Bittensor Novelty Search :: SN4 Targon :: Confidential Cloud Compute

This episode reveals how Targon is launching an enterprise-grade confidential cloud platform on Bittensor, leveraging trusted execution environments to solve the critical security and privacy challenges in decentralized AI.

Introduction: Six Months of Confidential Compute

• Rob from Targon kicks off by recapping the project's progress since their last update, highlighting six months of live operations with Trusted Execution Environments (TEEs). A TEE is a secure, isolated area within a processor that protects code and data from being accessed or modified by the host system, ensuring confidentiality. This capability was a game-changer, allowing Targon to offer stable, secure virtual machines by enabling direct SSH access for workload management.
• Rob contrasts Bittensor's iterative nature with the rigidity of smart contract-based platforms like Golem, emphasizing Bittensor's strength in allowing constant evolution.
• The subnet has successfully transitioned from a speed-based competition to a price-based one, enhancing stability for enterprise customers.
• Security is paramount, with workloads protected not only from Targon but also verifiable by third parties like Intel, AMD, and Nvidia through their attestation endpoints.

Technical Breakthroughs and Hardware Support

• The discussion details the technical advancements that underpin Targon's platform. The team has achieved fully encrypted support for both Intel (TDX) and AMD (SEV) CPUs, significantly broadening the range of compatible hardware. This allows them to support any recent-generation CPU server.
• Josh, providing technical depth, explains that on the GPU side, support is currently focused on Nvidia's Hopper series, with Blackwell compatibility expected in one to two months.
• Rob emphasizes the non-negotiable nature of security for enterprise adoption. He states, "If you could break if the host machine could break into the VM and see what the workload that was being run on there, you know that's just not something that is really marketable or palatable to most customers."
• Targon is working closely with Intel to shape next-generation TDX features. Josh reveals that Intel is incorporating hardware-level functionality requested by the Targon team, demonstrating a deep, collaborative partnership.

Product Launch: Self-Serve Platform and Serverless SDK

• Targon announces the launch of its fully self-serve platform for container and CPU server rentals, accessible at targon.com. The user experience is designed to be familiar to users of platforms like RunPod, allowing anyone to deploy workloads without direct interaction with the Targon team.
• GPU Offerings: The platform currently supports H200s, with plans to introduce GPU virtualization. This will allow users to combine multiple GPUs for larger VRAM (e.g., two H200s for 282 GB of VRAM) or fractionalize a single GPU to get H200 performance at a lower price point.
• CPU Offerings: A large inventory of virtualized CPU servers is available, supporting Docker with BusyBox and Sysbox to provide a VM-like experience within containers.
• Serverless SDK: Targon is also releasing a serverless SDK for deploying Python functions, VLM instances, web servers, and web scrapers. This creates a comprehensive cloud solution for customers to build entire pipelines, from A/B testing models to fine-tuning and deployment.

Upcoming Features: RDMA Clusters and Network Volumes

• RDMA Clusters: Support for RDMA (Remote Direct Memory Access) will be available for pre-training and large-scale fine-tuning. RDMA enables direct memory access between servers, bypassing the operating system to achieve high-throughput, low-latency networking crucial for distributed training.
• Network Volumes: Multi-node network volumes will be supported across both the rental and serverless platforms. This allows users to connect datasets and model weights to multiple containers simultaneously, ensuring data persistence and ease of use.

Enterprise Traction and Strategic Partnerships

• Rob highlights Targon's growing enterprise adoption and key partnerships that strengthen its market position.
• The team has signed its first 12-month enterprise contract and has a strong pipeline of Web3 and Frontier AI companies in active proof-of-concepts.
• To accelerate adoption, Targon is offering up to $100,000 in credits to eligible enterprise customers.
• A key strategic partner is Dippy, which processes millions of requests daily on Targon for its 8 million users.
• Targon has also partnered with major data centers in the United States and Latin America, securing access to thousands of GPUs with room for 10x expansion and achieving gross margins between 40% and 90%.

The Vision for a Transparent Compute Market

• A core part of Targon's strategy is to create a transparent, efficient market for compute. They have developed an order book for interruptible compute, matching sellers with their incentive mechanism to provide customers with an uninterruptible experience at a lower price.
• Rob draws a parallel between the opaque pricing in the US healthcare system and the current compute market, arguing that a lack of transparency prevents efficiency.
• The long-term vision includes futures markets (spot, 1-month, 3-month, 12-month) and financial derivatives built on top of compute, which they see as a potential trillion-dollar market.
• This permissionless financial layer aims to unlock new financing models for data centers, allowing them to monetize idle hardware on the interruptible market.

Const's User Experience and the TEE Revolution

• Const, the podcast host, shares his firsthand experience using Targon's new self-serve platform. He needed CPU servers and found Targon's offering to be simpler, faster, and more powerful than AWS EC2, highlighting the superior user experience of the decentralized solution.
• Const notes the revolutionary nature of TEEs for the decentralized physical infrastructure (DePIN) space, recounting a conversation with a Golem representative who didn't believe the technology was possible.
• He sees TEEs as a foundational layer for Bittensor, enabling subnets to offer monetizable, private, and verifiable services to enterprise customers. He states, "All of a sudden you become...able to give assurance to...Fortune 500 companies that no other team in the world [can]."

The Technical Mechanics of Attestation and Security

• Josh provides a detailed explanation of how TEE attestation works. Attestation is a cryptographic process where a chip uses a secret key, burned into the hardware itself, to sign a report containing metadata about the running process. This allows Targon to verify that the correct, secure VM is running on a specific machine.
• The process verifies the VM's startup flags, GPU configuration (via Nvidia SMI and NVCC checks), and other critical metadata.
• To prevent relay attacks, a VM is hard-locked to its IP address upon its initial boot-up attestation. If the VM is moved or its location is misrepresented, it will fail verification and refuse to run.
• Josh explains that this hardware-level trust is essential for a network of anonymous and untrusted compute providers, as purely software-based solutions are fundamentally vulnerable.

Targon's Competitive Edge and Evolution

• The conversation explores Targon's journey and what gives it a competitive advantage. Rob recounts the evolution of Subnet 4 from its origins as "Sibyl," an inference and search mechanism, to its current focus on confidential compute.
• The initial challenge of verifying LLM inference non-determinism and ensuring user privacy led them to pivot to TEEs in November 2023.
• Rob argues that Targon's experience fending off malicious miners has forged a battle-hardened, security-obsessed culture. This allows them to focus on customer experience rather than constantly patching exploits.
• The team's software expertise enables them to offer superior products, such as virtualized GPUs that provide the performance of an H200 (including the Transformer Engine) at the price point of a 4090 by limiting VRAM.

Connecting Revenue to the Bittensor Ecosystem

• Const raises a critical question about how fiat-based revenue from enterprise contracts will flow back to the Bittensor network and token holders.
• Rob explains that while customers paying in TAO makes buybacks straightforward, fiat payments introduce significant tax and legal complexities for a US-based team.
• The team is actively working with legal counsel to develop a compliant and automated solution for converting fiat revenue to on-chain value, aiming to create a legal precedent for other on-chain businesses.
• Targon's ultimate goal is to build a robust, permissionless compute market that benefits the entire ecosystem, including other subnets, by providing a foundational service for secure, low-cost compute.

Conclusion

This episode demonstrates that verifiable, confidential compute is a foundational primitive for decentralized AI. Targon's launch of a self-serve TEE-powered platform positions it as a critical infrastructure layer, creating immediate opportunities for enterprise adoption and providing a secure, efficient compute marketplace for the entire Bittensor network.