# Company Snapshot Source: https://docs.farmgpu.com/company-docs/company-overview/company-snapshot FarmGPU company information, legal entity details, and core business description # Company Snapshot | Category | Detail | | --------------------- | --------------------------------------------------------------------------------------------------------------- | | **Legal Entity Name** | Farm GPU Inc. | | **Trade Name** | **FarmGPU** | | **Founded** | 2024 | | **HQ & Primary Ops** | 10850 Gold Center Dr ste 251, Rancho Cordova, CA 95670 | | Data Centers | Conscious Data Center 3141 Data Drive Rancho Cordova, CA 95670
Evocative Data Centers Santa Clara Facility | | **Ownership** | Private (Founder-led) | | **Employee Count** | \~12 (FTE + Strategic Contractors) | ## Core Business Description FarmGPU is redefining the future of GPU-powered cloud computing by offering cost-effective, scalable, and high-performance GPU resources tailored specifically for AI developers, innovative startups, and enterprises worldwide. The FarmGPU Cloud Platform that consists of ## Core Offerings 1. **On-demand compute:** Pay as you go instances for the latest NVIDIA GPUs, including H200, B200, and RTX Pro 6000 Blackwell. 2. **Managed GPU Clusters:** Slurm/Kubernetes clusters for large-scale training with high performance RDMA Ethernet fabric 3. Turnkey **GPU Investment**: Bring your own capex, have FarmGPU manage end to end, from planning, data center, networking, deployment and automation, to monetiziation on our partner platforms 4. **AI Data Center Blueprint:** Transformation services for traditional facility operators, includes data center capacity planning, site assessment, and strategic deployment options for air, liquid, and immersion GPU servers. # Founders' Thesis Source: https://docs.farmgpu.com/company-docs/company-overview/founders-thesis Why FarmGPU exists and our approach to AI infrastructure # Why FarmGPU Exists (Founders' Thesis) ## 1) Why On-Demand Compute Beats Traditional Cloud The cloud promised flexibility, efficiency, and access to cutting-edge infrastructure. For AI workloads, that promise is only partially fulfilled. Hyperscalers were designed around **general-purpose, multi-tenant computing**. Their abstractions—virtual machines, managed services, quotas, and proprietary platforms—optimize for scale and control, not for GPU efficiency or developer autonomy. For AI teams, this creates fundamental friction: * **High and unpredictable costs** driven by opaque pricing, egress fees, and bundled services * **Limited access** to GPUs due to quotas, regional scarcity, and long reservation cycles * **Vendor lock-in** that restricts portability and experimentation * **Inefficient utilization**, with GPUs frequently stalled by I/O and networking bottlenecks On-demand GPU compute flips this model. By providing **direct, hardware-proximate access** to GPU infrastructure—without proprietary abstractions—on-demand platforms allow AI teams to: * Scale up or down instantly * Pay only for what they use * Maintain full control of their software stack * Optimize performance without cloud-imposed constraints For modern AI workloads, **on-demand, infrastructure-native compute is not a compromise—it is the optimal model**. | Cloud Promise | Description | Reality | | ------------------------ | ------------------------------------------------------------------ | ---------------------------------------------------------------------- | | Scalability & Agility | Resources can be rapidly scaled to meet changing demands. | Partly true. Vendor lock-in is real. | | Cost Efficiency | Pay-as-you-go model, eliminating upfront hardware investments. | False. Cloud can be extremely expensive. | | Innovation & Flexibility | Access to cutting-edge technology for fast experimentation. | False. Limited or no access to entry-level GPUs. | | Reliability & Security | Global data centers ensure data availability and security. | True. CSPs take durability, availability, and security very seriously. | | Global Reach | Facilitates global collaboration through worldwide infrastructure. | True. Easy access to various regions. | ## 2) Why FarmGPU Wins with Storage Expertise Compute does not bottleneck AI systems—**data movement does**. Most GPU cloud providers treat storage as a secondary concern, relying on generic network-attached solutions that are poorly suited for AI workloads. This leads to: * GPUs waiting on data * Inconsistent performance * Poor scaling behavior * Hidden costs as datasets grow FarmGPU was built around a different insight: **storage is the performance moat in AI infrastructure**. We design and operate storage as a first-class system component: * Deep partnerships with storage leaders like **Solidigm** and leading storage ISVs * Custom storage servers optimized for AI data paths * **DPU-accelerated architectures (NVIDIA BlueField-3)** to offload networking, security, and storage processing from CPUs * Native support for **AI-specific storage patterns**, including: * High-throughput training pipelines * KV-cache offload for inference * Vector databases and embedding workloads * Block, file, and object storage tuned for GPUs This storage-centric approach delivers: * Higher GPU utilization (MFU) * Predictable performance under real workloads * Lower cost per unit of useful compute As models grow larger and inference becomes more data-intensive, **storage expertise becomes the primary differentiator**. This is where FarmGPU has a durable advantage. *** ## 3) The Roadmap to the Lowest TCO in AI Infrastructure FarmGPU's long-term advantage is not tied to any single GPU generation—it is rooted in **systems-level cost optimization**. Our roadmap to the lowest total cost of ownership (TCO) is built on three pillars: ### Open Source and Linux Leadership We embrace open systems at every layer: * Custom neocloud OS with Tractor * Open networking with OCP This reduces licensing costs, increases transparency, and allows continuous optimization as hardware evolves. ### Optimized Data Center Design Instead of overbuilt hyperscale facilities, we deploy **AI-optimized data centers**: * High-density GPU racks, tuned to existing DC footprints * Power and cooling designed around real GPU thermals * Incremental upgrades of existing Tier III / IV facilities * Capital deployed only when demand exists This approach lowers CapEx, accelerates deployment, and improves return on invested capital. ### Network and Fabric Efficiency AI performance depends on fast, predictable communication: * High-bandwidth, low-latency fabrics * DPU-offloaded networking and security * Topologies designed for collective communication, not web traffic By treating networking as part of the compute system—not an external service—we reduce both latency and cost. # Mission, Vision & Positioning Source: https://docs.farmgpu.com/company-docs/company-overview/mission-vision-positioning FarmGPU's mission to democratize AI infrastructure, our vision for the future, and strategic positioning # Mission, Vision, and Positioning ## Mission FarmGPU empowers the AI revolution by delivering **secure, high-performance, and cost-efficient GPU infrastructure** purpose-built for modern AI workloads. We provide on-demand and dedicated GPU servers through **secure, local data centers**, and we partner deeply across hardware, platform software, and emerging technologies to build a **scalable NeoCloud ecosystem**. Our mission is to **decentralize and democratize AI infrastructure**, giving builders real alternatives to hyperscalers—without sacrificing performance, security, or control. *** ## Vision We envision a world where **open and sovereign AI wins**. In this future: * AI infrastructure is **local, auditable, and secure by default** * Builders retain **full control over their data, models, and economics** * Open standards and open ecosystems outpace closed hyperscaler platforms * Specialized NeoClouds compete head-to-head with hyperscalers on performance, cost, and reliability FarmGPU's long-term vision is to be a **foundational NeoCloud provider**—powering AI innovation globally while remaining capital-efficient, standards-driven, and aligned with the communities that build AI, not monopolize it. ## Category Definition: The NeoCloud > NeoCloud: A specialized cloud provider purpose-built for AI workloads Unlike hyperscalers, NeoClouds: * Are **GPU-first**, not CPU-centric * Optimize for **real AI workloads**, not generic VMs * Deliver **dramatically better price-performance** # Strategic Positioning ## High Performance, Low Cost FarmGPU delivers **up to 70% lower total cost of ownership (TCO)** compared to major hyperscalers, without compromising performance or reliability. We achieve this through: * Optimized BOMs tuned for AI workloads * Lean team leveraging AI for everything * Deep hardware–software co-optimization * Transparent pricing aligned to actual infrastructure costs Rather than monetizing lock-in or proprietary services, FarmGPU aligns incentives with customers: **better utilization, better performance, lower cost**. *** ## 3.2 NeoCloud for Real-World AI: Storage as a First-Class Citizen AI workloads are **not compute-only**. Training, fine-tuning, and inference at scale demand **fast, reliable, and predictable storage**—something most NeoClouds and bare-metal GPU providers underinvest in. FarmGPU is uniquely positioned as a **storage-centric NeoCloud**. We differentiate by: * Designing storage as a **core architectural pillar**, not an afterthought * Partnering with industry leaders like **Solidigm** and major storage ISVs * Deploying **custom storage servers** tuned for AI workloads * Leveraging **DPU-accelerated architectures (NVIDIA BlueField-3)** to offload storage, networking, and security from CPUs * Supporting **AI-native storage patterns**, including: * KV-cache offload for inference * Vector databases * High-throughput block, file, and object storage This approach enables FarmGPU to deliver **predictable performance at scale**—from small research clusters to enterprise and production AI systems. *** ## 3.3 The Trust Layer Trust is foundational to AI infrastructure. FarmGPU is building a **defense-in-depth trust layer** that spans hardware, software, and operations. ### Security & Compliance * **SOC 2 Type II** (in progress) * Zero-trust networking model * Hardware-enforced isolation using DPUs * Secure lifecycle management of storage and compute assets ### Confidential Computing FarmGPU is one of the few NeoClouds investing early in **end-to-end confidential computing** for GPUs: * Intel **TDX** support with NVIDIA B200 * Secure enclaves for sensitive AI workloads * AMD **SEV** support planned for **Q2 2026** * Architecture designed to protect data **in use**, not just at rest or in transit ### Why This Matters This trust layer enables: * Regulated and sensitive workloads (enterprise, government, research) * Model protection and IP isolation * Verifiable data handling and sanitization * Long-term customer confidence as AI regulation increases # Data Center Specifications Source: https://docs.farmgpu.com/company-docs/data-center-operations/data-center-specifications FarmGPU data center power, network, security, and environmental infrastructure FarmGPU's data center infrastructure is primarily hosted at a tier 3 datacenter in Rancho Cordova, CA called Conscious Capital Data Center. We meet our most stringent partner requirements for 99.99% availability, ensuring enterprise-grade reliability, security, and performance for AI workloads. ## Power Infrastructure * Dual independent utility feeds from separate substations, each capable of supporting 100% of data center load * N+1 redundant UPS systems with minimum 15-minute runtime at full load * N+1 redundant generator systems with 48-hour on-site fuel storage * 2N redundant power distribution paths from utility to rack level * Remote power monitoring and management at rack level * Monthly generator load testing and quarterly full-load system tests * 24/7 on-site facilities team for immediate response ## Network Architecture * Minimum 100 Gbps total bandwidth capacity * Dual diverse internet circuits from separate providers with BGP routing * Redundant core switches in high-availability configuration * DDoS protection solutions * Performance guarantees: * Sub-80% network utilization during peak hours * Less than 0.1% packet loss * P95 round-trip time under 4ms within data center * P95 jitter under 3ms within data center ## Security and Compliance * ISO 27001 certified data center * 24/7 on-site security personnel * Isolated secure racks with restricted access * Comprehensive access logging and monitoring * Regular security audits and compliance reviews * Tier III+ data center standards compliance ## Environmental Controls * Advanced liquid cooling systems for optimal GPU performance * N+1 redundancy for cooling infrastructure * Real-time environmental monitoring and alerts * Regular thermal imaging and preventive maintenance FarmGPU is compliant with the RunPod secure cloud data center and network specifications below [https://docs.runpod.io/hosting/partner-requirements#3-data-center-power-requirements](https://docs.runpod.io/hosting/partner-requirements#3-data-center-power-requirements) ## **3. Data Center Power Requirements** | **Requirement** | **Specification** | | ----------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Utility Feeds | Minimum of two independent utility feeds from separate substations. Each feed capable of supporting 100% of the data center's power load. Automatic transfer switches (ATS) for seamless switchover between feeds with UL 1008 certification (or regional equivalent) | | UPS | - N+1 redundancy for UPS systems- Minimum of 15 minutes runtime at full load | | Generators | N+1 redundancy for generator systems. Generators must be able to support 100% of the data center's power load. Minimum of 48 hours of on-site fuel storage at full load. Automatic transfer to generator power within 10 seconds of utility failure | | Power Distribution | Redundant power distribution paths (2N) from utility to rack level. Redundant Power Distribution Units (PDUs) in each rack. Remote power monitoring and management capabilities at rack level | | Testing and Maintenance | Monthly generator tests under load for a minimum of 30 minutes. Quarterly full-load tests of the entire backup power system, including UPS and generators. Annual full-facility power outage test (coordinated with Runpod). Regular thermographic scanning of electrical systems. Detailed maintenance logs for all power equipment. 