Excerpts from Subnet 64’s Chutes & Answers AMA

In crypto, the most important shifts rarely arrive as announcements. They surface gradually, through technical conversations, design tradeoffs, and moments where builders begin to articulate constraints that no longer scale. The first AMA from Chutes, Bittensor Subnet 64, was one of those moments.

What might initially appear to be a routine community session evolved into something more consequential. It revealed not only how Chutes is positioning itself within the decentralized AI stack, but also how the team is thinking about one of the hardest problems in the space today, which is how to scale AI training beyond the limits of centralized infrastructure.

At the center of the discussion were two voices: Timon (communications and sales at Chutes), who framed the product and its philosophy, and Jon Durbin (backend dev. at Chutes), who provided a detailed breakdown of the technical direction. Around them, a live audience of builders and contributors shaped the conversation in real time, grounding theory in practical questions.

Positioning Chutes: Infrastructure, Not Interface

The session opened with a clear framing that Chutes is not trying to be another application layer, it is positioning itself as infrastructure, specifically:

a. A decentralized inference as a service layer,

b. A confidential compute provider, and

c. A backend system for AI applications operating on Bittensor.

This distinction matters because it defines where value accrues. Rather than competing in user interfaces or model branding, Chutes is building at the layer where computation, privacy, and execution converge.

The guiding principle which is rarely enforced across the industry is that privacy should be a default condition of AI usage, not a premium feature.

This is a direct response to the current paradigm, where most inference providers retain visibility into prompts, outputs, and usage patterns. Chutes is attempting to invert that model entirely.

Confidential Compute as a First Principle

The technical foundation of Chutes is built around eliminating visibility across the entire inference pipeline.

Instead of relying on policy guarantees, the system enforces privacy at the hardware and protocol level.

1. Trusted Execution Environments

Chutes leverages Intel TDX-based Trusted Execution Environments, where computation is isolated inside hardware enclaves. This has several implications:

a. Even node operators cannot access data,

b. Physical access to machines does not expose workloads, and

c. The infrastructure layer itself is blind to user activity.

The result is a system where trust is minimized not through reputation, but through architecture.

2. End-to-End Encryption at the Edge

The encryption model extends beyond the server as data is encrypted on the user’s device before transmission, routed directly to a verified enclave, processed and returned without ever being decrypted in transit

Decryption ‘can’ only occurs at the user endpoint. This effectively removes the entire attack surface associated with intermediary visibility, including internal monitoring, logging systems, or compromised nodes.

3. Encrypted Interconnects

Even within the system, communication between components remains encrypted, including traffic between CPUs and GPUs. The combined effect is a closed loop of encrypted computation, where every stage is verifiably opaque.

As Algorod summarized during the session, intercepting the system yields no usable data, only encrypted outputs.

From Product to Platform: Serving Builders, Not Just Users

The AMA moved quickly from architecture into usage, driven by a question from a developer building a multi model AI application.

The operational concern was that public model availability is inherently dynamic, constrained by hardware supply and cost efficiency. For production applications, this introduces instability.

The response revealed a critical layer of the Chutes stack: Private Chutes as Serverless AI Infrastructure.

Developers can deploy dedicated environments, referred to as private Chutes, which function as configurable compute instances.

These environments allow developers to:

a. Deploy any model of their choice,

b. Control scaling behavior, and 

c. Maintain consistent availability independent of public offerings.

The pricing model reflects infrastructure economics rather than API abstraction:

a. Billing is based on GPU usage, calculated at a granular level,

b. There are no per token inference costs within private deployments, and

c. Developers can layer their own pricing models on top.

This effectively turns Chutes into a backend layer for AI native applications, where teams can build, deploy, and monetize models without relying on centralized providers.

Operational Reset and Revenue Recovery

The team also addressed recent fluctuations in platform metrics. Structural changes were made to the user base and subscription model, removing segments that were not aligned with long term sustainability.

While this led to a temporary decline in visible activity, the outcome was intentional. The system has since stabilized, with:

a. Daily revenues ranging between $15,000 and $20,000,

b. Improved efficiency in resource allocation, and

c. Higher quality demand driving usage.

This reflects a broader trend across decentralized infrastructure, where short-term growth is often sacrificed to maintain long-term viability.

Governance Under Pressure: Designing Against Failure

A notable portion of the AMA focused on governance risks, particularly in light of failures observed across other subnets. Jon Durbin addressed this directly, emphasizing that: Failures in one part of the network affect the credibility of the entire system.

