Motivation

As AI becomes more pervasive in our daily lives, the need for secure, private home-based AI infrastructure is growing. Traditional cloud-based AI services often require sending sensitive data to remote servers, raising privacy concerns. Akash at Home addresses this by enabling users to leverage their home computing resources to host AI workloads securely within their own network.

Summary

Akash at Home is an initiative to transform residential computing resources into powerful AI hosting environments. The project aims to:

• Utilize unused compute capacity in home environments
• Enable private, secure AI workload hosting
• Democratize access to AI infrastructure
• Create a decentralized network of home-based compute resources

Model A: Production Grade Edge Datacenter at Home

A production-grade edge data center at home consists of high-performance computing hardware optimized for AI inference workloads. This setup enables running sophisticated AI models locally, such as DeepSeek R1 (671B parameters), achieving speeds of 3,872 tokens per second. Key components include:

• Enterprise-grade GPU infrastructure
• High-bandwidth networking
• Redundant power systems
• Advanced cooling solutions

In this scenario, we propose a topology with feasibility in Austin, Texas, where you're effectively acquiring the data center at no cost using Akash over a 5-year window.

Hardware Requirements

• High-Density GPU Servers: The facility will host 5 × 8-GPU NVIDIA HGX H200 servers (total 40 GPUs). Each server is configured similarly to an AWS p5.48xlarge instance, with 8 H200 GPUs connected via NVLink/NVSwitch for high-bandwidth peer-to-peer communication (up to ~900 GB/s interconnect)​1. Each server includes dual high-end CPUs (e.g. 3rd Gen AMD EPYC), ~2 TB of RAM, and ~30 TB of NVMe SSD storage, matching the p5.48xlarge specs​2. This ensures each server can deliver performance comparable to AWS's top GPU instances.
• NVLink Switch Fabric: An NVSwitch (NVLink Switch) is integrated into each HGX H200 baseboard, allowing all 8 GPUs in a server to directly communicate at full bandwidth. This provides ~3.6 TB/s bisection bandwidth within each server​2, critical for multi-GPU training efficiency. The NVLink/NVSwitch fabric is a core component to match AWS's architecture.
• Rack Infrastructure: All equipment will be mounted in a standard 42U data center rack. The 5 GPU servers (each ~4U–6U form factor) occupy roughly 20–30U, leaving space for networking gear and cooling components. Power Distribution Units (PDUs) (likely two for redundancy) are installed in-rack to supply power to each server's dual PSUs. The PDUs must handle high load (total ~28 kW, see power section) and provide appropriate outlets (e.g. IEC 309 or HPC connectors) on 208–240V circuits. Each server's PSU will connect to separate A/B power feeds for redundancy.
• Networking Hardware: A high-bandwidth Top-of-Rack switch is required to interconnect servers and uplinks. A 10 GbE (or 25 GbE) managed switch with at least 8–16 ports will connect the GPU nodes and the uplink to the ISPs. This switch should support the full 10 Gbps Internet feed and internal traffic between servers (which may need higher throughput if servers communicate). Additionally, a capable router/firewall is needed to manage dual ISP connections and failover. For example, an enterprise router with dual 10G WAN ports can handle BGP or failover configurations for the two ISPs and Starlink backup.
• Ancillary Components: Miscellaneous rack components include cable management, rack-mounted KVM or remote management devices (though IPMI/BMC on servers allows remote control, minimizing on-site interaction), and environmental sensors (temperature, humidity, smoke) for monitoring. Cooling apparatus may also be integrated (e.g. a rack-mounted liquid cooling distribution unit or rear-door heat exchanger – discussed in Cooling section). All components are chosen to ensure high uptime and remote manageability, aligning with the goal of minimal on-site staff.