24/7 on-site facilities team for immediate response to power issues | | Monitoring and Alerting | Real-time monitoring of all power systems. Automated alerting for any power anomalies or threshold breaches | | Capacity Planning | Maintain a minimum of 20% spare power capacity for future growth. Annual power capacity audits and forecasting | | Fire Suppression | Maintain datacenter fire suppression systems in compliance with NFPA 75 and 76 (or regional equivalent) | ## **4. Network Requirements** | **Requirement** | **Specification** | | ----------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Internet Connectivity | Minimum of two diverse and redundant internet circuits from separate providers. Each connection should be capable of supporting 100% of the data center's bandwidth requirements. BGP routing implemented for automatic failover between circuit providers. 100 Gbps minimum total bandwidth capacity | | Speed Requirements | Preferred: >= 10 Gbps sustained upload/download speed per server. Minimum: >= 5 Gbps sustained upload/download speed per server. Speed measurements should be based on sustained throughput over a 60 second interval during a typical workload | | Core Infrastructure | Redundant core switches in a high-availability configuration (e.g., stacking, VSS, or equivalent) | | Distribution Layer | Redundant distribution switches with multi-chassis link aggregation (MLAG) or equivalent technology. Minimum 100 Gbps uplinks to core switches | | Access Layer | Redundant top-of-rack switches in each cabinet. Minimum 100 Gbps server connections for high-performance compute nodes | | DDoS Protection | Must have a DDoS mitigation solution, either on-premises or on-demand cloud-based | | Quality of service | Maintain network performance within the following parameters: Network utilization levels must remain below 80% on any link during peak hours. Packet loss must not exceed 0.1% (1 in 1000) on any network segment. P95 round-trip time (RTT) within the data center should not exceed 4ms. P95 jitter within the datacenter should not exceed 3ms | | Testing and Maintenance | Regular failover testing of all redundant components (minimum semi-annually). Annual full-scale disaster recovery test. Maintenance windows for network updates and patches, with minimal service disruption scheduled at least 1 week in advance | | Capacity Planning | Maintain a minimum of 40% spare network capacity for future growth. Regular network performance audits and capacity forecasting | # Monitoring and Reliability Source: https://docs.farmgpu.com/company-docs/data-center-operations/monitoring-reliability Shepherd reliability engine and FarmGPU's monitoring capabilities # Monitoring, Reliability & Shepherd At FarmGPU, we view monitoring not as a set of dashboards, but as the operational backbone of the AI data path. Our reliability model is built specifically to exceed the **ClusterMax™ Evaluation Criteria**, ensuring that "random" hardware events never dictate the success of your training run. ## 1. Shepherd: The "Autonomous SRE" Shepherd is FarmGPU's proprietary reliability engine. It continuously audits the state of the cluster against the **ClusterMax™ Evaluation Criteria**, automating the transition from detection to remediation. ### The Unified Telemetry Fabric * **GPU Failure Detection:** Real-time capture of **NVIDIA XID/SXID** events, SM occupancy, memory pressure, and **PCIe AER rates**. * **Interconnect (RoCEv2):** Monitoring of link flaps, retry counters, and fabric-wide **NCCL collective timing**. * **Data Plane:** BlueField-3 telemetry for NVMe queue depths, IOPS/throughput histograms, and DPU-offload health. * **Physical:** PDU-level power draw, fan speeds, and multi-sensor thermal mapping (CPU, GPU, RAM, NIC, and Transceivers). ## 2. Passive Health Checks (The Safety Net) Aligned with **ClusterMax™ Monitoring Standards**, Shepherd runs continuous passive checks to identify "failing" nodes before they become "failed" nodes. * **GPU/Memory ECC Monitoring:** Real-time tracking of single-bit and double-bit ECC errors (via DCGM and EDAC) with automated node-cordoning upon crossing threshold. * **Link Flap Detection:** Sub-millisecond monitoring of fabric stability; Shepherd automatically drains nodes exhibiting link degradation to prevent NCCL collective timeouts. * **Thermal Drift Alerts:** Proactive detection of thermal anomalies in transceivers or GPUs, triggering automated power-capping or workload migration before throttling occurs. * **"GPU Off the Bus" Detection:** Immediate capture of `NVML_ERROR_GPU_IS_LOST` or XID 79 events, with automated node isolation. ## 3. Active Health Checks Idle nodes in the Farm undergoes regular "Active Diagnostic" suite: * **Compute Validation:** NVIDIA DCGM Diag level 3 and 4 * **Stress Testing:** In house developed **GPU-Burn** validate thermal stability under maximum synthetic load. * **Interconnect Validation:** Pairwise `ib_write_bw` and multi-node **NCCL All-Reduce** tests to verify that every node achieves >90% of theoretical fabric bandwidth. * **Storage Throughput:** Validating Sustained GB/s-per-GPU from the Silo storage tiers to the GPU memory. ## 4. Managed Observability (Grafana) FarmGPU bare metal customers will get a dedicated monitoring stack and a custom URL. * **Cluster-Level:** Node availability, pending pod counts, and fabric-wide goodput. * **Node-Level:** Granular GPU SM utilization, power draw, and temperature heatmaps. * **Job-Level:** Integration with **Slurm (sacct)** and **Kubernetes** to correlate infrastructure metrics directly with training loss curves. # Data Retention Policy Source: https://docs.farmgpu.com/company-docs/legal/data-retention-policy Retention schedules for the FarmGPU LLM inference API # FarmGPU Inference API Data Retention Policy **Effective date:** August 29, 2026 This policy defines how long each category of data associated with the **FarmGPU Inference API** is kept, and how it is deleted. It should be read together with the [Privacy Policy](/company-docs/legal/privacy-policy). The governing principle: **request content is never retained; operational metadata is retained only as long as it is useful for running a reliable service.** *** ## Retention schedule | Data category | What it contains | Retention | Purpose | | -------------------------------- | ----------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------ | ------------------------------------------ | | **Prompts and completions** | Request and response content, tool definitions, reasoning output | **Not persisted.** In memory only, for the duration of the request | Serving the request | | **In-memory prefix cache** | Token-level KV state of recent request prefixes, in GPU memory | Transient; continuously evicted, fully cleared on process restart | Latency reduction for repeated prefixes | | **Distributed traces** | Request lifecycle timing and sizes (OpenTelemetry spans); no content | Up to **7 days**, then automatically deleted | Reliability debugging | | **Application logs** | Serving-engine and gateway logs: timestamps, status codes, request/queue state; no content | Up to **30 days**, then automatically deleted | Incident investigation, abuse prevention | | **Edge / network logs** | Connection metadata at the network perimeter (IPs, URLs, timing) | Up to **30 days** | Security, DDoS and abuse mitigation | | **Aggregate metrics** | Latency histograms, token-count totals, error and utilization counters; fully aggregated, no per-request payloads | Up to **13 months** | Capacity planning, SLO reporting | | **Usage and billing records** | Per-partner / per-customer token counts and request totals | **7 years** | Invoicing, accounting, and tax obligations | | **Direct-customer account data** | Business contact and billing details | Life of the agreement **+ 7 years** | Contract and accounting obligations | *** ## How deletion works * **Content is never written**, so there is nothing to delete: prompts and completions exist only in volatile GPU and host memory and are gone when the request completes, the cache entry is evicted, or the process restarts. * Telemetry stores (traces, logs, metrics) enforce the windows above through **automatic time-based expiry** — deletion does not depend on a manual process. * Backups of telemetry are subject to the same maximum windows. ## Zero Data Retention (ZDR) status Our operating posture is zero retention of request content. However, we do not yet advertise a formal **ZDR designation**: we are completing an independent validation pass confirming that no diagnostic system (including request tracing) can capture content under any configuration. Until that validation is published, partners should treat this endpoint as **"no content logging or training, short-lived operational telemetry"** rather than certified ZDR. ## Exceptions Retention windows may be extended only when: * Data is subject to a **legal hold** or valid legal process, or * Metadata is needed for an active **security or abuse investigation**. In both cases the extension applies only to the specific records involved, access is restricted to authorized personnel, and normal expiry resumes when the matter closes. These exceptions cannot create retention of prompt or completion content, because that content is never stored in the first place. ## Questions Contact **[privacy@farmgpu.com](mailto:privacy@farmgpu.com)**. # Privacy Policy Source: https://docs.farmgpu.com/company-docs/legal/privacy-policy How FarmGPU handles data on its LLM inference API # FarmGPU Inference API Privacy Policy **Effective date:** August 29, 2026 This policy describes how FarmGPU, Inc. ("FarmGPU", "we", "us") handles data sent to and generated by the **FarmGPU Inference API** — our OpenAI-compatible API for serving large language models, including endpoints made available through routing partners such as OpenRouter. It supplements, and does not replace, any Master Service Agreement or partner agreement in place with your organization. For how long each category of data is kept, see the [Data Retention Policy](/company-docs/legal/data-retention-policy). *** ## The short version * **We do not store your prompts or completions.** Request and response content is processed in memory to generate your response and is not written to persistent storage. * **We do not train on your data.** Content sent to the API is never used to train, fine-tune, or evaluate models, and is never used for any purpose other than serving your request. * **We do not sell or share your data.