TomCon.net was built with several recycled computers working on a local network. One recycled ThinkPad T440s runs the web server, Gemini server, and encryption handling. One recycled desktop PC runs Kobold.
There are a few ways to measure power consumption. The most straightforward method uses a wattage metering device (sold for $10-30 USD) that plugs directly into your home's AC outlet as a middleman, metering power draw. This is what I used for the following calculations. Another option would be to directly measure voltage and current, then calculate the precise wattage. I may consider doing this later.
Power consumption is measured in watts (W), which represents the rate of energy use at any given moment. Energy consumption over time is measured in watt-hours (Wh). For example, a 10-watt LED lightbulb turned on for one hour consumes 10 watt-hours of energy. A battery that stores 256Wh of energy can power that same 10-watt lightbulb for about 25.6 hours (256 ÷ 10 = 25.6). This is the same unit your electric bill uses to calculate your energy costs.
The ThinkPad hosting the website, Gemini site, and Caddy consumes approximately 10 watts at idle. Under light load, it increases to 25 watts. It doesn't perform enough computational work to reach its maximum power rating, which would max out around 50-75 watts. I'll need to do stress testing, but upper wattage usage isn't a factor for our setup since the system is idle almost all the time.
Running 24/7 while idle and serving pages, the ThinkPad uses 10 watts continuously, which equals 240Wh per day (10W × 24 hours = 240Wh).
The Kobold desktop handles the majority of heavy computing at TomCon.net. This desktop consumes 30 watts at idle. During light tasks like loading models into GPU memory, it consumes around 60-70 watts. When performing inference under load, wattage spikes up to 200 watts.
The desktop sits idle most of the day. On average, it performs no more than one hour of heavy compute work per day. So the daily calculation is: 30W × 23 hours + 200W × 1 hour = 690Wh + 200Wh = 890Wh used per day.
Combining the laptop and desktop, the total system averages about 47 watts continuously, using 1,130Wh per day (240Wh + 890Wh = 1,130Wh).
Power Consumption Summary ┌─────────────────┬──────────┬─────────────┬──────────────┐ │ Component │ Idle (W) │ Load (W) │ Daily (Wh) │ ├─────────────────┼──────────┼─────────────┼──────────────┤ │ ThinkPad T440s │ 10 │ 25 │ 240 │ │ Desktop PC │ 30 │ 200 │ 890 │ ├─────────────────┼──────────┼─────────────┼──────────────┤ │ Total System │ 40 │ 225 │ 1,130 │ └─────────────────┴──────────┴─────────────┴──────────────┘
For perspective, a plug-in fan uses 80-100Wh per hour. An electric space heater in North America is typically rated at 1,500W. TomCon.net uses less power than a fan on low speed. Over a whole day, TomCon.net consumes less energy than running a space heater for just 45 minutes.
Power Comparison Chart ┌──────────────────────┬─────────────┬─────────────────────┐ │ Device │ Power (W) │ Daily Energy (Wh) │ ├──────────────────────┼─────────────┼─────────────────────┤ │ TomCon.net (avg) │ 47 │ 1,130 │ │ Desktop Fan (low) │ 80 │ 1,920 │ │ LED TV (50") │ 150 │ 3,600 │ │ Electric Heater │ 1,500 │ 36,000 │ │ Home Refrigerator │ 400 │ 9,600 │ └──────────────────────┴─────────────┴─────────────────────┘
The power consumption reflects TomCon.net's dual purpose as both a web platform and an AI research service. The ThinkPad handles all web serving, Gemini protocol connections, and SSL certificate management with remarkably consistent power draw. The desktop's variable consumption comes from hosting the kobold.cpp service, which provides free access to open source language models like COSMOS-REASON and DeepHermes for fellow researchers and hobbyists.
Model inference creates the most significant power spikes, particularly during the initial loading of large language models into GPU memory. Once loaded, inference requests maintain moderate power levels, though multiple concurrent users or complex reasoning tasks can sustain higher consumption. The system spends most time idle, waiting for inference requests rather than actively processing them.
There's room for efficiency gains:
The machine learning workload presents unique optimization opportunities. Model quantization techniques explored in my semantic manifold research suggest that smaller, more efficient models can maintain capability while reducing computational overhead. GPU power management could potentially reduce idle consumption, though this must be balanced against response time requirements for the public kobold service.
Thermal management also affects efficiency. During extended inference sessions, the desktop's cooling requirements increase power consumption beyond the GPU and CPU loads alone.
Running AI models locally presents interesting trade-offs compared to cloud-based alternatives. While my desktop consumes more power per inference than a highly optimized data center GPU, the comparison becomes complex when considering transmission costs, data center cooling, and infrastructure overhead. Local inference also eliminates the privacy costs of sending data to third-party services.