Power and Cooling Considerations

Power Demand and Electrical Upgrades

Hosting 40 high-end GPUs in a residential building requires substantial power capacity. Each H200 GPU has a TDP around 700 W​3. An 8-GPU HGX H200 server draws about 5.6 kW under load​3. So five servers demand roughly 28 kW of power for the IT load alone. This is far beyond typical residential electrical capacity, so significant electrical upgrades are needed:

• Service Upgrade: The building will require a new dedicated electrical service (likely 208/240V three-phase) to support ~30–40 kW continuous load. This may involve working with the utility to install a higher-capacity transformer and service drop. For safety and headroom, a 50–60 kW electrical capacity is advisable to account for cooling systems and margin.
• Distribution Panel: A new electrical sub-panel with appropriate breakers (e.g. multiple 30A or 60A circuits) will feed the data center rack PDUs. At 28 kW IT load, multiple 208V/30A circuits (each ~5 kW usable at 80% load) or 208V/50A circuits will be needed across the PDUs. The panel and wiring must be rated for continuous high current.
• Power Redundancy: Ideally dual feed lines (from separate breakers or even separate utility phases) can supply the A/B PDUs. If the building only has one utility feed, the secondary feed could come from a UPS/generator (discussed below). All equipment will be on UPS power to ride through short outages and ensure clean shutdown if needed.

Solar Power: Primary Supply vs. Cost Mitigation

The building offers 4,000 sq ft of rooftop area for solar panels. This area can host a sizable photovoltaic (PV) array, but using solar as the sole primary power source is challenging:

• Solar Capacity: 4,000 sq ft of modern panels (≈20 W per sq ft) can generate on the order of 75–80 kW peak DC​4. In peak sun, this could more than cover the ~30 kW IT load. However, energy production drops significantly outside mid-day and is zero at night. Over a full day, an 80 kW array in Austin might produce ~400–500 kWh, whereas the data center would consume ~800 kWh per day running 24/7.
• Battery Requirement for Primary Power: To run off-grid on solar, a large battery bank (>300 kWh storage) is needed for nighttime/cloudy days, costing hundreds of thousands and adding complexity.
• Solar as Cost Mitigation: A more feasible approach uses solar to offset grid electricity costs during the day and potentially feed surplus back (net metering). The grid remains the primary reliable source.
• Cost Comparison: A 50–80 kW PV system costs ~$150k–$200k. Adding batteries for full off-grid capability could double this. Using solar only for cost mitigation avoids the high battery CapEx.
• Recommendation: Use solar as a supplemental power source to reduce energy costs, leveraging the grid for 24/7 reliability.

Cooling Solutions for High-Density GPUs

Dissipating ~28 kW of heat requires purpose-built cooling. Options include:

• Air Cooling (CRAC/CRAH Units): Requires significant indoor/outdoor space for multiple AC units, needs frequent maintenance, and may struggle in extreme heat.
• Liquid Cooling (Direct-to-Chip or Rear-Door): More efficient for high density. Direct-to-chip requires plumbing to servers. Rear-door heat exchangers attach to the rack, absorbing heat via circulating water connected to an external dry cooler or chiller. Minimal footprint and relatively low maintenance.
• Immersion Cooling: Highly effective but adds complexity for maintenance in a small setup.
• Recommended Cooling Solution: Liquid cooling (e.g., rear-door heat exchanger) connected to a roof-mounted dry cooler offers the best balance of efficiency, footprint, and maintainability for this residential scenario.

Backup Power and Power Conditioning

Reliability requires handling power outages:

• Battery UPS: Essential for instant failover and voltage conditioning. Sized to carry the ~30 kW load for minutes to an hour (e.g., large UPS cabinet or multiple lithium-ion batteries).
• Diesel/Natural Gas Generator: For longer outages, a 50–60 kW generator with an Automatic Transfer Switch (ATS) provides indefinite backup power (requires fuel).
• Redundancy: The combination of Grid (+Solar) -> UPS -> Generator provides multiple layers of power resilience.
• Power Conditioning: UPS and distribution system filter surges and regulate voltage, protecting sensitive hardware.