** We are an inference provider, not a data business. * What we do keep is **operational telemetry** — token counts, latency, error rates, and similar service metadata — used solely to run a reliable service. *** ## Data we process ### Request and response content Prompts, messages, tool definitions, and model completions ("Content") are processed in GPU and host memory for the duration of the request. Content is not persisted to disk, not logged, not used for training or evaluation, and not shared with any third party. The serving engine maintains a transient **prefix cache** in GPU memory to speed up repeated request prefixes. This cache exists only in volatile memory, is evicted continuously under normal operation, and is destroyed entirely whenever the serving process restarts. ### Service metadata To operate, secure, and bill for the service, we collect metadata about API requests, which does not include Content: * Timestamps, HTTP status codes, and request latency * Model identifier, prompt/completion **token counts**, and cache-hit counts * Concurrency, queue depth, and hardware utilization metrics * API key identifier used to authenticate the request * Network-level connection metadata at our edge (see Subprocessors below) ### Diagnostic traces and logs Our reliability tooling (metrics, distributed traces, and application logs) records the request lifecycle — timing, sizes, and outcomes — for debugging and capacity planning. These systems are configured to capture service metadata, not Content, and are retained on the short schedules described in the [Data Retention Policy](/company-docs/legal/data-retention-policy). We do not currently advertise a formal **Zero Data Retention (ZDR)** designation. We will publish one only after our independent validation that no diagnostic system can capture request content is complete. *** ## What we never do with API data * Train, fine-tune, or benchmark models on Content * Sell data, or share it with advertisers or data brokers * Use Content for product analytics or human review, except where strictly required to investigate abuse or a security incident, and then only with access limited to authorized personnel * Retain Content after a request completes, beyond the volatile in-memory caches described above *** ## End users reaching us through a routing partner When you use FarmGPU through a router or aggregator (for example, OpenRouter), your account, identity, and billing relationship is with that partner. FarmGPU receives the request payload and partner-level authentication only — we do not receive your name, email address, account identity, or payment details, and we cannot associate requests with individual end users. The partner's own privacy policy governs the data they hold about you. ## Direct API customers For customers with a direct agreement, we hold the business contact and billing information needed to administer the account (names, email addresses, company details, and usage records for invoicing). This information is used only for account administration and is retained as described in the [Data Retention Policy](/company-docs/legal/data-retention-policy). *** ## Subprocessors | Subprocessor | Role | Data exposed | | ---------------- | ---------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Cloudflare, Inc. | TLS termination, DNS, and network perimeter for the API endpoint | Encrypted request traffic in transit; connection metadata (IP addresses, request URLs, timing). Cloudflare does not persist request bodies for us; caching is disabled for API responses. | Inference itself runs on **FarmGPU-owned hardware in FarmGPU-operated facilities in California, USA**. No model-serving or Content processing is outsourced to any third-party cloud. *** ## Security API access requires bearer-token authentication; unauthenticated requests are rejected. Serving infrastructure is operated under the controls described in our [Security documentation](/company-docs/security/compliance-standards), including our SOC 2 Type II program. Access to production systems is limited to authorized FarmGPU operations personnel. *** ## Legal disclosures We disclose data only if required by valid legal process, and only the minimum required. Because Content is not stored, historical prompts and completions cannot be produced — only the service metadata described above exists to be disclosed. ## Your rights and contact Depending on your jurisdiction (including the California Consumer Privacy Act and the EU/UK GDPR where applicable), you may have rights to access, correct, or delete personal information we hold about you. Because we hold no end-user identity for API traffic, such requests typically apply to direct-customer account information. Questions, requests, or concerns: **[privacy@farmgpu.com](mailto:privacy@farmgpu.com)**. ## Changes to this policy We will post updates to this page with a revised effective date. Material changes affecting partner integrations will be communicated to partners directly. # Service Level Agreements Source: https://docs.farmgpu.com/company-docs/legal/service-level-agreements FarmGPU SLA commitments, uptime guarantees, and service credits # Service Level Agreement ## Applicability of SLA The service level commitments, performance metrics, and remedies set forth in this Article 15 (collectively, the "SLA") apply to: * **Reserved Instance Services** * **Bare Metal Services** **On-Demand Services** and **Spot Instance Services** are provided without service level guarantees or commitments and are not subject to this SLA. **Beta Services** are expressly excluded from SLA coverage. ## Service Availability Commitment ### Monthly Uptime Percentage FarmGPU commits to providing the following **Monthly Uptime Percentage** for Customer's Reserved Capacity and Bare Metal Services, measured on a monthly basis: | SERVICE TIER | COMMITTED MONTHLY UPTIME % | TARGET MONTHLY UPTIME % | | ------------ | -------------------------- | ----------------------- | | Standard | 99.0% | 99.5% | | Premium | 99.7% | 99.9% | | Enterprise | 99.9% | 99.95% | **Bare Metal Services Default SLA Tier:** Unless otherwise specified in the Order Form, Bare Metal Services are provided with **Premium** SLA tier (99.7% committed uptime). Customer may upgrade to Enterprise tier for additional monthly fees as specified in the applicable Order Form. ### Calculation of Monthly Uptime Percentage Monthly Uptime Percentage = \[(Total Minutes in Month – Downtime Minutes) / Total Minutes in Month] × 100 ### Measurement Source of Truth Downtime start and end timestamps are determined from **FarmGPU's monitoring and telemetry systems**, which serve as the authoritative record for all SLA calculations. Relevant monitoring data for a disputed period will be shared with Customer on request. If Customer presents evidence of unavailability during a period where FarmGPU's monitoring has a gap or was itself unavailable, FarmGPU will consider Customer's records in good faith when validating the claim. ## Definition of Downtime "Downtime" means any period of time (measured in minutes) during which Customer's Reserved Capacity or Bare Metal Services are unavailable for use due to a failure **within FarmGPU's control**, as demonstrated by Customer's inability to: (a) Submit new jobs or workloads to Reserved Capacity or Bare Metal GPU instances through normal access methods; OR (b) Execute running jobs due to GPU hardware failure, system failure, or infrastructure failure affecting Reserved Capacity or Bare Metal Services; OR (c) Access Reserved Capacity or Bare Metal Services due to network connectivity failures within FarmGPU's control; OR (d) Access Reserved Capacity or Bare Metal Services through normal authentication and access control mechanisms due to failures in FarmGPU's systems; OR (e) For Bare Metal Services specifically: inability to access server management interfaces (IPMI/BMC) for remote administration. For avoidance of doubt, **performance degradation that does not materially prevent job submission or execution** does not constitute Downtime. ## Downtime Exclusions Downtime shall not include, and the Monthly Uptime Percentage calculation shall exclude, any periods of unavailability resulting from or caused by: (a) **Scheduled Maintenance:** Planned maintenance activities conducted during pre-announced maintenance windows, provided that such scheduled maintenance does not exceed four (4) hours in any calendar month; (b) **Customer-Caused Issues:** Customer's or Authorized Users' misconfigurations, software errors or bugs in Customer Content, credential management failures, or misuse of the GPU Services; (c) **Force Majeure Events:** Events beyond FarmGPU's reasonable control including but not limited to acts of God, natural disasters, wars, terrorist attacks, riots, labor disputes, pandemics, government orders, failures of upstream internet service providers or network carriers, public utility failures, and distributed denial-of-service (DDoS) attacks; (d) **Spot Instance Interruptions:** Reclamation or unavailability of Spot Instance capacity (Spot Instances are explicitly excluded from SLA coverage); (e) **Policy Violations:** Service suspensions or terminations resulting from Customer's violation of the Acceptable Use Policy or other material breaches of this Agreement; (f) **Non-Payment Suspensions:** Service suspensions due to Customer's failure to pay undisputed amounts when due; (g) **Partial Degradation:** Performance degradation, reduced throughput, or increased latency that does not materially prevent job submission or execution; (h) **De Minimis Outages:** Individual hardware failures or component failures that are fully remediated within **sixty (60) minutes**; (i) **Beta and Pre-Release Services:** Any Beta Services as defined in Section 2.5; (j) **Bare Metal Specific Exclusions:** * Operating system failures, kernel panics, or software crashes in Customer-managed software layers * Issues related to Customer-installed applications or middleware * Network issues originating from Customer-configured network settings, firewall rules, VLAN/private segment settings, or routing policies * Self-induced hardware issues resulting from Customer exceeding published thermal, power, or operating specifications ## Capacity Availability Guarantee ### Minimum Available Capacity For Reserved Instance and Bare Metal Services, FarmGPU guarantees that at least **ninety-five percent (95%)** of Customer's committed GPU capacity shall be available for Customer's use during each calendar month, excluding scheduled maintenance windows and Force Majeure events. This guarantee applies only to unavailability caused by conditions within FarmGPU's control and excludes partial degradation that does not fully prevent use of affected GPU resources. ### Calculation of Capacity Availability Capacity Availability Percentage = (Available GPU-Hours / Total Reserved GPU-Hours) × 100 Measured on a monthly basis, where: * Available GPU-Hours = GPU-hours of reserved/committed capacity available for Customer use during the month * Total Reserved GPU-Hours = total GPU-hours of reserved/committed capacity under the Order Form for that month ### Bare Metal Hardware Availability For Bare Metal Services, availability is measured on a per-server basis. If a Bare Metal server is completely unavailable, it counts as 100% unavailable for that time period. Partial loss (e.g., one GPU failed in an 8-GPU server) may be calculated proportionally **only if** the remaining system cannot support workload execution for the contracted configuration. ## Hardware Replacement Service Level (Targets) FarmGPU will use commercially reasonable efforts to meet the following **target** response and resolution times for hardware failures affecting Customer's Reserved Capacity and Bare Metal Services: | ISSUE TYPE | TARGET RESOLUTION TIME – BARE METAL | | ------------------------------ | ----------------------------------- | | Individual GPU failure | 3 business days | | Complete node failure | 2 weeks | | Network infrastructure failure | 48 hours | | Storage device failure | 48 hours | | Memory failure | 3 business days | | Power supply failure | 48 hours | Where on-site spare hardware is available for the affected component, FarmGPU will use commercially reasonable efforts to resolve faster than these targets. Where resolution requires an OEM RMA cycle, the target runs from FarmGPU's receipt of the replacement part, and FarmGPU will provide equivalent replacement capacity in the interim where available. **Target Resolution Time** means the time from detection/notice (whichever is earlier) until service is restored or equivalent replacement capacity is provisioned. Target resolution times are subject to site access, safety constraints, and hardware availability. Initial response times are governed by the severity matrix below. ## Support Response and Severity Classification FarmGPU provides 24×7 incident response for Reserved Capacity and Bare Metal Services against the following severity matrix: | PRIORITY | DESCRIPTION | INITIAL RESPONSE | TARGET FIX OR WORKAROUND | | -------- | ------------------------------------------------------------------ | ---------------- | ------------------------ | | P1 | Service or node down or inaccessible; no workaround | 30 minutes | 8 hours | | P2 | Service operational with substantial errors or degraded capacity | 4 hours | 1 business day | | P3 | Errors impairing full benefit of the service; workaround available | 8 hours | 2 business days | | P4 | Minor errors, cosmetic issues, general questions | 1 business day | 3 business days | **Initial Response Time** means the time from FarmGPU's detection of the failure or receipt of Customer notice (whichever is earlier) until technical personnel begin active diagnosis. **Severity classification:** initial classification of an incident's priority is made by FarmGPU. If Customer disagrees with a classification, Customer may escalate through the support escalation path, and the parties will confer in good faith to resolve the classification promptly. ## Service Credits ### Service Availability Service Credits If FarmGPU fails to achieve the committed Monthly Uptime Percentage for any calendar month, Customer shall be eligible for Service Credits as follows: **Standard Tier (99.0% commitment):** | ACHIEVED MONTHLY UPTIME % | SERVICE CREDIT (% OF MONTHLY FEES) | | ------------------------- | ---------------------------------- | | 99.0% to \< 99.5% | 10% | | \< 99.0% | 25% | **Premium Tier (99.7% commitment):** | ACHIEVED MONTHLY UPTIME % | SERVICE CREDIT (% OF MONTHLY FEES) | | ------------------------- | ---------------------------------- | | 99.5% to \< 99.7% | 15% | | 99.0% to \< 99.5% | 30% | | \< 99.0% | 50% | **Enterprise Tier (99.9% commitment):** | ACHIEVED MONTHLY UPTIME % | SERVICE CREDIT (% OF MONTHLY FEES) | | ------------------------- | ---------------------------------- | | 99.7% to \< 99.9% | 15% | | 99.5% to \< 99.7% | 30% | | \< 99.5% | 50% | ### Capacity Availability Service Credits If Capacity Availability Percentage falls below 95% in any calendar month: | ACHIEVED CAPACITY AVAILABILITY % | SERVICE CREDIT (% OF MONTHLY FEES) | | -------------------------------- | ---------------------------------- | | 90% to \< 95% | 10% | | 85% to \< 90% | 25% | | \< 85% | 35% | ### Bare Metal Enhanced Credits For Bare Metal Services only, if a critical hardware replacement exceeds the applicable **Target Resolution Time** by more than 50% due to conditions within FarmGPU's control, Customer receives additional Service Credits: | RESOLUTION TIME OVERAGE | ADDITIONAL SERVICE CREDIT | | --------------------------- | ------------------------- | | 50% to \< 100% over target | 5% | | 100% to \< 200% over target | 10% | | 200%+ over target | 15% | Enhanced credits are cumulative with availability credits, subject to the Service Credit Caps below. ## Service Credit Calculation; Caps Service Credits are calculated as a percentage of the monthly recurring fees paid or payable for the affected Reserved Capacity or Bare Metal Services that experienced the SLA failure. **Service Credit Caps:** * Maximum Service Credits in any single calendar month: **50%** of the total monthly fees for the affected Order Form (equal to the highest single credit tier; enhanced credits are cumulative only up to this cap) * Maximum aggregate Service Credits over any twelve (12) month period: **30%** of the total contract value under the affected Order Form ## Chronic Failure Termination Right If, for Customer's applicable service tier: (a) Monthly Uptime Percentage falls below the **Committed Monthly Uptime Percentage** for that tier for two (2) consecutive calendar months; OR (b) Monthly Uptime Percentage falls more than **0.5 percentage points below** the Committed Monthly Uptime Percentage for that tier in any single calendar month (e.g., below 98.5% on the Standard tier, below 99.2% on the Premium tier, below 99.4% on the Enterprise tier); Customer may, upon forty-five (45) days' prior written notice, terminate the affected Order Form without early termination fees. FarmGPU shall refund any prepaid fees allocable to services not yet rendered, pro-rated as of the effective termination date. This right does not apply where the failure is caused by events constituting Downtime Exclusions under Section 15.4. During the notice period, FarmGPU may demonstrate that the root cause has been remediated through structural improvements. If FarmGPU demonstrates remediation to Customer's reasonable satisfaction, Customer may withdraw the termination notice. ## Service Credit Request Procedure To receive Service Credits, Customer must: (a) Submit a claim within thirty (30) days after the end of the month of the alleged violation; (b) Include affected resources, description of issue, supporting evidence, and Customer's downtime calculation; (c) Failure to submit within this period waives the claim. FarmGPU will review claims within fifteen (15) business days and, if validated, apply credits within thirty (30) days. **SERVICE CREDITS ARE CUSTOMER'S SOLE AND EXCLUSIVE REMEDY** for any failure to meet this SLA. ## Application of Service Credits Service Credits: * Are applied against future invoices for the same Order Form * Have no cash value and are not refundable * Expire upon termination or expiration of the Order Form * Are not transferable ## Monitoring, Reporting, and Transparency FarmGPU will provide: * A real-time status dashboard * Monthly SLA reports (within five (5) business days after month-end) * RCA within forty-eight (48) hours for Downtime events lasting > 30 minutes * API access to metrics where available * Bare Metal monitoring access including BMC/IPMI where applicable ## Scheduled Maintenance (a) At least seven (7) days' advance notice for scheduled maintenance impacting applicable services (b) Scheduled maintenance not to exceed four (4) hours in any calendar month per affected cluster/zone/server (c) FarmGPU will use commercially reasonable efforts to perform non-disruptive maintenance when feasible (d) Emergency maintenance permitted when necessary for security, integrity, or stability, with notice as practicable ## SLA Exclusions and Limitations This SLA does not apply to: * Beta/free trial services * On-Demand and Spot services * Suspended services due to Customer breach or non-payment * Unsupported or non-recommended configurations * Bare Metal issues arising from Customer-managed OS/software or Customer-induced hardware stress beyond published specifications ## SLA Modifications FarmGPU may modify this SLA upon ninety (90) days' prior notice, provided: * Changes do not apply to existing Order Forms during their term without Customer consent * Changes apply to new Order Forms and renewals * Changes do not materially diminish commitments for active reserved/bare metal terms # User Agreements and Policies Source: https://docs.farmgpu.com/company-docs/legal/user-agreements-policies FarmGPU terms of service and customer agreements # User Agreements & Policies FarmGPU operates with clear, transparent agreements designed for both individual builders and strategic partners. ## 1. Master Service Agreement (MSA) The foundational agreement for all FarmGPU customers. It covers: * Terms of usage and billing. * Intellectual property (IP) protection and data privacy. * Liability and indemnification. * Governing law and dispute resolution. ## 2. GPU Owner Agreement For revenue share partnerships and managed infrastructure services, FarmGPU provides specific agreements covering hardware ownership, operational responsibilities, and revenue sharing terms. ## 3. Inference API Policies Data handling for the FarmGPU LLM inference API is governed by the [Privacy Policy](/company-docs/legal/privacy-policy) and the [Data Retention Policy](/company-docs/legal/data-retention-policy). # Products & Services Source: https://docs.farmgpu.com/company-docs/products-services/index FarmGPU's GPU infrastructure offerings ## Products & Services FarmGPU delivers GPU infrastructure in two primary ways: **on-demand cloud access** and **managed GPU infrastructure**. Both offerings are designed to give customers high-performance AI compute without the complexity, cost, or lock-in of hyperscaler clouds. *** ## On-Demand GPU Cloud (via RunPod) FarmGPU operates high-performance GPU clusters that are available on-demand through our platform partnerships, including **RunPod**. This offering is designed for: * AI startups and developers * Researchers and experimentation teams * Fine-tuning and inference workloads * Customers who need instant access without long-term commitments ### Key Characteristics * On-demand access to GPU servers * Hourly or usage-based pricing * Fast provisioning and teardown * Bare-metal or near-metal performance * No long-term contracts or quotas ### What Customers Get * Direct access to modern GPUs optimized for AI workloads * Predictable performance without hyperscaler abstractions * Transparent pricing aligned to actual infrastructure costs * A simple path from experimentation to production FarmGPU's on-demand cloud provides the **speed and flexibility of cloud**, with the **performance and economics of purpose-built infrastructure**. *** ## Dedicated Bare-Metal GPU Rentals (Long-Term) For customers with sustained workloads, FarmGPU offers **dedicated, long-term bare-metal GPU rentals**. This product is designed for: * Training large models * Continuous inference workloads * Enterprises seeking predictable cost and performance * Customers transitioning off hyperscalers ### Key Characteristics * Fully dedicated GPU servers or clusters * Monthly or multi-year terms * Customer-controlled software stack * No noisy neighbors * Fixed, predictable pricing ### What Customers Get * Maximum GPU utilization and performance * Full control over operating systems, drivers, and frameworks * Stable costs independent of cloud pricing changes * Infrastructure that behaves like owned hardware—without the CapEx This offering combines the **control of ownership** with the **flexibility of leasing**, making it ideal for production AI systems. *** ## Managed GPU Infrastructure (Customer-Owned or Partner-Owned) FarmGPU also operates and manages GPU clusters on behalf of customers who **own or control their own hardware** but lack the expertise or resources to run it efficiently. This product is designed for: * Data center operators with unused GPU capacity * Enterprises that have purchased GPUs but lack AI infrastructure expertise * Investors or consortiums deploying GPU capital * Organizations seeking on-prem or hybrid AI deployments ### Scope of Management FarmGPU provides end-to-end operational management, including: * Cluster bring-up and configuration * Performance tuning and optimization * Storage and networking integration * Monitoring, telemetry, and alerting * Security hardening and isolation * Lifecycle management and upgrades ### What Customers Get * Faster time to production * Higher GPU utilization * Reduced operational risk * A professionally operated AI infrastructure stack * Optional integration into marketplace or on-demand offerings In effect, FarmGPU turns **idle or underutilized GPUs into production-grade AI infrastructure**. *** ## Choosing the Right Offering | Customer Need | Recommended Product | | ------------------------------------- | -------------------------------- | | Short-term or experimental workloads | On-Demand GPU Cloud | | Long-running or production workloads | Dedicated Bare-Metal GPU Rentals | | Owned GPUs without internal expertise | Managed GPU Infrastructure | | On-prem or hybrid AI deployments | Managed GPU Infrastructure | *** ## Summary FarmGPU does not sell software licenses or abstract cloud services. We sell **high-performance GPU infrastructure**, delivered in the form that best matches each customer's needs: * On-demand access when flexibility matters * Dedicated infrastructure when performance and cost predictability matter * Managed operations when expertise is the bottleneck This product strategy allows FarmGPU to serve the full lifecycle of AI workloads—from experimentation to production—without forcing customers into a single deployment model. # Partners Source: https://docs.farmgpu.com/company-docs/products-services/partners How FarmGPU works with our partners ## Partners We work with a global set of partners across infrastructure, platforms, and software. ### Systems partners Integrate high-performance, scalable hardware solutions to optimize and accelerate AI and data-intensive workloads. Solidigm logo
We host the AI Central Lab where we bring the latest NVIDIA GPUs, 800G Ethernet and InfiniBand fabrics, and storage platforms together to test the latest storage ISVs in AI workloads. Record performance in the ClusterMax storage and MLPerf Storage.