The sustainability equation extends beyond immediate power consumption. By extending the useful life of recycled hardware that might otherwise become e-waste, the system provides environmental benefits that pure efficiency metrics miss. The 8-year-old desktop continues serving a productive role rather than contributing to landfills.
The solar system powering the ThinkPad demonstrates the potential for completely carbon-neutral web hosting. While the current setup only covers the web server, scaling renewable capacity to handle the AI workload represents an achievable goal. The predictable nature of the base web serving load makes it ideal for solar power, while the variable AI workload could potentially be scheduled around renewable energy availability.
Once the laptop runs entirely on solar power, its energy operation cost becomes essentially free, with no ongoing electricity bills beyond the initial investment in solar equipment. My current solar system isn't sufficient to handle the desktop's power requirements, so it remains on grid power and costs a few dollars per month to operate.
Hardware Investment Breakdown ┌─────────────────────┬─────────────┬─────────────┬─────────────┐ │ Component │ Cost (USD) │ Age (years) │ Condition │ ├─────────────────────┼─────────────┼─────────────┼─────────────┤ │ ThinkPad T440s │ $100 │ 5 │ Used │ │ Desktop PC │ $1,000 │ 8 │ New │ │ Solar Panels (200W) │ $200 │ 2 │ New │ │ Battery Storage │ $100 │ 2 │ New │ │ Charge Controller │ $20 │ 2 │ New │ ├─────────────────────┼─────────────┼─────────────┼─────────────┤ │ Total System Cost │ $1,420 │ - │ - │ └─────────────────────┴─────────────┴─────────────┴─────────────┘
The desktop represents an interesting case study in long-term technology investment. Purchased new in 2017 for $1,000, it has now provided 8 years of continuous service.
Cost vs Value Over Time Analysis ┌──────┬─────────────┬──────────────┬─────────────┬─────────────┐ │ Year │ Market Value│ Annual Cost │ Cloud Alt. │ Value Ratio │ ├──────┼─────────────┼──────────────┼─────────────┼─────────────┤ │ 2017 │ $1,000 │ $125.00 │ $182.50 │ 1.46x │ │ 2018 │ $650 │ $125.00 │ $182.50 │ 1.46x │ │ 2019 │ $450 │ $125.00 │ $182.50 │ 1.46x │ │ 2020 │ $320 │ $125.00 │ $182.50 │ 1.46x │ │ 2021 │ $250 │ $125.00 │ $182.50 │ 1.46x │ │ 2022 │ $200 │ $125.00 │ $182.50 │ 1.46x │ │ 2023 │ $150 │ $125.00 │ $182.50 │ 1.46x │ │ 2024 │ $120 │ $125.00 │ $182.50 │ 1.46x │ │ 2025 │ $100 │ $125.00 │ $182.50 │ 1.46x │ └──────┴─────────────┴──────────────┴─────────────┴─────────────┘ Total Cost of Ownership: $1,311.86 (hardware + electricity) Cloud Computing Alternative: $1,460.00 (same workload) Break-even Point: 7.2 years Current Net Savings: $148.14
The analysis reveals several insights about long-term infrastructure investment. While the desktop has depreciated 90% in market value, it continues providing full utility for AI workloads. The annual cost of $125 for hardware plus approximately $39 for electricity compares favorably to equivalent cloud GPU services at roughly $182 per year.
More significantly, the system broke even against cloud alternatives at 7.2 years and now operates at a net saving. The decision to keep running "legacy" hardware rather than following typical 3-4 year replacement cycles has avoided approximately $1,600 in additional hardware costs while maintaining adequate performance for the intended workload.
Rather than replacing the entire system, I've made targeted upgrades to extend capability and optimize for AI workloads. Notably, these upgrades all occurred in 2025, meaning the original 2017 system provided 8 years of service before requiring additional investment:
Desktop Upgrade History (2025) ┌─────────────────────────────────┬─────────────┬─────────────────────────┐ │ Component │ Cost (USD) │ Purpose │ ├─────────────────────────────────┼─────────────┼─────────────────────────┤ │ Tesla P100 GPU │ $200 │ AI inference upgrade │ │ Corsair PSU 850W │ $120 │ Power dual GPU setup │ │ Adapter cables │ $15 │ GPU connectivity │ │ Dual GPU trigger board │ $15 │ Multi-GPU coordination │ │ PCIe 16x to 4x4x4x4 adapter │ $30 │ Expand PCIe slots │ │ 600mm PCIe riser extension │ $30 │ Physical GPU placement │ ├─────────────────────────────────┼─────────────┼─────────────────────────┤ │ Total 2025 Upgrade Investment │ $410 │ Single upgrade cycle │ └─────────────────────────────────┴─────────────┴─────────────────────────┘ Investment Timeline: 2017-2024: $1,000 original system (7 years, $142.86/year) 2025: $410 upgrade investment Total: $1,410 over 8 years = $176.25/year average
This timeline reveals even stronger value than initially calculated. The original $1,000 investment provided 7+ years of continuous service before requiring any additional hardware investment. The 2025 upgrades represent a strategic capability expansion rather than maintenance of a degrading system.