Chutes has implemented several mechanisms to mitigate these risks:

a. Smart contract based management of owner emissions,

b. Locked principal funds to prevent unauthorized extraction,

c. Multi-key control systems with separation of responsibilities, and

d. Time-locked emergency mechanisms to handle edge cases.

These safeguards are designed to eliminate unilateral control over funds, even at the operator level.

The underlying philosophy is that decentralization is not achieved through intent but enforced through constraint.

The Constraint That Matters Most: Training

While inference and privacy form the current product, the most significant part of the AMA emerged when the conversation shifted toward training.

This is where the limits of existing systems become most apparent. Training large models today is constrained by several factors:

a. Scarcity of high end GPUs such as H100 and B200,

b. Dependence on large-scale data centers,

c. Massive dataset requirements, and

d. Limited participation due to hardware thresholds.

Even decentralized approaches often replicate these constraints, relying on clusters of powerful nodes rather than truly distributed systems. The result is a paradox: Decentralized AI exists in theory, but remains practically centralized in execution.

A New Architecture: Decoupling Scale from Hardware

Chutes is attempting to address this by rethinking how training workloads are distributed. The approach is built around Mixture of Experts architectures, where different parts of the model can be trained independently. Core component of this architecture involves:

a. Composer Node: Maintains the full model and dataset, handles routing and coordination, performs validation and aggregation, and operates on high performance infrastructure.

b. Expert Workers: Train smaller segments of the model, operate on distributed hardware, and send lightweight updates back to the composer

This separation allows the system to offload the majority of computation to a distributed network, while maintaining coherence through a centralized coordination layer.

Lowering the Barrier to Participation

The key innovation is that participation no longer requires access to data centers or high end clusters.

Instead, contributors can:

a. Use consumer grade GPUs,

b. Download small subsets of model weights,

c. Process local computations, and

d. Submit compact updates over low bandwidth connections.

Even devices with limited resources can contribute, provided they meet minimal requirements. This transforms training from an institutional activity into a globally accessible process.

Validation as a First Class Constraint

One of the historical challenges in decentralized training is ensuring that contributions are valid and non malicious. Chutes addresses this through:

a. Real-time validation of gradient updates,

b. Replay based auditing mechanisms,

c. Lightweight verification techniques that do not require full recomputation, and

d. Slashing conditions for invalid or adversarial contributions.

This allows the system to maintain integrity without relying on trust or centralized oversight, and also enables immediate exclusion of bad actors, preventing long term corruption of training runs.

The Market Implication: Unlocking Idle Compute

A contributor during the AMA articulated the broader implication succinctly by noting that AI development today is constrained not by demand, but by access to compute, and if idle hardware across the world can be aggregated into a functional training network, the supply side of AI changes entirely.

This, however, introduces several second order effects:

a. Reduced dependence on centralized providers,

b. Increased competition from open source models, and

c. A new economic layer where compute become directly monetizable.

Chutes is effectively attempting to tap into this latent resource.

Toward a Distributed Training Economy

The long-term vision extends beyond technical feasibility. Participants in the network could:

a. Earn tokens by contributing compute,

b. Use those tokens for inference, and

c. Participate in both production and consumption within the same system.

This creates a closed loop economy where value flows between contributors, developers, and users. It also aligns incentives across the network, encouraging sustained participation.

What Comes Next

The audio chat revealed that the system is already functional at smaller scales, with ongoing work focused on:

a. Scaling the architecture to larger models,

b. Publishing a detailed whitepaper,

c. Building onboarding tools for distributed participants, and

d. Integrating training outputs with inference infrastructure.

The roadmap is ambitious, but grounded in iterative progress.

Conclusion: From Infrastructure to Coordination Layer

What emerged from the Chutes AMA is not just a product update, but a shift in perspective. Chutes is evolving from a confidential inference provider into a coordination layer for decentralized AI computation.

This transition has broader implications for the entire ecosystem as it suggests a future where AI infrastructure is privacy preserving by default, model training is globally distributed, and participation is open to anyone with compute.

And perhaps most importantly, where the development of frontier models is no longer constrained by the boundaries of centralized capital and hardware access.

In a landscape increasingly defined by concentration, Chutes is exploring what it would mean to build AI systems that scale in the opposite direction.