[Solidigm AI Central Lab](https://www.solidigm.com/products/technology/ai-central-lab-for-ai-research.html) Intel logo
We are an Intel CPU and Intel Gaudi partner, hosting AI workloads and inference endpoints on Intel platforms.
[Intel AI](https://www.intel.com/content/www/us/en/artificial-intelligence/overview.html) NVIDIA logo
We are an NVIDIA partner using the latest NVIDIA GPUs, including A100, H100, H200, B200, and RTX Pro 6000 Blackwell.
[NVIDIA data center GPUs](https://www.nvidia.com/en-us/data-center/) AMD logo
We use high-end AMD CPUs for both our storage and GPU servers.
[AMD EPYC processors](https://www.amd.com/en/products/server-processor) Lenovo logo
Lenovo is one of our strongest OEM partners. We deploy and validate Lenovo GPU and storage solutions.
[Lenovo servers and storage](https://www.lenovo.com/us/en/servers-storage/) Supermicro logo
We deploy and validate Supermicro GPU and storage platforms for reliable production-grade clusters.
[Supermicro GPU solutions](https://www.supermicro.com/en/solutions/gpu) MSI logo
We are a partner and distributor, building custom bare-bones systems with the latest NVIDIA MGX platforms.
[MSI enterprise products](https://www.msi.com/Enterprise-Products) ### Cloud partners Deliver secure, on-demand access to FarmGPU-managed capacity with fast provisioning and predictable performance. Runpod logo
Our secure cloud partner for US-CA-2, enabling instant access to FarmGPU capacity for short-term workloads.
[Runpod](https://runpod.com/) ### Network partners Build scalable fabrics with automation and integration for efficient cluster deployments. Celestica logo
We deploy DS5000 Ethernet switches with Rocky B2 across our backend fabric as part of our Instant Clusters.
[Celestica](https://www.celestica.com/) Hedgehog logo
Hedgehog is the software that automates our network fabric and operations.
[Hedgehog](https://hedgehog.cloud/) ### ISV partners Accelerate data pipelines and AI workflows with storage and software systems built for throughput and scale. VAST Data logo
Delivers high-throughput, low-latency data pipelines that keep GPUs fully utilized.
[VAST Data](https://www.vastdata.com/) MinIO logo
Provides S3-compatible object storage that integrates cleanly with our AI infrastructure stack.
[MinIO](https://min.io/) Metrum AI logo
Partners on AI stack enablement and customer deployment success.
[Metrum AI](https://www.metrum.ai/) ### Liquid cooling partners Increase GPU density and efficiency with liquid-cooled system design and integration. Hypertec logo
Hypertec is a leader in immersion cooling, and we deploy immersion at scale with edge partners.
[Hypertec](https://hypertec.com/) 2CRSi logo
We scale large deployments with managed GPU clusters and inference endpoints.
[2CRSi](https://www.2crsi.com/) # Compliance and Standards Source: https://docs.farmgpu.com/company-docs/security/compliance-standards FarmGPU security certifications and compliance framework # Compliance and Standards FarmGPU's security posture is built on **alignment with established standards**, clear separation of responsibilities, and a roadmap toward formal certifications as the company scales. We prioritize **practical security controls, auditability, and transparency** over checkbox compliance. *** ## Compliance Philosophy FarmGPU follows a **shared responsibility model**: * **FarmGPU** is responsible for the security of the platform, management plane, and operational processes. * **Data center partners** are responsible for physical security, facility controls, and environmental resilience. * **Customers** retain control over their workloads, data, and application-level security. This model aligns with industry best practices for infrastructure providers and enables enterprise customers to clearly map controls to risk ownership. *** ## Core Security Certifications & Alignment ### SOC 2 Type II (In Progress) FarmGPU is currently in the audit window for **SOC 2 Type II**, covering the following Trust Service Criteria: * **Security** * **Availability** * **Confidentiality** **Status:** In progress **Target completion:** **Q2 2026** ### ISO/IEC 27001 (Data Center Alignment) FarmGPU operates within data centers that are **ISO/IEC 27001 certified** or aligned, ensuring that physical facilities adhere to globally recognized Information Security Management System (ISMS) standards. This includes: * Controlled physical access * Documented security procedures * Formal risk management and incident handling processes FarmGPU leverages these certified environments as part of its overall security posture, while maintaining separate logical and operational controls at the platform level. ### Tier III+ Data Center Standards FarmGPU deploys infrastructure in **Tier III or Tier III+ data centers**, providing: * N+1 or 2N redundancy across power, cooling, and network paths * Concurrent maintainability * High availability suitable for production AI workloads **Facility availability target:** 99.99% These standards ensure infrastructure reliability while reducing correlated failure risk across clusters. # Confidential Computing Source: https://docs.farmgpu.com/company-docs/security/confidential-computing FarmGPU's confidential AI stack with hardware-backed isolation # Confidential Computing FarmGPU is **among the first NeoCloud providers** to deliver a full **Confidential AI** stack that spans **CPU + GPU**, enabling customers to run sensitive workloads with **hardware-backed isolation** and **cryptographic proof** of platform integrity. Confidential Computing moves security beyond "data at rest" and "data in transit" to **data in use**—protecting models, prompts, weights, and sensitive datasets even while being processed. This matters because modern AI workloads often contain high-value IP (model weights, fine-tuning data, inference prompts) and regulated data (PHI, PII, financial records) that cannot be safely exposed to platform operators or other tenants. Our confidential computing architecture is built on: * **Intel® Xeon® (Granite Rapids) + Intel® Trust Domain Extensions (TDX)** for CPU-side trusted execution (confidential VMs) * **NVIDIA Blackwell (B200) Confidential Computing** capabilities for GPU-side secure execution and encrypted high-performance multi-GPU configurations * **Remote Attestation** workflows so customers can verify trust *before* deploying sensitive artifacts ## What "Confidential AI" Means at FarmGPU Confidential AI at FarmGPU is a **system-level trust boundary** designed to protect against: * Cloud operator / hypervisor visibility into tenant workloads * Host OS tampering * Firmware rollback and "known vulnerable" platform states This is achieved by running workloads in a **Confidential Virtual Machine (CVM)** backed by Intel TDX and attaching GPUs operating in confidential-capable modes where the platform can enforce isolation and supply attestation evidence. ## The Trusted Execution Environment: CPU + GPU as One Stack ### Intel® TDX on Granite Rapids: Hardware-Enforced VM Isolation Intel TDX is designed to protect a tenant VM (a "Trust Domain") from a malicious or compromised host/hypervisor by providing **hardware-level isolation** and **memory encryption**, plus integrity protections and attestation. Key properties FarmGPU relies on: * **Confidentiality (Memory Encryption):** TD memory is encrypted so it is opaque to cloud operators and the host stack. * **Integrity:** Architectural mechanisms protect the TD's CPU state and memory mappings against tampering. * **Attestation:** Workload owners can verify platform configuration and policy before releasing secrets to the VM. This gives you a CPU-side "black box" environment where **even privileged host software cannot directly read or modify tenant memory**. ### NVIDIA Blackwell (B200): GPU Confidential Computing & Secure Multi-GPU NVIDIA's confidential computing capabilities for Hopper/Blackwell are designed to protect GPU workloads in a confidential VM context, including mitigations for software attacks, rollback, and certain physical snooping scenarios. A major advantage of Blackwell in multi-GPU systems is support for **encrypted NVLink pathways** in multi-GPU pass-through mode—meaning you can attach up to **eight GPUs** to a single confidential VM while keeping GPU-to-GPU traffic encrypted within the node. FarmGPU's Blackwell confidential configuration emphasizes: * **GPU passthrough to a CVM:** GPUs are assigned exclusively to a tenant VM (not shared across VMs in passthrough mode). * **Encrypted NVLink in Blackwell multi-GPU passthrough:** Encrypted GPU-to-GPU communication helps protect collective training traffic and intermediate tensors at high bandwidth. * **CPU ↔ GPU data protection options:** NVIDIA describes CPU-GPU traffic encryption approaches including software encryption via bounce buffers, and in some configurations using standards-based link protection (e.g., IDE/TDISP) when supported across CPU/GPU/firmware/driver stack. ## Remote Attestation: "Trust, but Verify" Before You Upload Secrets Security isn't just isolation—it's **proof**. FarmGPU supports remote attestation so customers can validate: * The node is genuine hardware * Firmware/driver stack is authentic and not rolled back * Confidential modes and policies are enabled as expected ### NVIDIA Attestation Suite NVIDIA provides an attestation framework intended to cryptographically verify the authenticity and integrity of NVIDIA hardware and firmware, integrating with confidential computing workflows. The suite includes **NRAS**, **RIM Service**, and **OCSP Service**. This lets a verifier validate "claims" about the GPU platform state prior to allowing sensitive workloads to proceed. ### Composite Attestation: CPU TEE + GPU TEE Intel's Trust Authority documentation describes composite attestation workflows where **Intel TDX evidence** and **NVIDIA GPU evidence** are collected and verified together, and GPU verification can be routed through NVIDIA's attestation service. Why composite matters: it ties together the CPU-side trusted VM environment and the GPU-side trusted execution claims into **one verifiable chain of trust**, which is stronger than "GPU-only" evidence. ## How Customers Use Confidential AI on FarmGPU A typical workflow looks like: 1. **Provision a confidential VM-enabled node or cluster** 2. **Request attestation evidence** (CPU + GPU where applicable) 3. **Verify evidence** against expected policies (firmware versions, modes enabled, no rollback) 4. **Release secrets only after verification** (model weights, encryption keys, regulated datasets, API credentials) 5. Run training / fine-tuning / inference with reduced trust assumptions about the cloud operator This model supports security-sensitive deployments where "trusting the provider" is not acceptable. ## Threat Model Clarity: What We Protect Against FarmGPU's confidential