The upgrade strategy demonstrates practical sustainability principles. Instead of discarding the original system when AI workloads demanded more GPU memory, targeted component upgrades extended its useful life. The Tesla P100, while several generations old, provides 16GB of HBM2 memory ideal for hosting larger language models that wouldn't fit on the original GTX 1070ti's 8GB.
The additional infrastructure components (PSU upgrade, adapters, trigger boards) represent the engineering overhead of maintaining older hardware in non-standard configurations. While this adds complexity compared to buying a new AI-optimized system, it demonstrates how incremental investment can preserve the utility of existing infrastructure.
The financial analysis above only captures direct compute costs, but local AI infrastructure provides additional value that's difficult to quantify:
**Privacy and Data Sovereignty:** Unlike cloud AI services that log conversations and may use data for training, local inference ensures complete privacy. All processing happens on-premise with zero data transmission to third parties. For research work involving sensitive or proprietary information, this privacy guarantee has significant value.
**Unlimited Access:** Cloud AI services typically impose rate limits, usage caps, or per-token pricing that can become expensive with heavy usage. The local system provides truly unlimited access - researchers can run thousands of inference requests for experimentation without worrying about escalating costs or hitting API limits.
**Model Flexibility:** Local infrastructure allows testing of any open-source model, including experimental or fine-tuned versions not available through commercial APIs. This flexibility is crucial for research applications where specific model capabilities are needed.
**Always-On Availability:** The system provides 24/7 availability without dependency on external service providers, internet connectivity, or third-party infrastructure uptime. For time-sensitive research or international collaborators in different time zones, this reliability has substantial value.
Total hardware investment now stands at $1,410 over 8 years, or $176.25 annually. While this approaches the raw compute cost of cloud alternatives, the privacy, unlimited access, and research flexibility provided make the local infrastructure significantly more valuable than direct cost comparisons suggest. For the TomCon.net research mission, these factors justify the incremental investment in maintaining and upgrading older hardware rather than relying on external services.
To put the upgrade investment in perspective, building a new system with equivalent AI capabilities would cost significantly more in today's market:
New System Build Cost Estimate (2025) ┌─────────────────────────────────┬─────────────┬─────────────────────────┐ │ Component │ Cost (USD) │ Specification │ ├─────────────────────────────────┼─────────────┼─────────────────────────┤ │ Modern CPU (Ryzen 5/Intel i5) │ $200-300 │ Current gen processor │ │ Motherboard (ATX) │ $150-200 │ PCIe 4.0, modern I/O │ │ 32GB DDR4/5 RAM │ $100-150 │ Adequate for AI work │ │ 1TB NVMe SSD │ $80-120 │ Fast storage │ │ Case + Cooling │ $100-180 │ Proper airflow │ │ 850W PSU (80+ Gold) │ $120-160 │ Efficient power supply │ │ GPU (16GB+ VRAM) │ $700-1,600+ │ Used 3090: $700+ │ │ │ │ RTX 4080: $1,000+ │ │ │ │ RTX 4090: $1,600+ │ ├─────────────────────────────────┼─────────────┼─────────────────────────┤ │ Total New System Cost │$1,450-2,710+│ Modern equivalent build │ ├─────────────────────────────────┼─────────────┼─────────────────────────┤ │ TomCon.net Upgraded System │ $1,410 │ 8-year total investment │ │ Cost Savings vs New │ $40-1,300+│ Massive GPU savings │ └─────────────────────────────────┴─────────────┴─────────────────────────┘
The comparison reveals that the current GPU market makes the Tesla P100 upgrade an exceptional bargain. Even used RTX 3090s with 24GB VRAM command $700+ on eBay, while new GPUs with adequate VRAM for AI workloads start around $1,000 for an RTX 4080 (16GB) and exceed $1,600 for an RTX 4090 (24GB).