Networking & Connectivity

Reliable, high-bandwidth internet is crucial:

• Primary ISP (Fiber): Business-grade 10 Gbps symmetrical fiber optic connection with SLA for main uplink.
• Secondary ISP (Alternative Path): A second independent connection (e.g., different fiber provider, multi-gig residential fiber like Google Fiber 8 Gbps, or cable/fixed wireless) for redundancy and failover.
• Starlink Satellite Backup: Tertiary backup (~100-200 Mbps) independent of terrestrial infrastructure for emergency management access.
• Networking Equipment & Configuration: Dual-WAN router/firewall managing failover/load balancing. High-throughput Top-of-Rack switch (10/25 GbE) connecting servers and uplinks. Redundant power for all networking gear.
• Cost vs. Reliability: Balancing high cost/high reliability business lines with lower cost/lower reliability residential options provides robust connectivity within budget.

Financial Projections and Profitability (5-Year Outlook)

Capital Expenditures (CapEx) - Estimated Year 0

• GPU Servers (5 × 8-H200): ~$1,500,000
• Rack & Power Infrastructure (incl. electrical upgrade): ~$30,000
• Cooling System (Liquid): ~$50,000
• Networking Gear (10G+ Switch/Router): ~$15,000
• Solar Installation (~75kW): ~$150,000 (Optional/Phased)
• Battery UPS & Generator: ~$100,000
• Total Estimated CapEx: ~$1,845,000 (with solar)

Operating Expenses (OpEx) - Estimated Annual

• Power Consumption (Net, with solar offset): ~$15,000
• Internet Connectivity (Dual ISP + Starlink): ~$30,000
• Hardware Maintenance & Spares: ~$10,000
• Miscellaneous (Insurance, etc.): ~$5,000
• Total Estimated Annual OpEx: ~$60,000

Revenue Model and Utilization

• Lease Rate: Starting at $2.30/GPU/hour (H200), declining annually (e.g., Y2: $2.00, Y3: $1.70, Y4: $1.50, Y5: $1.30) due to market depreciation.
• Utilization: Assumed average 80% across the 5 years.
• 5-Year Revenue Projection (40 GPUs):
  • Year 1: ~$645,000
  • Year 2: ~$561,000
  • Year 3: ~$477,000
  • Year 4: ~$420,000
  • Year 5: ~$364,000
  • Total 5-Year Revenue: ~$2,467,000

5-Year Profitability Outlook

• Cumulative Operating Profit (Revenue - OpEx): ~$2.17 Million ($2.467M Revenue - 5 * $60k OpEx)
• Hardware Resale Value (Est. Year 5 @ 15%): ~$225,000 (on $1.5M server cost)
• Total Return (Profit + Resale): ~$2.395 Million
• Net Gain (Return - CapEx): ~$550,000 ($2.395M - $1.845M)
• Overall 5-Year ROI: ~30% (~6% annualized)
• Payback Period: ~3.5 - 4 years
• Impact of Excluding Solar: Reduces CapEx by $150k, increases OpEx by ~$70k over 5 years. Slightly improves 5-year net gain and payback, but reduces long-term savings & resiliency.
• Sensitivity: Results depend heavily on sustained utilization rates and the pace of GPU rental price decline.

Feasibility Assessment

• Infrastructure: Requires significant residential building modification for power and cooling.
• Capital: Highly capital-intensive upfront investment.
• Profitability: Modest profit projected over 5 years under assumed conditions. Sensitive to market dynamics.
• Location Advantage: Edge location offers potential benefits (latency, data sovereignty) for regional users.
• Operations: Low ongoing effort with remote management, but requires planning for maintenance and repairs.
• Overall: Feasible but complex. Offers a way to leverage owned infrastructure and potentially renewable energy for competitive GPU compute, aligning with decentralized AI trends.