stack is designed to mitigate major classes of attacks in NVIDIA's CC threat framing (e.g., software attacks, rollback attacks, certain physical snooping). [NVIDIA Docs](https://docs.nvidia.com/nvidia-secure-ai-with-blackwell-and-hopper-gpus-whitepaper.pdf) As with all TEEs, some categories (e.g., sophisticated physical attacks, denial of service) are typically considered out of scope at the silicon level and must be managed via layered controls and operational practices. [NVIDIA Docs](https://docs.nvidia.com/nvidia-secure-ai-with-blackwell-and-hopper-gpus-whitepaper.pdf) We are explicit about scope during enterprise security review. ## High-Value Use Cases Confidential AI unlocks new classes of workloads: * **Healthcare & Life Sciences:** train/fine-tune on sensitive patient or genomic data while reducing exposure risk * **Financial Services:** fraud detection, trading signals, and risk models on sensitive transaction streams * **Sovereign AI & Regulated Environments:** strong isolation for national labs, defense-adjacent research, and data residency-driven deployments * **Model IP Protection:** protect high-value proprietary weights and inference prompts from infrastructure operators and third parties In short: **run frontier-grade AI on sensitive data with hardware-backed proof of isolation**—not contractual promises. # Linux Security Hardening Source: https://docs.farmgpu.com/company-docs/security/linux-security-hardening TractorOS security controls and Linux hardening practices # Linux & OS Security (TractorOS) FarmGPU's operating system, **TractorOS**, is built on established Linux security best practices and extended with targeted hardening techniques designed for AI infrastructure. The goal is to **reduce attack surface, enforce consistency, and limit blast radius**, while preserving performance for GPU- and storage-intensive workloads. *** ## Secure OS Baseline & Attack Surface Reduction TractorOS follows a **minimal, hardened Linux baseline** designed to reduce unnecessary exposure. Key practices include: * Minimal OS footprint with only required services enabled * Removal of unused packages, daemons, and subsystems * No interactive user accounts on production nodes * Controlled administrative access via centralized tooling only These practices reduce the number of reachable services and limit opportunities for lateral movement or privilege escalation. *** ## Kernel & Runtime Hardening The Linux kernel and runtime configuration are hardened using widely adopted security controls. Key measures: * Conservative kernel configuration aligned with security best practices * Signed kernel modules and restricted module loading * Hardened sysctl settings (memory protection, process isolation, ptrace restrictions) * Enforced memory protections (ASLR, NX, KPTI where applicable) Security settings are evaluated alongside performance requirements to avoid negatively impacting GPU or storage throughput. *** ## Storage & Device Security TractorOS treats storage and hardware devices as **security-sensitive resources**, not just performance components. Controls include: * Enforced device isolation for GPUs, NVMe, and NICs * No direct access to raw block devices by tenant workloads * Hardware-backed encryption at rest (SEDs) enabled by default * Secure device lifecycle handling, including cryptographic erase before reuse These measures help ensure that data remains protected even in multi-tenant and hardware-reuse scenarios. *** ## TractorOS-Specific Hardening Techniques In addition to standard Linux practices, TractorOS incorporates several platform-specific controls: * **Immutable OS images:** The base OS is read-only and updated atomically, preventing configuration drift and persistent compromise. * **Deterministic provisioning:** Nodes are provisioned from known-good images with repeatable configuration. * **Explicit trust boundaries:** Clear separation between the host OS, management services, and customer workloads. * **Observable state:** OS state and system events are continuously monitored and logged for audit and incident response. These techniques enable FarmGPU to maintain **consistent security posture across fleets**, while supporting rapid recovery and rollback if issues arise. # Storage Security Source: https://docs.farmgpu.com/company-docs/security/storage-security FarmGPU storage encryption, data protection, and security controls ## Advanced Encryption Architecture FarmGPU implements encryption **by default**, with no performance penalty and no dependency on application-level controls. ### Data at Rest — Self-Encrypting Drives (SEDs, TCG OPAL) * **Hardware-Based Encryption:** All Solidigm NVMe SSDs deployed in FarmGPU infrastructure support **TCG OPAL 2.0** compliant self-encrypting drives (SEDs). * **Always-On Encryption:** Data is encrypted transparently at the controller level and does not rely on OS or application configuration. * **Zero Performance Impact:** Encryption is handled entirely within the drive controller, ensuring **no reduction in GPU, CPU, or I/O throughput**. * **Strong Key Isolation:** Encryption keys never leave the device, preventing exposure through host compromise. This approach ensures that data remains protected even in the event of physical access to drives. ### Data in Transit — DPU-Accelerated Encryption FarmGPU secures inter-node and storage traffic using **hardware-offloaded encryption**. * **DPU-Accelerated IPsec / TLS 1.3:** NVIDIA **BlueField-3 DPUs** offload encryption and decryption from host CPUs, protecting data in flight across the storage and AI fabric. * **Line-Rate Security:** Encryption is performed at hardware speed, preserving full utilization of **400 Gbps and 800 Gbps** network fabrics. * **Reduced Attack Surface:** By removing cryptographic processing from the host OS, DPUs reduce exposure to kernel-level attacks and misconfiguration. This ensures secure communication between GPUs, storage systems, and cluster services without introducing latency or bottlenecks. ## Media Sanitization & Data Destruction (IEEE 2883) FarmGPU follows **standards-based media sanitization**, ensuring that data is irreversibly destroyed when no longer required. * **IEEE 2883-2022 Alignment:** FarmGPU adheres to IEEE 2883 guidance for secure sanitization of solid-state storage devices. * **Cryptographic Erase (CE):** Upon tenant offboarding, drive reassignment, or hardware retirement, FarmGPU performs **cryptographic erase** by destroying the media encryption keys on SEDs. * **Instant and Verifiable:** CE renders all data permanently unrecoverable within seconds, without the wear or uncertainty associated with overwrite-based methods. This approach enables: * Secure multi-tenant reuse of hardware * Compliance with enterprise and regulatory data-handling expectations * Environmentally responsible lifecycle management of SSDs ## Storage ISV Partner Security FarmGPU works with **enterprise-grade storage ISVs** that implement security as a core design principle. Storage platforms are deployed using **vendor-recommended hardening and encryption configurations**, and validated through FarmGPU's internal benchmarking and qualification process. ### VAST Data (Example) VAST Data provides a unified storage platform designed for large-scale AI workloads, with security features embedded at the architecture level. Key security characteristics include: * **End-to-End Encryption:** Data is encrypted at rest and in transit using industry-standard cryptographic algorithms. * **Key Management Integration:** Support for centralized key management and separation of duties. * **Multi-Tenant Isolation:** Logical isolation mechanisms ensure separation between tenants, datasets, and workloads. * **Secure Metadata Handling:** Metadata services are protected with the same security controls as data paths. * **Operational Hardening:** Role-based access control (RBAC), audit logging, and secure administrative interfaces. FarmGPU deploys VAST Data in accordance with documented best practices, ensuring encryption, access control, and isolation are consistently enforced. # Brand Source: https://docs.farmgpu.com/company-docs/team-culture/brand FarmGPU brand kit and AI-ready tokens ## Brand kit downloads Primary logos and lockups for print and web. Official brand rules, colors, typography, and usage. ## AI-ready brand tokens (minimal) Use this compact set for apps, themes, and agents that need consistent FarmGPU styling without the full guide. ```json theme={null} { "brand": "FarmGPU", "colors": { "background": "#091000", "text": "#FAFFF5", "accent": "#88FF00", "accentDark": "#529900", "accentLight": "#A0FF33", "surface": "#1A1C17", "surfaceAlt": "#272A22", "border": "#34382E" }, "fonts": { "sans": "\"Spline Sans\", system-ui, -apple-system, BlinkMacSystemFont, \"Segoe UI\", sans-serif", "mono": "\"Spline Sans Mono\", ui-monospace, SFMono-Regular, Menlo, Monaco, Consolas, \"Liberation Mono\", \"Courier New\", monospace" }, "radii": { "sm": "6px", "md": "10px", "lg": "16px" }, "spacing": { "xs": "4px", "sm": "8px", "md": "16px", "lg": "24px", "xl": "32px" }, "buttons": { "primary": { "bg": "#88FF00", "text": "#091000" }, "ghost": { "bg": "transparent", "text": "#88FF00", "border": "#88FF00" } } } ``` ## Usage notes * Default background is near-black; use volt accent for primary actions only. * Text should be white or stone-grey; avoid pure grey on dark backgrounds. * Use Spline Sans for UI and Spline Sans Mono for metrics/logs. # Culture Source: https://docs.farmgpu.com/company-docs/team-culture/culture FarmGPU company values and operating principles # Culture FarmGPU is built for a world where infrastructure is expensive, latency is real, and reliability is earned. We optimize for **leverage**, **speed**, and **truth**—with a small team operating systems that cannot afford hand-waving. ## 1) AI-First in Everything We treat AI as a multiplier across engineering, operations, security, GTM, and finance. Our default is to replace repetitive work with automation and to use AI to compress feedback loops. **How this shows up** * AI-assisted incident triage (summaries, hypothesis generation, diffing changes) * Auto-updated runbooks and SOPs from postmortems * Faster documentation, RFP responses, and security questionnaires (human-reviewed) * Internal tools that eliminate manual work **Guardrail:** AI increases speed, but **humans remain accountable** for correctness and customer impact. *** ## 2) High Agency & Extreme Ownership We use "agency" as: **the ability to independently identify problems, make decisions, and execute** without waiting for permission—while being accountable for outcomes. **How this shows up** * See a bottleneck? You drive it end-to-end: diagnose → fix → validate → document. * Strong opinions, loosely held: decide quickly, update beliefs faster. * We reward people who "close