**Current GPU Market Reality:**
- **Tesla P100 (16GB HBM2): $200** - Incredible value for AI inference
- **Used RTX 3090 (24GB): $700+** - Still 3.5x more expensive
- **RTX 4080 (16GB): $1,000+** - 5x more expensive than P100
- **RTX 4090 (24GB): $1,600+** - 8x more expensive than P100
The $200 Tesla P100 upgrade now appears as one of the best price-to-VRAM ratios available in the current market. While newer GPUs offer better performance per watt and more modern features, the P100's 16GB of high-bandwidth memory remains perfectly adequate for hosting large language models and provides exceptional value in today's inflated GPU market.
**Key Advantages of the Upgrade Approach:**
- **Proven Stability:** The base system has 8 years of reliable operation
- **Incremental Risk:** $410 upgrade vs $1,150+ complete rebuild
- **Environmental Impact:** Extends hardware lifecycle vs creating new e-waste
- **Performance Adequacy:** Tesla P100's 16GB HBM2 remains excellent for inference workloads
The upgrade strategy demonstrates that with careful component selection, older systems can be cost-effectively enhanced to meet modern AI research requirements without the expense and environmental impact of complete replacement.
The manufacturing of these computer components involved releasing CO2 and consuming non-renewable resources through industrial processes. Precious metals like gold and lithium were mined through intensive human labor to create circuit boards and batteries. High-grade silica was processed for transistors and accelerator cards. Non-renewable oil was extracted, refined, and synthesized into non-biodegradable plastics for many components.
However, by using recycled hardware and renewable energy where possible, TomCon.net minimizes its ongoing environmental impact while extending the useful life of existing technology.
This power measurement exercise connects directly to broader questions in AI research about efficiency and accessibility. Local inference democratizes access to large language models, allowing researchers and hobbyists to experiment without relying on expensive cloud services or compromising data privacy.
The relationship between model size, quantization techniques, and power consumption offers practical insights for sustainable AI development. Smaller, well-optimized models can often achieve comparable results to larger ones while requiring significantly less computational resources. This aligns with research directions explored in my semantic manifold work, where archetypal attractors suggest efficient compression methods that preserve model capability.
Future iterations of this analysis could benefit from more granular monitoring. Tracking GPU utilization, thermal states, and request patterns would provide deeper insights into optimization opportunities. The current measurement approach using AC power meters captures total system consumption but obscures the individual component contributions during different workload states.
Documenting these power characteristics also serves the broader open source community. Other researchers building similar local AI infrastructure can use these measurements as baseline references for their own efficiency optimization efforts.
TomCon.net's power consumption profile reflects its dual mission as both a sustainable web platform and an accessible AI research service. The modest energy requirements demonstrate that meaningful digital infrastructure can operate within reasonable power budgets, especially when built on recycled hardware and renewable energy.
The measurements also highlight the practical considerations for anyone interested in hosting local AI services. While the variable power demands of machine learning workloads present optimization challenges, they remain manageable for hobbyist and research applications. The combination of efficient web serving with on-demand AI capabilities creates a platform that balances accessibility, privacy, and sustainability.
As both the renewable energy capacity and AI optimization techniques continue evolving, future iterations of this infrastructure will likely achieve even better efficiency ratios while maintaining or expanding service capabilities.
For readers interested in measuring their own systems, here are the basic electrical formulas used in these calculations:
Energy (Wh) = Power (W) × Time (hours) Daily Energy = Power × 24 hours Monthly Energy = Daily Energy × 30 days Annual Energy = Daily Energy × 365 days
Cost = Energy (kWh) × Rate ($/kWh) Where: kWh = Wh ÷ 1000 Monthly Cost = (Daily Energy ÷ 1000) × 30 × Rate Annual Cost = (Daily Energy ÷ 1000) × 365 × Rate
Average Power = (Power1 × Time1 + Power2 × Time2 + ... + PowerN × TimeN) ÷ Total Time Example: 30W × 23h + 200W × 1h = (690 + 200) ÷ 24 = 37.08W average
Efficiency Ratio = Useful Output ÷ Total Input Power per Unit Work = Total Power ÷ Work Units Cost per Unit Work = Total Cost ÷ Work Units Example: Cost per AI inference = Daily AI Cost ÷ Daily Inference Count
Power (W) = Voltage (V) × Current (A) Power (W) = Voltage² (V²) ÷ Resistance (Ω) Power (W) = Current² (A²) × Resistance (Ω)
Required Solar Capacity = Daily Energy Consumption ÷ Peak Sun Hours Battery Capacity = Daily Energy × Days of Autonomy ÷ Depth of Discharge Example: For 240Wh daily with 5 peak sun hours: Solar Panel Size = 240Wh ÷ 5h = 48W minimum
These formulas provide the foundation for analyzing any system's power consumption and costs. Remember to account for conversion losses, seasonal variations, and safety margins when designing renewable energy systems.