoots, Gradients, and subnet 19) and operates at a management level, coordinating the three subnets.
• John himself is responsible for the rapid deployment of new AI models on Shoots, stating, "it's my fault those models get up so fast because I'm like obsessed."

This episode underscores Shoots' pivotal role in making AI compute accessible and its strategic efforts to onboard traditional AI users to the decentralized Bittensor ecosystem. Investors and researchers should monitor Shoots' progress in revenue generation, privacy solutions like TEE, and the broader evolution of Bittensor's tokenomics.