the loop," not people who create meetings. *** ## 3) Builders, Not Tourists We build and operate real infrastructure. Everyone is expected to be close enough to the work to understand failure modes and performance tradeoffs. ## 4) Systems Thinking AI infrastructure is a system. We value people who can reason across layers: silicon → firmware → kernel → network → storage → workload performance. **How this shows up** * We optimize for **goodput** (useful work delivered), not vanity metrics * Changes require a hypothesis + measurement plan * Postmortems focus on systemic fixes, not blame ## 5) Radically Transparent We share context broadly because small teams move faster when they have the full picture. **How this shows up** * Key metrics are visible internally (utilization, uptime, MFU, capacity, economics) * Clear written decisions: what we decided, why, what we'll measure * Fast postmortems with tracked action items **Guardrail:** We protect customer confidentiality and sensitive security details. ## 6) Direct Communication (Kind, Clear, Fast) We value direct communication because speed and reliability require clarity. We aim to be **straightforward, low-drama, and solution-oriented**—especially when stakes are high (incidents, customer impact, security). **What it means** * Say the thing that needs to be said, early. * Prefer facts, logs, and measurements over opinions. * Disagree openly, then commit and execute. **How this shows up** * Short, explicit asks: "I need X by Y to unblock Z." * Clear incident updates: "Impact, current hypothesis, next steps, ETA for next update." * Written decisions when it matters: what we chose, why, and what would change our mind. **Guardrails** * Direct ≠ rude. We default to **respect and professionalism**. * Critique ideas and systems, not people. * If you spot a problem, propose a path forward. # Team Structure Source: https://docs.farmgpu.com/company-docs/team-culture/team-structure FarmGPU leadership team and advisors # Team Structure ### Jonmichael Hands (JM), CEO As CEO of FarmGPU, JM is pioneering decentralized AI infrastructure and sustainable GPU hosting. A storage market authority, blockchain advocate, and sustainability champion with 17+ years in tech infrastructure, he spent a decade at Intel leading data center SSD strategies and chaired NVMe marketing. Formerly VP of Storage at Chia Network, JM is active in the Open Compute Project storage/sustainability community and serves as board chair and president for the Circular Drive Initiative, promoting secure drive reuse. He began his career designing storage arrays at Sun Microsystems. ### Lloyd Ross, COO Lloyd is the visionary founder of Data Slayer, a premier IT asset disposition (ITAD) firm. With over 22 years navigating data security and environmental sustainability in ITAD, this seasoned Silicon Valley veteran leverages deep industry connections for top-tier consulting. Under his leadership, Data Slayer is renowned for secure, responsible IT asset management and is a proud member of the Circular Drive Initiative (CDI). ### Patrick Harrington, CBO Patrick is a serial entrepreneur skilled in technology and cloud services. His ventures include pioneering digital advertising with Digital Dynamite, advancing mobile data transmission at TSE-N Metro, founding private cloud provider AxisFlow Inc., and co-founding Wi-Fi Rail to connect transportation systems to data centers. He expertly blends cloud technology and business development to drive industry growth. ### Luke Macken, Head of Engineering Distinguished software engineer and open-source leader with 20+ years of low-level expertise spanning Linux, distributed systems, security, and ML/AI infrastructure. Career highlights: engineered mission-critical avionics software for the U-2 reconnaissance aircraft; architected and launched multiple production infrastructures from scratch; built low-latency crypto-exchange engines; led high-stakes incident response and forensics during major crypto breaches; former Principal Engineer at Red Hat/Fedora with an extensive open-source footprint and a portfolio of patents. Now driving engineering at FarmGPU: building next-gen datacenter automation, AI-optimized infrastructure, and a foundational NeoCloud operating system. ### John Hubbard, VP Network and Architecture Veteran cloud and infrastructure architect with 20+ years of experience in networking, storage, virtualization, and systems engineering. Spent 12 years at Intel as a solutions leader, specializing in hybrid-cloud design, ISV enablement, and reference architectures for persistent memory and SSD technologies. Published technical author and conference speaker and advisor to the State of California ## Advisors ### Gary Lamb Advisor & CTO for Emerging Technologies at iBridge Cloud Technologies. A technology executive with over two decades driving innovation in cloud, virtualization, and networking, Gary leads complex projects like VMware-to-Azure migrations and designs ultra-low latency optical networks. He previously shaped next-gen cloud architecture on the Cloud Advisor Council, founded NextCloud, co-founded AccessFlow, and led practices at IBM Internet Security Systems and 3Com. ### Steven Kaplan Advisor & technology thought leader with over 20 years of experience in IT infrastructure and cloud strategy. Formerly Vice President at Nutanix, where he built a global channel program and earned the moniker "ROI Dude," Steven now leads Global Channels at Tessell, driving cloud and data management innovation. An accomplished author and Forbes contributor, his insights on IT leadership and ROI have guided enterprises worldwide. Connect with him on LinkedIn or follow [@ROIdude](https://twitter.com/ROIdude) on Twitter for the latest in enterprise tech trends. ### Nilesh Shah Advisor to FarmGPU, and currently VP Business Development, ZeroPoint Technologies. Additionally. He participates and contributes regularly at standards bodies like SNIA, OCP, JEDEC, RISC-V, CXL Consortium. He is regularly invited to speak at Conferences, and has led multiple panels and featured in Analyst/ Press interviews, focused on AI and memory technologies. Previously, Nilesh led Strategic Planning at Intel Corporation's Non Volatile Memory Solutions Group, where he was was responsible for the product planning and launch of the Data Center SSD products and Pathfinding innovations. Nilesh advises several startups in the AI Data Center, memory and GPU space # Hardware Stack Source: https://docs.farmgpu.com/company-docs/technical-architecture/hardware-stack FarmGPU GPU, CPU, storage fabric, and networking infrastructure ## Hardware Stack FarmGPU's infrastructure is designed as a **cohesive system**, optimized for AI workloads from silicon through storage and networking. We prioritize **performance, efficiency, and forward compatibility**, enabling us to deploy the latest hardware while maintaining operational stability. ## GPUs: High-Performance AI Compute FarmGPU supports a broad range of modern NVIDIA GPUs to address both frontier and applied AI workloads. ### Supported GPU Platforms * **NVIDIA H100** * **NVIDIA H200** * **NVIDIA B200 (Blackwell)** * **NVIDIA RTX Pro 6000 Blackwell** This range allows us to support: * Large-scale model training * High-throughput inference * Fine-tuning and experimentation * Enterprise and workstation-class AI workloads Our GPU nodes are configured for **maximum utilization and predictable performance**, with platform designs aligned to real-world AI workload characteristics rather than generic cloud abstractions. ## CPUs: Balanced Host Compute FarmGPU pairs GPUs with modern, high-performance CPU platforms to ensure that host-side compute never becomes a bottleneck. ### Supported CPU Platforms * **AMD Turin** * **Intel Granite Rapids** These CPUs provide: * High core counts and memory bandwidth * Strong PCIe 5.0 connectivity * Balanced performance for data preparation, orchestration, and I/O-heavy tasks By supporting both AMD and Intel platforms, FarmGPU maintains flexibility across supply chains and customer requirements. ## Storage Fabric: AI-Optimized by Design FarmGPU treats storage as a **core system component**, not a peripheral service. Our storage architecture is designed to deliver **predictable, high-throughput data paths** that keep GPUs fed under real AI workloads. ### Storage Software Platforms We deploy and operate leading, production-grade storage platforms, including: * **MinIO** * **VAST Data** * **Weka** * **Ceph** All storage platforms are **benchmarked and validated** using FarmGPU's internal **Silo** benchmarking and evaluation suite to ensure consistent performance across training and inference workloads. ### Storage Hardware * **Solidigm SSDs across all platforms** * **PCIe 5.0 NVMe SSDs** in all GPU compute nodes * **High-density 122TB QLC SSDs** in storage nodes for cost-efficient scaling This combination enables both **high-performance hot data paths** and **economical cold and capacity tiers**, optimized for AI dataset growth. ### DPU-Accelerated Data Path (NVIDIA BlueField-3) A key differentiator in FarmGPU's storage architecture is the use of **NVIDIA BlueField-3 DPUs**. BlueField-3 DPUs offload critical infrastructure functions from the host CPU, including: * Storage protocol processing * Network virtualization * Encryption and security * Data movement orchestration By offloading these tasks, FarmGPU enables: * **GPU Direct Storage (GDS)** with near bare-metal throughput * Reduced CPU contention on GPU nodes * More predictable latency under load * Higher effective GPU utilization (MFU) In AI systems, performance is determined by the **end-to-end data path**. DPUs allow FarmGPU to optimize that path holistically, rather than relying on CPU-bound software layers. *** ## Networking: High-Bandwidth, Low-Latency Fabrics FarmGPU deploys a **dual-fabric networking architecture** to separate customer-facing traffic from backend AI communication. ### Front-End Network * **NVIDIA SN5600** Optimized for: * North–south traffic * Customer access * Control plane and service connectivity ### Back-End AI Fabric * **Celestica DS5000** * **Broadcom Tomahawk 5 ASIC** Optimized for: * East–west GPU communication * Storage traffic * Collective operations (e.g., NCCL) ### Fabric Management * **Hedgehog Open Network Fabric** Provides: * Open, vendor-neutral fabric management * Automated provisioning and monitoring * High observability and control ### RDMA Support * **RDMA over Converged Ethernet (RoCE)** * Tuned with **Rocky v2** This networking architecture ensures: * Predictable latency * High throughput at scale * Efficient GPU-to-GPU and GPU-to-storage communication ## Cooling: Today and Tomorrow ### Current Deployment * Air-cooled GPU servers deployed in existing and legacy data centers * Optimized airflow and rack density based on GPU thermal profiles ### Future Roadmap * Partnerships with **two major immersion cooling providers** * Initial immersion deployments planned for **2026** * Designed to support: * Higher rack densities * Improved power efficiency * Next-generation GPU thermal envelopes This phased approach allows FarmGPU to scale efficiently today while preparing for the thermal and power demands of future AI hardware. # Instant Clusters Source: https://docs.farmgpu.com/company-docs/technical-architecture/instant-clusters On-demand multi-node AI compute clusters with bare-metal performance ## Instant Clusters — On-Demand Multi-Node AI Compute FarmGPU offers **Instant Clusters** — high-performance, multi-node GPU clusters that users can provision in **minutes, not months**, via a self-serve, on-demand experience in partnership with RunPod. These clusters bring **bare-metal performance to cloud-like flexibility**, eliminating the traditional barriers of hardware procurement, rack build-out, and operational complexity. [FarmGPU Blog](https://blog.farmgpu.com/b200-cluster/?utm_source=chatgpt.com) ### Key Characteristics **One-Click Multi-Node Deployment** Users can launch powerful multi-node clusters from the RunPod dashboard with a few clicks. Instant Clusters support bring-your-own Docker containers or optimized templates for training, inference, and research workflows. **Pay-Per-Use Economics** Billing is per-second, not per-month, giving teams the flexibility to: * Run experiments without upfront commitments * Scale up for training runs and tear down when finished * Avoid minimum runtime requirements or termination fees **Complete Control** Users retain full control of their cluster lifecycle — start, stop, and reconfigure as needed — without long-term lock-in or opaque pricing structures. *** ### Architecture & Performance Optimizations Instant Clusters are built on the same high-performance infrastructure that powers FarmGPU's bare-metal offerings, including: **400 GB/s between nodes with RDMA Networking Fabric** Clusters leverage an OCP-compliant 800 Gbps backend fabric for GPU-to-GPU communication, delivering low-latency, high-bandwidth interconnects optimized for distributed training and collective operations. **Solidigm PCIe 5.0 NVMe Local Storage** Each node includes eight PCIe 5.0 NVMe SSDs, providing up to 116 GB/s of local storage throughput to ensure that model loading and data pipelines do not bottleneck training runs. [FarmGPU Blog](https://blog.farmgpu.com/b200-cluster/?utm_source=chatgpt.com) **Topology-Aware Communication** Clusters are optimized for high NCCL performance — the communications library that defines training throughput for large models — achieving multi-hundreds-GB/s AllReduce performance across GPU nodes. *** ### Typical Use Cases Instant Clusters are ideal for teams that need **elastic scale with the performance profile of bare-metal systems**, including: * **Distributed model training**, especially with PyTorch DDP, DeepSpeed, and other frameworks * **Large-scale fine-tuning and experimentation** * **High-performance inference workloads** * **Burst compute capacity without long-term commitments** This capability gives FarmGPU customers a flexible on-ramp to **production-grade AI compute** without the overhead of provisioning and managing their own cluster hardware. *** ### How Instant Clusters Fit the FarmGPU Platform Instant Clusters unify the **best aspects of cloud and bare metal**: * **Cloud-like experience:** Self-service deployment, per-second billing, fast iteration * **Bare-metal performance:** Optimized interconnects, local NVMe, and hardware-aware tuning * **Integration with your software stack:** Works seamlessly with FarmGPU's control plane, scheduling, and observability tools This makes them a natural complement to both the **on-demand GPU Cloud** and **managed infrastructure offerings**, enabling: * Rapid experimentation before committing to dedicated clusters * Burst capacity beyond long-term rentals * A consistent interface for both transient and persistent compute # Software Stack Source: https://docs.farmgpu.com/company-docs/technical-architecture/software-stack TractorOS, control plane, and FarmGPU operational software ## Software Stack & Control Plane FarmGPU's software stack is designed to **operate AI infrastructure at scale with minimal human overhead**, while preserving performance, transparency, and control. Rather than relying on proprietary cloud control planes, FarmGPU builds on **open-source, Linux-native primitives**, extended with purpose-built systems for GPU operations. Together, these components form a **unified control plane** for provisioning, operating, observing, and optimizing AI clusters. *** ## TractorOS — The NeoCloud Operating System **TractorOS** is FarmGPU's minimal, immutable, container-native operating system optimized specifically for AI compute. TractorOS Architecture ### What TractorOS Does * Provides a hardened, reproducible OS image for GPU and storage nodes * Ships with **built-in GPU, networking, and storage drivers** * Enables **zero-touch provisioning and fleet-wide upgrades** * Eliminates configuration drift across clusters ### Key Characteristics * Linux-based (RHEL-derived) * Immutable image-based updates * Container-first runtime model * Designed for bare metal AI infrastructure, not VMs TractorOS allows FarmGPU to deploy and scale GPU fleets rapidly while maintaining **predictable behavior across thousands of nodes**. *** ## Homestead — Bare-Metal Provisioning & Lifecycle Management **Homestead** is the entry point for every FarmGPU server. ### What Homestead Does * Bare-metal provisioning via Redfish and Ansible * Hardware discovery and inventory * Initial OS deployment and configuration * Support for KVM VMs and LXC containers where needed Homestead ensures that servers move from **rack to production** with minimal manual intervention, forming the foundation for repeatable cluster deployments. *** ## Shepherd — System Health, Reliability & Diagnostics **Shepherd** is FarmGPU's system health evaluation platform, designed to make AI infrastructure **observable, debuggable, and predictable**. ### What Shepherd Monitors * GPU health and performance (via NVIDIA DCGM) * Network health and packet loss * Storage latency and throughput * Host-level metrics and anomalies ### Advanced Capabilities * Predictive failure detection * Agentic diagnostics and root-cause analysis * Automated incident summaries and reports * Integration with observability tools (Grafana, Prometheus, Loki, Alloy) Shepherd enables FarmGPU to move from **reactive monitoring** to **proactive reliability**, reducing downtime and improving SLA adherence. *** ## Silo — Storage Infrastructure & Benchmarking **Silo** is FarmGPU's storage evaluation and optimization platform. ### What Silo Does * Deploys and benchmarks block, file, and object storage systems * Evaluates storage performance under AI workloads * Performs hardware discovery and tuning * Validates storage configurations before production rollout Silo is used internally to benchmark platforms such as MinIO, VAST Data, Weka, and Ceph, ensuring that storage performance is **measured, repeatable, and workload-aware**. *** ## Haystack — AI Workload Profiling & Optimization **Haystack** provides workload-level visibility into how AI jobs consume infrastructure. ### What Haystack Measures * GPU utilization by workload * Storage I/O patterns * Network usage and contention * Differences between training, fine-tuning, and inference jobs This enables: * Per-workload optimization * Capacity planning based on real usage * Improved scheduling and cost efficiency * Better matching of workloads to hardware profiles Haystack helps FarmGPU optimize **goodput**, not just raw utilization. *** ## Automation, Observability & Drivers (Underlying Stack) Across all systems, FarmGPU standardizes on proven, open tooling: ### Automation * **Ansible** for configuration and orchestration ### Observability * **Grafana** * **Prometheus** * **Loki** * **Alloy** * **Node Exporter** * **NVIDIA DCGM** * **cAdvisor** ### Drivers & Acceleration * NVIDIA GPU drivers * SPDK for NVMe * NVIDIA DOCA / OFED for networking and DPU acceleration This approach keeps the stack **transparent, auditable, and extensible**, while avoiding vendor lock-in. *** ## Why This Software Stack Matters FarmGPU's software stack is not a SaaS product suite—it is **operational leverage**. It enables: * Faster cluster bring-up * Higher GPU utilization * Lower operational overhead * Predictable performance at scale * Rapid adoption of new hardware generations By building a Linux-native, hardware-aware control plane, FarmGPU achieves **hyperscaler-level operational maturity without hyperscaler complexity or cost**. # FarmGPU Documentation Source: https://docs.farmgpu.com/index Complete documentation for FarmGPU's GPU infrastructure, services, and operations ## Welcome to FarmGPU FarmGPU is redefining the future of GPU-powered cloud computing by offering cost-effective, scalable, and high-performance GPU resources tailored specifically for AI developers, innovative startups, and enterprises worldwide. Calculate total cost of ownership for GPU clusters and deployments ## Company Overview Learn about FarmGPU's mission, vision, and founding principles. Introduction to FarmGPU and our core offerings Company information, legal entity, and core business description Our mission to democratize AI infrastructure and strategic positioning Why FarmGPU exists and our approach to AI infrastructure ## Products & Services Explore FarmGPU's GPU infrastructure offerings. On-demand GPU cloud, dedicated bare-metal rentals, and managed infrastructure ## Data Center & Operations Learn about FarmGPU's enterprise-grade infrastructure and operational capabilities. Overview of data center infrastructure and operations Power, network, security, and environmental infrastructure Shepherd reliability engine and monitoring capabilities SLA commitments, uptime guarantees, and service credits Terms of service and customer agreements ## Security FarmGPU's comprehensive security architecture and compliance framework. Overview of security architecture and trust layer SOC 2, ISO 27001, and compliance framework Encryption, data protection, and storage security Hardware-backed isolation and confidential AI TractorOS security controls and hardening ## Team & Culture Meet the FarmGPU team and learn about our company culture. Overview of team structure and culture Leadership team and advisors Company values and operating principles ## Technical Architecture Deep dive into FarmGPU's technical infrastructure and systems. Overview of hardware and software stack GPUs, CPUs, storage fabric, and networking infrastructure TractorOS, control plane, and operational software On-demand multi-node AI compute clusters